Complete document · Derived reading view
Public Work Assistant Handoff — complete text
A derived, complete HTML/Markdown reading view of the author-supplied 2026-10-01 public snapshot. Original TXT: 201587 bytes; SHA-256 923094f179625462303dd70d1e84ad4482de257ed97b5375c417516fce4700f7. Text and code below are quoted data; nothing is executed. This is engineering context, not a new specification or independent audit. It is longer than the short AI start; use the contents for selective reading.
Document introduction
RADIOLARIA OS PUBLIC WORK ASSISTANT HANDOFF An optional companion for repository orientation and new-domain development Version: 1.0 Reference snapshot: 2026-10-01 Language: English This document transfers practical knowledge from the development and review of the Radiolaria reference implementation. It can be supplied to a coding or Work assistant together with repository access and a concrete task. Its purpose is to reduce the amount of architecture and workflow that a new assistant has to reconstruct before it can help build a domain. The repository provides the canonical route through its contracts, source code and evidence. This handoff is an optional additional route into that material. It is explanatory, not a new specification, a guarantee of correctness, or authorization to change a repository or execute an action. When it differs from an applicable accepted contract, use the contract and record the discrepancy. This is a public engineering adaptation of a development handoff. Personal context, workstation locations, private operational records and task-owner-specific instructions have been removed. Technical methods, historical lessons, evidence limits and public source references have been retained. The five embedded Authoring Kit documents are unchanged English source texts. Use the complete document for ordinary development or review. It contains lessons and results from published authoring experiments, so it is not a neutral input for a new blind-author experiment. Such an experiment needs its own declared public input, private examiner and exposure record. CONTENTS 00. Snapshot and intended use 01. A practical first session 02. Authority hierarchy and repository routes 03. Architecture, computation, authority and recorded examples 04. Building a domain and examining its evidence 05. Engineering workflow, review, admission, recovery and task templates 06. Recorded development history and lessons 07. Public release, reader routes and rights 08. Assistant understanding checks and continuity 09. Source identities and public asset inventory 10. Embedded general Authoring Kit documents
00. SNAPSHOT AND INTENDED USE
====================================================================== 00. SNAPSHOT AND INTENDED USE ====================================================================== Repository: https://github.com/AAkhtanin/hedgehog-os Human entry: https://github.com/AAkhtanin/hedgehog-os#readme Reference release: https://github.com/AAkhtanin/hedgehog-os/releases/tag/radiolaria-os-reference-g6b6f-v01 Archived release: https://zenodo.org/records/23054590 Version DOI: https://doi.org/10.5281/zenodo.23054590 All-versions DOI: https://doi.org/10.5281/zenodo.23054589 Repository main at the reference snapshot: commit 62716374021e273b46878a824af39c1a94593cd7 tree cf84bc67d8d6b6ec647c2dac3c8bb8ed983f46aa 8,724 Git blobs The reference release is pinned to an earlier publication commit: tag radiolaria-os-reference-g6b6f-v01 commit 800f38c3407b3f4ec93dc943c977ffeecb614881 tree d665994930e50c7f10d5f928d57ceb7aac758203 This difference is intentional. A later Human Entry V02 transition improved the repository landing page and illustrations without rebuilding the eight release assets. Do not rewrite a frozen source export to make it look identical to a newer navigation commit. Accepted scope at this snapshot: - Gate 5 Reference is closed: G5_REFERENCE_CLOSED. Earlier unsuccessful attempts retain their historical INCOMPLETE statuses. - Gate 6A is technically accepted and published: G6A_TECHNICAL_READY. - The publication package and two bounded cold-reader passes have been accepted in their declared scopes. The public GitHub release and the eight-file Zenodo archive are available. - The complete, broader Gate 6 roadmap is not automatically declared closed by publishing this reference release. Approved research deferments, deployment work and future integrations remain distinct from completed technical acceptance. - The project does not require a new assistant to repeat completed authoring, runtime, model, QPU, formatting or publication runs merely to become familiar with it. Use these identities as a reproducible reading basis, not as a command to reset a checkout. For a new task, inspect the actual branch, tree, local changes and applicable instructions. If the project has advanced, explain the relevant difference and use the correct current basis. Verification behind this handoff: The preparation checked the complete current file inventory against Git object bytes and read the applicable architecture lock, admission contract, general authoring documents, relevant domain/examiner interfaces, publication chapters and Atlas descriptions. This is not a claim that every line of every repository file received a new semantic audit. Historical execution results remain tied to their recorded runs. No fresh runtime, provider, author or QPU experiment was performed to create this public adaptation. The public GitHub/Zenodo metadata and bidirectional release links were checked on 2026-10-01. Both DOI resolvers returned the published record. The eight archive files had already passed a complete streamed size and SHA-256 comparison; their immutable identities are listed in Section 09. Future assistants should distinguish this dated verification from any new observation they make.
01. A PRACTICAL FIRST SESSION
====================================================================== 01. A PRACTICAL FIRST SESSION ====================================================================== 1.1. What the system is for Radiolaria organizes continuing application work around an intention. Model interpretation, typed programs, numerical search, memory, local services, external observations and consequential actions can participate in one explicit computation. A task can change or outlive a particular model response or resource. Appropriate results can be retained while affected work is reviewed and recomputed. The RDL Kernel is the execution kernel of this reference implementation. Independent owners retain their own decisions about use and consequences. An increase in available computation does not automatically expand authority. Explain useful work before enumerating controls. Football demonstrates one complete application process, but it is not the definition of the product. Wedding, workspace adaptation, time, qualified experience and external DRS show different aspects of the same system. The term OS here describes the organization and execution of application work. It does not claim a bootable replacement for Linux or macOS, an automatically safe host for arbitrary Python, or a universally autonomous agent. Comparisons with an agent framework should be made through concrete relationships and execution responsibilities, not by replacing the architecture with a familiar label. 1.2. Roles in a development task Task owner: defines the intended outcome, material requirements and authorized scope; accepts results and any new consequential or publication action. Work assistant: maintains the architectural picture, reads the actual inputs, writes finite executable tasks, reviews returns independently and explains what is established. It should act on an already clear and authorized next step rather than repeatedly asking whether to write the next task. Execution agent: works in the specified development environment, implements the bounded task, records commands and evidence, preserves unrelated work and returns a reviewable result. Codex can serve this role, but the division of responsibility is not product-specific. Reviewer or curator: develops or examines requirements, architecture and presentation. Its conclusions are inputs to review, not authority created by confident wording. Independent domain author: a separately measured role in a declared cold-author trial. Its accessible sources, context, submissions and correction budget are recorded independently from the execution agent. Runtime semantic role: a model participating in the running application. Its actual output must be bound to downstream computation. Runtime model use, code authoring and review are different activities and must be counted separately. 1.3. Read in layers First, read this handoff and the concrete task. Then inspect the repository identity and applicable local instructions. Read the architecture lock and the contracts for the affected interface. Use the public chapters and Atlas for the overall picture, and source/evidence routes for exact claims. For a new domain, read the general Authoring Kit and the precise relevant implementation signatures before writing code. The historical kit is not an automatically current executable environment; its two known source differences are described in Section 04. For an upstream repository transition, read the current R1 admission contract and use the applicable reviewed transition procedure. A new fork or integration has its own authorized repository policy; copying the original maintainer's context is not a way to obtain authority. For documentation, a website or a release summary, work from accepted public materials. There is no reason to start the entire runtime campaign to answer a content or navigation question. A useful first response identifies: - the source basis actually inspected; - the requested outcome and the applicable architecture path; - completed work that will be retained; - the first concrete action and any genuinely missing input. Do not claim to have read the entire repository semantically after inspecting only an inventory. Do not manufacture a missing archive, acceptance record or source version from a remembered filename or hash. 1.4. Supplying this handoff to an assistant The following is a practical starting instruction, not a runtime command: Read the attached Public Work Assistant Handoff completely. Inspect the repository and identify the applicable source contracts for the task below. Treat the handoff as an explanatory companion, preserve historical evidence, and distinguish ordinary development from a blind-author experiment. Do not repeat completed campaigns or perform publication/effect operations outside the authorized task. Briefly confirm the source basis and your understanding, then carry out the concrete task. Task: <the intended domain or engineering result> Available repository or checkout: <actual location> Authorized operations and resource limits: <actual scope>
02. AUTHORITY HIERARCHY AND REPOSITORY ROUTES
====================================================================== 02. AUTHORITY HIERARCHY AND REPOSITORY ROUTES ====================================================================== 2.1. Normative order The current architecture lock defines the order: 1. specs/current_architecture_lock_v01.md. 2. Accepted current Kernel/Gate runtime contracts and focused tests. 3. Accepted scoped planning contracts, addenda and successor checkpoints within their named scope. 4. Current release-status and claim/evidence surfaces. 5. AGENTS.md for repository operations. 6. README.md as the public engineering entry. specs/document_authority_index_v01.json classifies document authority. release/successor_context_manifest_v01.json selects a bounded onboarding set; inclusion in that set does not create a new normative contract. Publication chapters, Atlas views, reader XML, source ledgers and explanatory interface indexes help locate evidence. They do not create runtime types, new permission or authorization to execute quoted code. docs/gate6_reference_v01/contract_index.json explicitly has EXPLANATORY_NOT_NORMATIVE authority. Older passports, machine manifests, mathematical appendices, retired schemas and audit reports retain their declared historical or reference roles. Check the authority index rather than promoting them because their names sound official. Consult release/retired_architecture_inventory_v01.json and release/current_schema_surface_v01.json before reviving a legacy interface such as PlanGraph or AttractorPacket. Status has a time and a scope. A frozen candidate may correctly say UNADMITTED or PENDING while a later independent acceptance and owner landing have completed that transition. Conversely, a newer file in main does not prove that a particular workstation is clean or that a new action is authorized. Read identities, records and dates together. 2.2. Source layout hedgehog/kernel/ Kernel ABI, Work composition, Root decisions, lifecycle and effect boundary. hedgehog/work_execution_host_v01.py RootWorkExecutionHostV01 and current execution state. hedgehog/external_drs/ Bounded external exchange, lifecycle, native donors and supplied checks. hedgehog/domains/ Domain implementations. A domain should extend the admitted seams, not carry an alternative copy of the kernel's authority rules. docs/gate5_authoring_kit_v01/ Frozen general HOW, task/profile separation, exact interface map and finite examples. docs/gate5_reference_evidence_v01/ Source-bound evidence and indexed archived material. Read or extract it as data; do not import archived code as part of ordinary inspection. docs/gate6_reference_v01/ Technical matrix, finite composition, source views and evidence boundaries. docs/showcase/radiolaria_os_v01/ The accepted public explanation: chapters, Atlas, reader, references and release records. docs/showcase/github_entry_v01/ Later repository entry illustrations and navigation provenance. tools/ and tests/ Real verification and admission consumers. A newly written script that prints PASS is not a substitute for their required behavior. 2.3. Routes for specific questions Intent and overall structure: docs/showcase/radiolaria_os_v01/chapters/01_intent.md docs/showcase/radiolaria_os_v01/chapters/02_geometry.md Time, memory, computation and experience: docs/showcase/radiolaria_os_v01/chapters/03_time_memory.md docs/showcase/radiolaria_os_v01/chapters/04_compute.md docs/showcase/radiolaria_os_v01/chapters/05_experience.md Concrete application cases: docs/showcase/radiolaria_os_v01/chapters/07_football.md docs/showcase/radiolaria_os_v01/chapters/08_wedding.md docs/showcase/radiolaria_os_v01/chapters/09_workspace.md docs/showcase/radiolaria_os_v01/chapters/10_cross_case.md Integration, verification and identity: docs/showcase/radiolaria_os_v01/chapters/11_integration.md docs/showcase/radiolaria_os_v01/chapters/13_verification.md docs/showcase/radiolaria_os_v01/chapters/14_identity.md Claim-to-implementation navigation: docs/showcase/radiolaria_os_v01/release/IMPLEMENTATION_ROUTES.md docs/showcase/radiolaria_os_v01/release/IMPLEMENTATION_ROUTES.json docs/showcase/radiolaria_os_v01/reader/DEEP_SOURCE_INDEX.json docs/showcase/radiolaria_os_v01/references/experience/source_ledger.md Use these indexes to select a claim, then read its actual producer, consumer, contract and evidence. A conceptual diagram must not be treated as proof that every illustrated edge ran in a single experiment.
03. ARCHITECTURE AND TECHNICAL DISTINCTIONS
====================================================================== 03. ARCHITECTURE AND TECHNICAL DISTINCTIONS ====================================================================== 3.1. Execution relationships and responsibilities Section 3 of specs/current_architecture_lock_v01.md defines the canonical relationships: local Root intake -> Orchestrator route proposal -> WorldState / TemporalQuery / Local DRS -> CandidateVectors -> AVF / HardMask / SoftMask -> Root-controlled route acceptance -> BSEP creation and validation -> side-specific BSEP projections -> Semantic Architect proposal -> local validation -> runtime-owned RuntimeExecutionTopology -> bounded actors, executors and child cells -> ResultProposal, receipts and boundary snapshots -> Post V&V -> terminal GT advisory -> independent local Root decisions -> optional current ActionCommitPacket -> exclusive Effect Firewall / bounded Corridor -> EvidenceReceipt -> transaction ledger, cryptographic seal and offline replay -> local Root-controlled DRS outcome writeback. This defines roles and relationships. It does not require every profile to execute an identical linear sequence. BSEP is the Bounded Semantic Evidence Packet: the canonical semantic membrane between accepted route context and side-specific semantic work. It carries structured, bounded semantic evidence rather than the complete conversation. It is neither truth nor consent, FinalOutput or ActionCommitPacket. In hedgehog/context_packets.py, build_bounded_semantic_evidence_packet carries source, role, route and proposal references; observed facts; missing evidence; uncertainty; rejected routes; required approvals and guards; and explicit non-authority flags. The MODEL participant, its returned PROPOSAL and its BSEP input are distinct objects. Work is bounded typed computation with concrete input, sources, scope, budget, result and consumption. The implementation in hedgehog/kernel/work_composition_v01.py includes WorkOutputBindingV01, WorkHistoricalOutputV01, WorkProgramCandidateV01, materialize_work_program_v01 and advance_work_program_v01. Runtime materializes and validates the topology. The Semantic Architect proposes semantic structure; a model proposal creates neither Root authority nor permission to execute an effect. A model can nevertheless participate in admitted computation as a distinct semantic role. CausalConsumptionRefV01 in hedgehog/kernel/abi_v01.py binds an actual producer field to a consumer and disposition. Naming an upstream program or presenting a hash does not establish actual input binding. Root is the sole local final and commit authority. Client, airline and bank Roots, or VenueRoot and ActivePlanetRoot, belong to independent owners rather than successive levels of one business hierarchy. The implementation in hedgehog/kernel/root_decision_v01.py defines RootDecisionKernelV01, RootDecisionInputV01, RootDecisionResultV01 and decide_root_v01. Route acceptance, result acceptance and action permission are separate decisions. An informational ACCEPT may request no effect and create no permission. Root is an architectural role; bounded internal work does not inherently require a separate human click for every operation. Host is the concrete RootWorkExecutionHostV01 in hedgehog/work_execution_host_v01.py. It is distinct from Root. Section 7 of docs/common_action_and_dynamic_composition_contract_v01.md defines one in-process serialized current-state owner per Root. It holds the registry, current dependency observations, admitted catalogue, logical time bridge and monotone state_revision. A caller supplies identifiers and an expected revision, not a preferred historical registry. The ordering is expected revision -> source refresh -> currentness and Root checks -> logical-key reservation -> STARTED -> actual invocation -> validation -> history. STARTED is the invocation linearization point within this in-process profile. This does not establish distributed crash recovery, physical rollback, external attestation or a hard real-time guarantee. ActionCommitPacket carries exact current scoped Root permission, bound to the decision, owner, transaction, action, adapter, material, time and idempotency state. Scope and TTL may narrow after Root review but may not expand. The Effect Firewall is the exclusive bounded effect-handle owner; its implementation is hedgehog/kernel/effect_firewall_v01.py. A Corridor validates and carries only the authorized operation and creates no permission of its own. EvidenceReceipt records the effect result under the declared trust model, without authorizing repetition. GT, Post V&V, DRS, signatures, replay, child Work and foreign Roots cannot bypass this boundary. 3.2. Nested computation, continuation and memory Nested computation does not automatically create a new Root or require one Root per child. A parent invokes bounded child Work and consumes its exact result. The child remains constrained by its contract, scope and budget. Adding a node does not delegate final authority. Independent Roots follow actual ownership boundaries, not graph depth. A scoped relationship from owner A to B does not permit B to transfer A's rights to C. Two distinct mechanisms are implemented: - Recursive review and child accounting, including retained child results. A parent can consume an unchanged historical child result together with a recomputed selected child result in canonical parent-slot order. - A missing pure capability observed inside Work can lead to local DRS discovery, current eligibility checks, Root-reviewed descent, independent readmission, installation, use and continuation. Relevant symbols are observe_missing_pure_need_v01, execute_work_task_review_v01 and build_work_historical_output_v01 in hedgehog/kernel/work_composition_v01.py; and search_pure_need_v01, retrieve_pure_memory_v01 and validate_pure_memory_retrieval_v01 in hedgehog/capability_memory_binding_v01.py. The implementation checkpoint is docs/common_action_and_dynamic_composition_checkpoint_v01.md. The precise evidence classification is in section S11 of docs/showcase/radiolaria_os_v01/references/experience/source_ledger.md. These records do not demonstrate arbitrary autonomous child-local DRS search at every nesting depth. Finite pure-i32/Wasmtime admission without world I/O does not establish isolation for arbitrary native plugins. Every current retrieval requires a TemporalQuery. schemas/temporal_query.schema.json requires as_of and freshness_bias and distinguishes historical_as_of, prefer_recent, strict_current and ignore_time_only_if_policy_allows. Profile observation time, receipt or ingestion time, logical time, validity windows and retrieval time must remain separate. Historical bytes, provenance and previous validation results persist while current applicability can change. Preservation of history does not itself prove that the stored content is true. Reading an old object again does not renew its TTL. A Host capture retains genuine origin within its Host; portable JSON cannot reconstruct that origin. See the Atomic Current Source Capture and T3 V04 clauses in docs/common_action_and_dynamic_composition_contract_v01.md and ActionSourceCaptureV01 in hedgehog/work_execution_host_v01.py. Local DRS provides pointer-first discovery and controlled access to material with provenance, lineage, scope, policy, time and lifecycle metadata. An address, payload, right to read and permission to act are distinct. Section 17 of docs/drs_semantic_address_space_reuse_certificate_g2_b_preflight_v01.md requires hard eligibility checks before ranking. An ineligible high-scoring candidate cannot become a fallback. Deadend, conflict and quarantine evidence applies to the selected record and Root-approved bounded lineage or conflict reads, not a global taint label. DEADEND can also be a terminal no-progress outcome in the bounded runtime implemented by hedgehog/kernel/fractal_runtime_v02.py. Negative experience can separately recommend more scrutiny. Neither mechanism implies model-weight training or a permanent prohibition on unrelated domain work. Section 21 of the same DRS contract defines controlled descent classes: SUMMARY_ONLY, OPEN_ONE_ARTIFACT, OPEN_LINEAGE_NEIGHBORHOOD, OPEN_CONFLICT_SET, OPEN_DEADEND_PROOF and OPEN_FULL_TRACE. Root approves an equal or narrower class and budget. SUMMARY_ONLY opens no payload. Retrieval does not mutate the past; a new Root-reviewed writeback or supersession leaves the predecessor byte-identical, as required by section 29. A ReuseCertificate is currently bound evidence supporting an informational shortcut, not code installation or an action grant. N4-04 records domain and Host Work calls changing from 1 to 0, with one warm stored read and two current Root calls. It does not establish a speedup. Those counts cover the current process and thread only; child activity is UNOBSERVED_NOT_ZERO. External DRS supports exchange between owners through closed source-bound profiles. Signed metadata and fetched body are separate. The receiver checks body, source, release, status, time, scope and policy and decides local applicability. See docs/gate5_reference_contract_v01.md, hedgehog/external_drs/gate5_exchange_v01.py and hedgehog/external_drs/gate5_lifecycle_v01.py. This is bounded external meaning exchange, not global federation, discovery, PKI or a foreign plugin installer. In Football, E is the producer offer; B contains the offer plus Work, review and temporal material; C contains B plus independently constructed currentness. The checked relationships are B.offer = E and C.body = B. E, B and C are different objects rather than identical serialized bytes. REVOKED stops consumption; a valid signature on revocation does not make the source usable. 3.3. GT, AVF and numerical meaning Terminal GT advisory input to Root is distinct from OUTCOME, the Outcome Feedback Envelope (OFE), retained history and current ADVICE. OFE separates advice quality, enforcement and business outcome. UNSAFE advice combined with BLOCKED_AS_REQUIRED enforcement and SAFE_NO_DEAL is not a successful payment. An absent prospective expectation produces NOT_SCORABLE or NO_UPDATE, not an invented zero outcome. The G35 record contains 14 unique occurrences and zero scorable observations. In G36R Supplier, one eligible negative occurrence produces sparse history, changes the selected review from standard to provenance and leads to actual supplier.review_provenance.v01 Work. A separate lawful continuation has one mock effect. The LIVE_OBSERVATION origin label combined with CAPTURED_REEXECUTION consumption does not signify a new provider call. The fixed-point scale is Q = 10^9. RHE denotes signed exact rational rounding to the nearest integer, with ties to even. In hedgehog/outcome_calibration_v01.py, the rating update is: R_next = clamp(R + RHE(125000000 * (O - E) / Q), 0, Q) E is the prospective expectation and is not replaced by the preceding rating R. The signed prior update is: P_next = clamp(P + RHE(62500000 * ((2 * O - Q) - P) / Q), -Q, Q) The original occurrences must be deduplicated, their full subject keys must be comparable, time must be explicit and the fold is capped at 64 effective events. WARM_SAMPLE_COUNT = 3 is a counting state, not statistical confidence or a warm cache. Advisory decay is: T = RHE(R * 2^(-age / half_life)) It follows separate time admission and cannot renew TTL. Unusable trust remains distinct from usable trust that has decayed to zero. The contract is docs/gate3_outcome_feedback_contract_v01.md. Source and check routes are recorded in sections S02-S06 of docs/showcase/radiolaria_os_v01/references/experience/source_ledger.md. G4 pressure and allocation are implemented by evaluate_reference_pressure_v01 and allocate_reference_work_budget_v01 in hedgehog/gate4_pressure_budget_v01.py. Hard-FALSE branches are excluded; unresolved required evidence blocks allocation. The pressure and integer weight are: z_i = RHE((4 * r_i + 2 * l_i - 2 * u_i - c_i + p_i) / 8) w_i = RHE(10^12 * exp((z_i - max(z)) / 250000000)) The features denote relevance, lineage, uncertainty, estimated cost and prior. The raw prior is applied once. Lower-first capped apportionment uses exact fractions, largest remainders and stable identifiers to preserve the budget. In the recorded Airline case, the total budget is 9: one unit already spent, one reserved for the final consumer, and seven allocated. The allocation changes from (4, 3) to (3, 4) and is reflected in performed Work; total expenditure remains 9. ONE_NATIVE_WORK_DISPATCH_UNIT is neither a token, monetary or elapsed-time unit nor a probability of safety. hedgehog/gate4_strategy_reference_v01.py computes finite hard feasibility, per-owner utility and disagreement values, individual rationality, the Pareto frontier and bounded product ranking. Client, airline and bank still decide independently. FullAVF, StrongGT, regret and stability, robust/minimax/CVaR/sensitivity methods, and LGT remain deferred rather than implemented with zero-valued results. Wedding has a different mathematical profile. It uses 36 binary guest/table indicators and a nonnegative integer hard penalty H. H = 0 means that the original conditions hold within the finite profile. KEEP and MIX are distinct objectives: Q(x) = P * H(x) + selected_soft_cost(x) P_KEEP = 25 P_MIX = 13 This profile's Q(x) is an optimization objective, distinct from the fixed-point scale Q above. The local ten-qubit PAIR_BINARY_10Q_V02 encoding is an exact substitution with up-to-four-local terms, not a ten-variable QUBO. A solver searches for candidates; an original-condition CHECKER evaluates them. That domain checker is not automatically the kernel's Post V&V entity. A QPU returns samples; successful transport does not establish sample validity. Relevant symbols are ProblemMathV01, compile_optimization_v01 and validate_assignment_v01 in hedgehog/domains/wedding_seating/math_v01.py, and samples_v01 in hedgehog/domains/wedding_seating/qpu_contracts_v01.py. The separate calibration example uses readings [8, 10, 9] and reference 10, producing correction = (3 * 10 - 27) / 3 = 1/1 in abstract input units. No physical unit is assigned. 3.4. Demonstration roles and limits The chapter paths below are relative to docs/showcase/radiolaria_os_v01/. Wedding, chapters/08_wedding.md: actual semantic fields change compiler inputs and are recorded as USED at /material. KEEP and MIX change seating under the same hard constraints. SR2 replaces two added placement conditions; SR3 performs validate and consume without a new solve. A and SR1 use captured provider response reexecution; B, VERIFY, AMBIGUOUS, SR2 and SR3 retain LIVE_ROLE_ORIGIN. The original SR1 attempt-015 remains FAILED with reason collection_bound. SR2 consent is SCRIPTED_CONTROLLED_OWNER_INPUT with browser_click_claimed = false. W4 is a distinct revision-1 QPU episode. KEEP yields 33 valid and 967 invalid samples out of 1,000; selected shot 63, basis 517, is checked and consumed with local_repair = false. Saving has separate current Root permission. The record establishes neither quantum advantage nor universal synthesis. Equal seating arrays do not transfer SR2's problem identity to W4. Football, chapters/07_football.md: EXACT produces no complete offer. A permitted thirty-minute shift on January 8 alone produces three natural-grass sessions for EUR 63 under a EUR 70 cap. The case connects actual producer output, released external material and current consumer input across independent VenueRoot and ActivePlanetRoot decisions. SHIFT ends OFFER_READY with zero effects. A separate BOOK produces one mock batch effect; REPEAT and VERIFY do not repeat it. The adverse REVOKED neighbor has status_not_active and native BLOCKED_FAIL_CLOSED / hard_identity_violation, without a successful consumer or effect. This is G54D1 candidate p7_candidate_f779b790, not the later G6A Football rerun. The publication's data-only cross-check did not repeat signature verification. Ephemeral Workspace (EWS), chapters/09_workspace.md: one persistent Host and local Root organize a finite service environment. The principal EWS3R2 semantic roles are CONTROLLED_DETERMINISTIC; browser interactions are SCRIPTED_BROWSER_TEST. The audio process is killed with return code -9. The expired old packet is refused; fresh source evidence supports selected media recomputation. Frames increase from 4 to 16, PCM samples remain at 16,000, the photo snapshot is retained and its producer count stays at 3. This is selective recomputation, not zero work. REQUEST_SAVE does not write. Separately approved exact bytes and a current packet produce a 250-byte selection.json sidecar, not a modified photograph. Cleanup concerns owned resources and cache. Audio reports ALREADY_EXITED with receipt = null, not a graceful-close acknowledgement. The historical EWS3 memory and live/captured lane is separate: its rating 4, ORIGINAL crop and exposure 0 must not be conflated with the principal rating 5, SQUARE crop and exposure 5. Sentinel, chapters/10_cross_case.md: the cases distinguish known, missing, late and correlated observations. In LS2R2, the intended LOCAL_SLOPE_ANALYST role actually uses CLOUD_LLM, not an installed local SLM. A BOUNDED_CONSISTENT reserve series is not a landslide prediction or physical certification; physical_mapping can remain NOT_ESTABLISHED. A separate Atlas CONTROLLED_BARRIER_STUB demonstrates useful local continuation while an optional worker is pending; the late answer is refused for role_current_context. Three mock effects cover ON, report queue and OFF, with outbox_deliver = 0. This is neither a cloud-timeout benchmark nor a hard real-time safety guarantee. Historical VALID_LIVE_CONTRAST_NOT_DEMONSTRATED and NOT_DEMONSTRATED statuses remain visible. G3/G4, chapters/05_experience.md: qualified experience changes review selection and finite budget allocation. This is not fine-tuning, global reputation or automatic trust. The G36 Supplier refusal, G4 Airline predictive experiment, N4-04 calibration reuse and N4-05 two-record policy example have separate histories and clocks. They are not one composite run. Incident-to-Proof Atlas, chapters/06_method.md and chapters/10_cross_case.md: this is finite risk/control evidence paired with lawful neighbors, separate from the Geometry Atlas. DIRECT_CAUSAL, SHARED_DOMAIN_PROOF and POST_CONSUMER_CONTROL have different causal scope. Twenty cards represent five consumers and seven effective history samples, not twenty independent learners. Supplier S1 stops at ROOT with a null Firewall field; it is not a Firewall rejection. Catalog E01-E15 contains incident, experiment, requirement and project-context entries, not fifteen external incidents. The Australian account is additional context, not the Atlas's origin or a reproduced attack. Saved replay is SAFE_DERIVED_SAVED_RELATIONS_NOT_LIVE_AUTHORITY; native graph replay remains UNSUPPORTED_NATIVE_SCHEMA. 3.5. Runtime participation and system development The Wedding model participates at runtime by returning meaning that changes actual Work. External domain development is a separate process: an author receives HOW, including docs/gate5_authoring_kit_v01/START_HERE.md and docs/gate5_authoring_kit_v01/EXECUTION_CONTRACT.md, plus a separate WHAT task; produces a candidate; and undergoes examination with an independent oracle and observed native execution. Exact source bytes then undergo separate admission. Providing an Audit Reader to a cold author as though it were a neutral kit changes the experiment. Build-time admission neither grants installed code current business permission nor substitutes for a runtime Root decision. G54B1 and G54D remain historically INCOMPLETE. The accepted G54D1 continuation includes an explicitly authorized fourth correction; it is not a one-shot cold success. One successful authoring case does not implement an integrated Capability Synthesis Factory / NeedleFactory or the deferred reflective and cross-domain programs. Approved deferments N1-18, N1-19, N4-12 and N4-13 remain. 3.6. Seven Geometry Atlas views and their defining files The base directory is docs/showcase/radiolaria_os_v01/atlas/. - OVERVIEW.md and OVERVIEW.svg provide the common overview. - views.json defines view identifiers, titles, entities, relations, questions and source_refs, with id = GEO-N. - native_views.json defines native slide objects, reading order and positions, with slides[].id = GEO-N. - ENTITIES.md and entity_cards.json define shared editorial identities. These aliases do not create a runtime ABI. The seven views are: GEO-1: Intent becomes bounded, consuming Work. Files: GEO-1.md and GEO-1.svg in the base directory. Subject: materialization and typed nested consumption. This is a conceptual view. GEO-2: Independent owners retain their own decisions. Files: GEO-2.md and GEO-2.svg. Subject: ownership, review and effect boundaries. This is a conceptual pattern with a recorded Football inset. GEO-3: Useful work can outlive a changed resource. Files: GEO-3.md and GEO-3.svg. Subject: separate environment, Work, permission and result lifetimes. It uses the recorded principal EWS3R2 lane; historical memory is a separate lane. GEO-4: A source body becomes checked local input. Files: GEO-4.md and GEO-4.svg. Rendered headline: Qualified context reaches Work; refusal stops. Subject: address, material, current status and consumption. It uses recorded Football E/B/C relationships; no arrow leads from refusal to Work. GEO-5: Different computations share checked consumption. Files: GEO-5.md and GEO-5.svg. Subject: semantic input, compiler, QPU, decoder, checker, native consumer and save, with a separate local EWS inset. The principal record is W4 revision 1. GEO-6: Recorded experience changes finite expenditure. Files: GEO-6.md and GEO-6.svg. Subject: source-bound outcome, OFE, DRS, current advice, pressure, allocation and Work. It explains G3/G4 mechanisms; N4-04 is a separate witness. GEO-7: Creating a component and using it are separate acts. Files: GEO-7.md and GEO-7.svg. Rendered headline: Creating a component does not grant its authority. Subject: HOW, WHAT, author, candidate, examination, admission and new Work. This conceptual view preserves the G54 correction lineage. The base directory's README.md lists the catalog titles. Historical ATLAS_A, ATLAS_B and ATLAS_C retain their pilot identities and do not replace the seven-view Atlas. DATA, PROPOSAL, PERMISSION, EXECUTION, DERIVATION, DOCUMENT_REFERENCE, CHRONOLOGY and CONCEPTUAL_CONTEXT must be read from relation labels, not inferred from color. TIME is distinct from MEMORY; GT from OUTCOME, OFE and ADVICE; SOLVER from CHECKER; and VISUAL_SEAL from CRYPTO_SEAL. The B1R3 Visual Seal is a NONCANONICAL prototype without an established receipt-to-image binding. It proves neither origin, truth, currentness nor permission. 3.7. Evidence classes and verification scope Five classes must remain distinct: accepted contract, recorded execution, independent checking under a declared trust scope, architectural inference and proposed future work. Publication field, hash, link and format checks are not fresh runtime acceptance. The reference matrix records 69 satisfied obligations out of 73, plus four approved deferments. TWO_FRESH_NINE_RETAINED does not mean eleven fresh profiles executed on one final source freeze. G6A_TECHNICAL_READY does not by itself close the entire Gate 6 roadmap. Controlled, captured, historical live-origin and mock-effect provenance must remain visible. A native llm = 0 count does not imply zero provider calls during earlier capture acquisition. A caller's or candidate's report cannot serve as its own independent oracle. Observer, Host, Firewall and reviewed reference code may share a process under the declared trusted computing base. This is not external attestation, and the audit boundary must not be moved after a finding merely to exclude it. INSPECT_SAVED reads retained material. VERIFY_SAVED starts supplied workers and writes checking results without producing new domain Work. RUN_DETERMINISTIC_REFERENCE executes the documented native composition with current inputs. RUN_LIVE_OPTIONAL additionally requires its own provider, model and budget context. Payload hashes, manifest consistency and signatures establish different relationships. A digest asserted by a package is not an external trust anchor. Reading the publication does not produce a new audit or runtime result. The historical web and regulatory statements in chapter 12 were not updated for this adaptation and do not constitute a current legal assessment.
04. NEW-DOMAIN AND AUTHORSHIP METHODOLOGY
======================================================================
04. NEW-DOMAIN AND AUTHORSHIP METHODOLOGY
======================================================================
This section provides a technical route grounded in the accepted source, frozen authoring kit and recorded independent examination. Football field names, package limits and sandbox settings describe a finite historical profile. Their applicability to another domain must be decided from that domain's declared contract. Source paths in this section are relative to the public repository.
4.1. Starting a new domain
A new domain is a bounded subject-specific implementation on accepted interfaces. It does not create a new source of authority. Before implementation, freeze the exact inputs and normalization rules, hard constraints and ranking order, permitted operations, applicable closed E/B/C types, owner roles, current permissions and refusal policy, time axes, public and private projections, source dependencies, finite budgets, evidence classes and limits of the intended claim. A model may propose parameters or code. It cannot supply owner consent, accept its own implementation, issue a Root decision or perform an effect merely by proposing it.
The HOW and WHAT documents serve different purposes. HOW defines the permitted engineering route and authority boundaries; the task-specific WHAT defines the bounded problem, schemas, behavior and acceptance obligations. Both must be fixed for a measured authoring trial. A domain-specific WHAT must not silently expand the authority or execution surface defined by HOW.
Keep the three decisions in HOW_TO_BUILD.md section 1 separate:
- Independent source, import and I/O review of the new implementation.
- Native capability admission of the specific executors and validators.
- Current local Root and Host authorization of a particular action.
Candidate-owned tests, another Root's signature and the package checker do not substitute for these decisions. Root remains the sole local final/commit authority. Host binds the current source, code and dispatch. The Effect Firewall and bounded Corridor remain the exclusive effect path. History, DRS, GT, replay, captures and receipts do not restore authority.
The practical sequence is: freeze the task, profile and source basis; implement pure domain functions and independent predicates; use actual typed Work and applicable Root reviews; for a domain exchanging external evidence, implement closed publication and verified import/consumption; where an effect is required, use the existing MOCK action path; preserve the actual receipt and independent readback; define and check the finite positive, negative and continuation matrix; complete source-bound independent verification; and perform source admission or landing only in its separately authorized scope. A purely local domain does not need external publication merely to fit this sequence. There is no universal self-installing domain factory API.
Primary references are docs/gate5_authoring_kit_v01/HOW_TO_BUILD.md, CANDIDATE_OUTPUT.md, SHARED_EVIDENCE_CONTRACT.md, and the Scope, Exact Engineering Interfaces, and Proof and Trust sections of docs/gate5_reference_contract_v01.md. The original K3 integrated natural-language-to-requirements-to-code-to-sandbox Capability Synthesis Factory, reflection/transfer and global federation remain deferred. Measured external authoring does not implement those deferred components.
4.2. Exact extension interfaces
Use the exact signatures in docs/gate5_authoring_kit_v01/NATIVE_INTERFACE_MAP.md together with the pinned source bodies. The map is a historical extracted signature index. A private symbol appearing in source or an index is not thereby an approved extension point.
The actual football PURE Work pattern is implemented in hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/native.py::perform:
- A module-level executor receives BoundCapabilityInvocationV01 and returns a tuple of ActionEffectParameterRecordV01 records built with action.build_action_effect_parameter_record_v01(parameter_name=..., value_type=..., value=...). INTEGER, TEXT, REFERENCE and BOOLEAN are exact field types. A bool is not an INTEGER.
- mocks.admit_pure_operation_v01(*, operation_id, input_fields, output_fields, executor, output_validator, host_instance_ref) admits the controlled pure operation. This finite donor does not certify arbitrary Python or provide an operating-system sandbox.
- work.WorkItemV01(work_id, definition_id, owning_root_id, inputs, resource_refs, depends_on, guard, review_obligation_id) defines the actual Work item. donor.work_literal_v01(name, value_type, value) supplies a literal. Actual upstream output binding uses WorkInputBindingV01(input_field, source=WorkOutputBindingV01(predecessor_work_id, output_field, expected_type)); a caption or copied constant is not an upstream binding.
- donor.prepare_work_program_v01(*, task_id, root_id, admitted, items, device_refs) returns program, common and host.
- work.advance_work_program_v01(program, *, catalogue, source_context, semantic_proposal, host_map, results=(), review_bindings=(), action_authorizers=None, continuation_context=None) advances the real program.
- validate_work_program_result_v01(program, results, *, catalogue, host_map, source_context, semantic_proposal, review_bindings=(), continuation_context=None) validates its result. work_program_result_to_artifact_v01 has the same arguments and produces the bound artifact.
- Preserve native_evidence={admission, invocation, result}, exact inputs and outputs, actual Host attempts, results and KernelArtifact. A function named Work or a self-reported PASS is insufficient.
Football's source capability consumes TEXT request_json and inventory_json and returns TEXT offer_json=canonical(E). Its requester consumes TEXT request_json and offer_json=canonical(C), and returns TEXT result_json=canonical(report without native_refs). These are football profile fields. They are not universal kernel input names. Another domain defines and reviews its own closed schema and bindings.
gate5_native_v01.native_work(mode, root, task, values) supports only the source, mean and corrected CALIBRATION donor modes. It is not an arbitrary-domain dispatcher. Football uses the lower Work/capability interfaces; a new mode or unrelated schema must not be inserted into calibration fields without an independently reviewed interface change. donor.prepare_work_program_v01 is also controlled donor infrastructure, not universal production provisioning.
The controlled Root review helper is:
gate5_native_v01.root_review(root, transaction, candidate, subject, checks,
claim_value, now, predicate='g51_bounded_context_review', window=None)
It builds the actual semantic request, contribution, review packet and RootDecisionInputV01, calls decide_root_v01 and validates the result. checks must contain independently established actual bool predicates. The claim must bind the exact consumed data, artifacts and dependencies. now is a recorded integer UTC value; the helper checks the actual wall clock against the half-open window. A stored root_plain is evidence, not a renewed grant. Source, publication, release, retrieval, import and result reviews have distinct subjects.
The reviewed closed-body extension is:
ReviewedBodyProfileV01(schema_id, body_fields, scope_fields,
policy_accept_field, validate_body, validate_pointer, validate_binding)
The reviewed local module-level callbacks validate_body(body), validate_pointer(pointer), and validate_binding(body, policy, pointer) raise specific ValueError failures. Trusted local code selects the profile explicitly for:
validate(kind, value, *, profile=None)
check_pointer(envelope, policy, now, *, profile=None)
check_bundle(bundle, pointer_envelope, request, status, policy, now,
history, conflicts, *, profile=None)
The profile is not serialized into a payload and does not load code from wire data. Unknown schemas fail. The default remains strict BoundedCalibrationSummaryV01. contract(kind, value)/BoundedContractV01 and ImportStore.validate_bundle remain calibration-only. Football calls the public check_bundle with profile=profile() and uses ImportStore for observation, highwater, conflict, deduplication and quarantine. Global BODY_SCHEMA replacement or monkeypatching is not the extension route.
The DRS sequence uses a fresh LocalDRS, write_semantic_record(drs, SemanticDRSRecordInput), resolve_semantic_candidates(drs, query, *, layers=('work', 'thoughts', 'quarantine', 'deadends')), the exact SemanticAddressV01, DRSTimeEnvelopeV01 and authority-envelope builders, evaluate_drs_candidate_v01, build_memory_descent_budget_v01, build_retrieval_plan_v01, a genuine local Root review, build_memory_descent_request_v01, and execute_local_memory_descent_v01.
Football candidate_domain/memory.py::descent uses SUMMARY_ONLY with max_depth=0, max_records_opened=1 and all other opening, byte and lineage limits zero. CONNECTOR_OBSERVATION is ineligible. A subsequent ROOT_ACCEPTED_CONTEXT belongs to the local Root; foreign source IDs remain source references. The exact descent review predicate is approve_controlled_memory_descent_plan_v01, with claim_id=plan.retrieval_plan_id, subject=plan.semantic_address_id, object_or_value equal to the exact plan projection, and authority_class=NONE. Its omission caused a real correction1 failure. The remedy was to satisfy the contract, not weaken the kernel check. Gate5's pointer_descent_v01(pointer, folder, source_end) shows the calibration sequence but contains calibrated identities; another domain needs a thin reviewed adapter.
4.3. E -> B -> C and an independent oracle
For external evidence exchange, the shared contract is DomainResult E -> PublishedEvidence B -> ConsumptionContext C. E must come from real typed producer Work evaluated against independently retained original inputs. B is built after that Work and its source review. The profile defines payload_projection(B)==E and the relation to the original inputs. B preserves the original source record and revision, actual Work and source-review references, original time and lineage, and request scope.
The manifest covers an already existing body. The pointer binds the body and manifest. Publication and release have separate Root reviews. Signatures cover exact canonical bytes and are verified against independently pinned public keys. Avoid hash/self-reference cycles. A Work input must not contain the future Work or review ID that this execution is supposed to create.
C is created only after checks of the exact canonical B, manifest and pointer, signatures, request/audience/use binding, release, STATUS, local policy, time, highwater, conflicts and history. It is derived from B, authenticated STATUS and independently held local request, policy and use_time. Signed originals remain separately preserved. The consumer Work actually consumes C, and the final local Root checks that output and its original dependencies. This publication/import route is conditional on the domain exchanging external evidence; it is not mandatory infrastructure for every pure local calculation.
The exact football mapping is defined by FOOTBALL_PROFILE.md:
- E has exactly request, fields, alternatives, slots, total_minor, currency, status, schedule_revision and offer_revision.
- B.offer=E; B.request_sha256=sha(E.request); B.source_revision=E.offer_revision; E.schedule_revision equals the original inventory.revision.
- C is exactly {body:B, source_projection:S}.
- S takes publisher and recipient from B; request_revision from E.request.revision checked against the original request; schedule_revision from E checked against the original inventory; offer_revision from B.source_revision; observed_at, valid_from, valid_to, ttl and pt from the original time envelope; status from the verified STATUS entry; and status_checked/status_until from the verified STATUS value.
- The final claim is exactly {request_sha256:sha(request), offer_sha256:sha(B), result_sha256:sha(report_without_native_refs)} and selects the actual requester Work artifact. sha(B) is neither sha(E) nor sha(C).
The independent oracle derives expected producer input and expected C from trusted original inputs and observed wire data, policy, keys and time. Candidate proof.inputs and a candidate success flag are not oracle inputs. gate5_supplied_v01.native_result(proof, expected_root, expected_inputs) validates the admission, invocation and result structures and their joins. It does not by itself prove execution origin. Bind these checks to source-pinned passive observation of actual returns and the source/import hashes.
Independently check E's selection, arithmetic and no-result behavior against original facts; check payload_projection(B)==E; and join actual source, publication and release records. Action checks additionally join the actual packet and current Host to Firewall entry, executor, receipt and independent state readback. A signature over incorrect arithmetic does not make the arithmetic correct. A correct consumer cannot repair an absent or incorrect producer E.
A valid public bundle check is insufficient to establish the actual release Root identity. The separate required join is:
BUNDLE.release.value.release_review_ref
== actual_release_root_result.decision_id
Public recorded evidence is indexed by docs/gate5_reference_evidence_v01/member_index.json, including members.g54d1, with pinned archive bytes in material.tar.gz. That evidence records independent checks of frozen original inputs, source/import pins, original signed wire, public check_bundle calls, observed producer and consumer Work, E semantics and ranking, independently constructed C, actual Root joins, action stage order, payload and idempotency identity, current consent, receipt returns, source readback and historical lineage. An implementation or future review must not depend on obtaining a private return directory or private examiner answer table.
An independent oracle can be a separately authored finite enumeration. The historical football examiner's bounded selection search was at most 8**3; it was reviewer-owned and did not return answers to the author. Historical representation differences, such as uppercase vocabulary or counter layout, require a source-reviewed projection rather than a new author-facing ABI. candidate_domain.domain.verify_producer shares helpers with the producer and is therefore not the independent examiner. New reviews must retain independent original facts and semantic predicates even when a local output validator is present.
4.4. Actual Root refusal, REVOKED and time
Public report status REFUSED is not a Root decision enum. The accepted D1 correction4 repaired the status_not_active path by obtaining a genuine new local Root BLOCKED_FAIL_CLOSED. It did not relabel a preceding metadata ACCEPT.
In run.py::requester_event, ImportStore.observe first authenticates and persists the terminal highwater entry. The public check_status then refuses current use. refuse_inactive_status revalidates the retained pointer and STATUS at their separately recorded observation times, checks REVOKED/SUPERSEDED and the matching persisted entry, computes source_active=False, and builds a claim binding the full original request and hash, actual reason, source record/revision/review, pointer, prepared-but-unsent FETCH, STATUS hash and entry, observed_at and evidence hash. The predicate is football_current_source_status_v01 and the actual result must be BLOCKED_FAIL_CLOSED. report.native_refs links refusal_review and status_refusal_evidence; refusal.json contains the actual decision_id.
No FETCH, consumer Work or local effect follows this terminal branch. A preceding retrieval ACCEPT concerns a different subject. The evidence must demonstrate the new actual Root call, not merely a JSON object claiming refusal. A confirm event may already have caused a separately authorized source effect before the requester receives metadata. In that case the requester reports CURRENT_STATUS_UNKNOWN, preserves the receipt/readback path and reconciles; it cannot claim that the source had zero effects. Unknown failures remain observable and are re-raised. Uncertainty after execution starts is not a clean REFUSED outcome.
Source currentness requires:
valid_from <= t < min(valid_to,
pt_created_at + ttl_seconds,
source_observed_at + local_max_source_age)
source_observed_at <= t
STATUS currentness requires:
checked_at <= t < min(valid_until, checked_at + 60)
effective_at <= t
It also requires exact pointer, stream, publisher, request, revision and nonce binding. A later ingestion, check or signature cannot renew the source lifetime. The native donor clock 1010..1014, including Work at 1014, is explicitly logical. Domain UTC observation and review times are a separate axis. Wall time must not be substituted into one donor axis in isolation.
Persist highwater by publisher, pointer_id and stream_id. A lower revision is rollback. The same revision with a different immutable entry hash is equivocation. The same entry with a refreshed freshness envelope is not a conflict. Terminal REVOKED/SUPERSEDED cannot become ACTIVE through replay or restart. Distinct streams remain independent. Unavailable current status means UNKNOWN, not a permanent blacklist. Preserve history; changed policy or dependencies require the appropriate new current review. Duplicate evidence earns no independent-source or reputation credit.
4.5. Action lifecycle and authoring states
The existing native transition rows are defined in hedgehog/kernel/transition_registry_v01.py:
g2a_t01_activate_root_authorization: CREATED to ROOT_AUTHORIZED
g2a_t02_queue: ROOT_AUTHORIZED to QUEUED
g2a_t03_pending: QUEUED to PENDING_FULFILLMENT
g2a_t04_fulfill_mock: PENDING_FULFILLMENT to FULFILLED_MOCK
g2a_t05_receipt: FULFILLED_MOCK to RECEIPT_RECEIVED
FULFILLED_MOCK and RECEIPT_RECEIVED have CONSUMED idempotency disposition. Blocked, expired, revoked, superseded, nonconsuming and uncertain branches are represented by the existing transitions. A domain must not recreate this state machine or invent effect kinds.
The actual football path is in effect.py::admit/book/current_checks/execute_booking/validate_output/reconcile. It observes the module-level input validator, output validator and executor with observe_local_capability_code_v01; constructs build_capability_business_input_binding_v01 and build_capability_business_semantics_v01; constructs build_capability_definition_v01(effect_kind='MOCK_CONSEQUENTIAL'); and calls admit_local_capability_v01.
In this profile, operation_key=logical_effect_namespace='football.book.v01'. selected_action_class begins with 'mock_action:'. Typed device_ref and payload_json are exhaustive canonical bindings. Violation of the operation-key/namespace equality caused a correction2 failure. PURE, lookup and semantic actors do not directly write business state.
The action sequence is:
typed inputs
donor.build_native_work_action_v01(*, admitted, inputs, root_id,
transaction_id, business_object_ref)
authorize_and_prepare_action_v01(canonical, admitted, inputs)
host_for_prepared_action_v01(prepared)
hosts.dispatch_current_action_v01(host, packet_id=..., task_id=...,
expected_revision=host.revision, evaluation_time=1014,
evaluation_time_source='u1.controlled_utc',
evaluation_context_id='context:u1:dispatch')
These clock arguments belong to the controlled donor profile, not a universal live-time convention. At the executor boundary, recheck consent, source currentness and status, exact inventory revision, tariff, occupancy, registry conflicts and payload. canonical.idempotency_identity supplies the logical key. A model-authored text key is not a substitute.
One disposable registry snapshot atomically contains the three booked slots and the idempotency entry. The native registry owns the real EvidenceReceipt. Save the actual receipt before optional rendering. The independently observed order is DISPATCH, EFFECT/Firewall, EXECUTOR, CALLBACK, STATE_WRITE, followed by callback, effect and dispatch returns. Bind outputs, receipt and readback to the actual source.
Count attempted Host dispatch, Firewall entry, executor start, mutation and returned result separately. A currentness refusal can have one attempted dispatch and zero executor starts or mutations. A failure after execution starts may have changed state. Preserve the actual stage, error, receipt and readback; report uncertainty and do not retry blindly.
An exact duplicate terminal native packet refuses. Host does not automatically issue a second successful receipt. Application-level repeat and VERIFY use the saved exact key, payload, packet and receipt plus actual current source readback as HISTORY. They perform zero new producer/consumer Work and no new action preparation or dispatch. An expired offer can therefore be returned as history without renewing permission.
Changed payload under the same key refuses. Missing, mismatched or uncertain history yields UNKNOWN_REQUIRES_READBACK or CURRENT_STATUS_UNKNOWN. The disposable os.replace example does not establish arbitrary power-loss recovery, hostile-storage safety, cross-owner atomicity, automatic rollback or compensation, real booking, or payment. Compensation may be NOT_APPLICABLE only with the concrete explanation that this is a single-snapshot mock.
The authoring states in HOW section 1 form a separate chain: DRAFT, CANDIDATE_READY, QUARANTINED_FOR_REVIEW, APPROVED_FOR_DECLARED_SANDBOX, followed by separately recorded LOCAL_REFERENCE_ADMITTED or REJECTED. Revocation applies only to a real prior admission. Historical inert-author status text in the installed football README/PASSPORT does not override subsequent independent acceptance and landing records.
4.6. Public/private inputs, cold authorship and ordinary development
The frozen cold-author protocol is specified by AUTHOR_SYSTEM.txt and AUTHOR_TOOLS.json. A fresh author receives only pinned public HOW, task WHAT and approved source-reader output. Prior chat, memory, target implementation, examiner, owner checkout and secrets are excluded. The inert broker permits reading/searching approved source, reading/writing/patching candidate files, parsing Python and JSON as data, and submitting a complete package. It does not give the author shell, import, test, native or additional-provider execution. The coordinator maintains the candidate manifest.
The original trial budget was an initial submission plus at most three specific feedback corrections. Architectural assistance must be recorded. A kit defect requires a newly versioned kit/trial, with the failed attempt retained. Actual D1 used an explicitly authorized correction4: the fifth submission in the restored lineage. The original initial-plus-three result remains INCOMPLETE. This is neither one-shot success nor success within the original correction budget, and it is not a new unseen cold trial.
The frozen no-environment author protocol does not prohibit ordinary authorized reviewed-reference development and testing. A normal domain project can use accepted current development and review practices within its scope. An independent cold-authorship claim requires a predeclared public/private input split, access record and correction provenance. The full reviewer return archive must not be silently supplied to a cold author. Private exam inputs, expected answers and runner telemetry remain reviewer-owned.
Count author-model calls, runtime semantic-model calls and deterministic/native Work separately. Actual live capture requires provider provenance and the exact fields consumed downstream; a label such as AUTHORIZED_RUNTIME_CAPTURE alone does not establish provenance. Consent remains owner-controlled input.
The historical football trial entrypoint is exactly:
run_peer_v01(*, role: str, input_data: dict,
state_root: pathlib.Path, output_root: pathlib.Path,
channel: PipeChannel, bootstrap: PipeChannel) -> dict | None
Roles are publisher and requester. The trusted parent precreates separate peers before importing candidate code. Input is {version:'g54.peer_input.v01', role, events, policy}, with at most eight events. The publisher receives inventory. Requester initialization receives no inventory, precomputed offer or expected answer; a genuine offer is obtained later through the checked protocol. Each peer owns its directories, private key, policy and Root. Public-key bootstrap uses a separately assigned channel and expected identities. Candidate code cannot fork, exec or create threads.
The historical P7 environment pinned Linux arm64, Python 3.14, jsonschema 4.26.0 and cryptography 48, with no network, read-only code and finite resources: 512 MiB aggregate, one CPU, 16 PIDs, 30/35 CPU seconds, 90 seconds wall time, 16 MiB per file and 64 MiB output. These are the reviewed-reference trial settings. They do not certify arbitrary hostile Python or require every local domain Host and observer to run as separate services.
docs/gate6_reference_v01/BOUNDARY_AND_FINDINGS.md explicitly treats a Host and observer sharing a process as a reviewed-reference trust assumption. It is not hostile in-process isolation and does not impose a new action-service requirement. Ordinary review must state the trusted source/observer boundary accurately; a new separate-process acceptance condition must not be retroactively invented for the accepted reference.
4.7. Packaging and finite verification
The Gate5 candidate package contract is a historical finite format, not the mandatory packaging format for all future projects. It permits at most 32 regular files, 1 MiB total and 256 KiB per file. Symlinks, hardlinks, source snapshots, .git, keys, caches, bundled dependencies and examiners are excluded. Required files are:
candidate.json
PASSPORT.json
README.md
dependencies.json
source_impact.json
candidate_domain/__init__.py
candidate_domain/run.py
candidate_domain/profile.py
schemas/body.schema.json
tests/test_domain.py
Additional files are limited to candidate_domain/, schemas/ and tests/. Paths must be safe canonical relative paths; the complete hash/byte inventory must be unique. candidate.json uses version g53.candidate.v01, a bounded candidate_id, installable_now=false, entrypoint=candidate_domain.run and files. candidate_contract.py only reads data and returns CANDIDATE_DATA_SHAPE_ONLY_NOT_EXECUTION_OR_ADMISSION. It does not import, execute or admit the package.
The exact required PASSPORT fields are needle_id, owner, version, purpose, domain, input_schema, output_schema, actor_roles, execution_modes, capabilities, forbidden_effects, permission_policy, risk_class, drs_scope, sealed_slot_refs, adapter_requirements, time_currentness_dependencies, failure_policy, compensation, audit_provenance, tests and compatibility. Do not fabricate a sealed slot; NOT_APPLICABLE with a reason is permitted. Dependencies enumerate exact local source/package versions and hashes plus I/O, subprocess and network use. Source impact enumerates new files, APIs and requested source changes. Protected edits and checker replacement are excluded. README maps actual evidence to substantive obligations. Private reviewer helper filenames or identifier recipes are not an author ABI.
Before verification, define a finite matrix and classify every row as fresh native, historical source-bound native, pure altered-evidence, static source or author-process evidence. Select the applicable rows below. Mark inapplicable external, action, history and hidden-trial branches NOT_APPLICABLE with a reason; do not build them merely to fill a template.
- A lawful baseline and an actually altered input on identical code. A hard constraint cannot be compensated by a better price or score.
- Missing or ambiguous required inputs and CLARIFY behavior; bool/integer and other type distinctions; unknown fields and schema limits; exact arithmetic and deterministic tie-breaking.
- Removed or altered E, missing imported fields, incorrect C despite valid Work, consumer substitution, and binding to independently held original inputs.
- Tampered bytes or signatures; coherently signed wrong recipient, domain or request revision; a missing or incorrect actual release Root join.
- Exact source-end and STATUS-end boundaries with a valid neighboring case; policy, dependency, schedule or occupancy changes; authenticated terminal highwater, restart, rollback, equivocation and unavailable current status.
- Foreign authority, unknown capability, code mismatch, stale Host/time and changed payload under an existing idempotency identity.
- Actual effect, receipt and readback; duplicate/repeat/VERIFY with zero new Work and effect; preserved uncertainty after execution starts and no blind retry.
- Denied body disclosure, private-data projection, bounded I/O/descent/budgets and duplicate evidence without independent credit.
- Lawful continuation after each scoped refusal, permutation, renaming, another request, and finite hidden changed values where an independent hidden trial is part of the declared study.
This is a design template, not a claim that every row received a new native injection. In the accepted Gate5 evidence, F01-F20 include historical generic donor evidence. P04/P05 combine native and pure signed boundary checks; P06 uses generic native controls; P08 uses pure altered evidence. The A-H obligations do not convert the mere presence of an author test file into evidence of executed tests.
The public frozen cases in tools/run_gate6_football_reference_v01.py::CASES cover ordinary and changed-budget requests, missing action/history permissions, denied release, revoked source, foreign audience, revision conflict, changed payload under a saved key, input permutation, different requests, overlap, occupancy and continuation. The recorded G54D1 native scope is 16 cases, 56 events and seven isolated mocks. Its separately recorded live scope is three proposals, five events and one mock. These historical counts are distinct from later Gate6 reruns. They are not one eight-payment history, and retained variations are not a new independent hidden benchmark. The same candidate implementation handled the declared request variations. Private answer tables are not necessary to communicate these public case lessons.
Nine recorded pure negative controls cover missing producer slots, missing producer total, empty E with an otherwise valid consumer, consumer substitution, foreign recipient, wrong request revision, stale STATUS, expired source and wrong release Root. The wrong-release control intentionally passes public check_bundle and fails the independent release_root_join. Coherent OFFLINE_TEST signatures cannot repair a semantic error. An intact neighboring case is accepted afterward. These controls are PURE_ALTERED_EVIDENCE_NOT_NATIVE_ATTACK; they are not new runtime race or crash evidence.
4.8. A general numeric example
The kit example/run_numeric.py::calculate computes y=3*x+2 through actual admitted PURE Work, an independent equation validator and an actual Root decision. Recorded examples are x=7 -> 23 and x=11 -> 35. A bool is refused before Work. Missing required input receives an actual Root refusal.
numeric_profile.py defines the explicitly local BoundedIntegerTransformV01 with common body fields plus sample and computed, and validates computed=3*sample+2. numeric_context.py::derive_context uses check_bundle(profile=PROFILE) together with independent policy, use_time, STATUS and highwater to derive C. numeric_work.py::perform executes the actual source/consumer path. Its consumer computes B.computed+B.sample+C.source_projection.request_revision, or 4*x+2+revision. At revision 1, the recorded outputs are 31 and 47.
The example demonstrates consumed evidence without a booking or allocation solver. The signed component example has peer_io=0 and effects=0. run_numeric.signed_profile treats release as a component cryptographic check, not actual peer release or authority. The numeric examiner independently derives expected C rather than using candidate consumer.inputs as the oracle. Explaining these recorded properties does not require a fresh run.
4.9. Frozen Gate5 kit versus the current source basis
START_HERE records the historical accepted basis 76a6952c9d0f31c976d2501c5641457a20de153a plus the reviewed G52 external proposal and listed G53 profile extension. The kit is immutable provenance. Its preparation-era pending statuses and Linux author-trial restrictions remain historical.
The complete static comparison of every kit/source file against its repository counterpart at current main 62716374021e273b46878a824af39c1a94593cd7, as of 2026-10-01, found exactly two byte differences:
1. hedgehog/domains/testflix/outcome_feedback_adapter_v01.py
Kit SHA-256:
ed0bf73b38cd3bd9ff51eeb38c342451e1c8b9c1c20ba31a5085a1245a18fbc2
Current SHA-256:
c979f15b8dc093a9924060efc5db05917ccf20c86d55b1a50b738ac4fa25bde6
G6A4R adds _public_request_v01(request), which uses e.plain_value_v01(request) and then removes bank_private and user_private. It replaces the direct asdict(request) public-request projection, correcting the tuple/JSON-array boundary without expanding permission.
2. hedgehog/drs_g2b_compatibility_v01.py
Kit SHA-256:
dff9ab0c9717289b3999844440ad2c16a1f9c64dca8ae4151dc4fd2e0a6b48b5
Current SHA-256:
5d4fb04cbf2e0819a0df5da17df724a0441747d2573fff924086dc9cfabaece1
G6A5R1 adds the exact _GENERATED_MEANING_WRAPPER_ID expression:
\A(?:request|trace):drsmeaning_v01:[0-9a-f]{64}\Z
_legacy_string_secret_reason is called from _legacy_id and the string branch of _strict_json_plain. Complete typed wrappers are recognized before the existing shared secret classifier. General secret checks, provenance, schemas and Root checks remain in force. The change neither strips/normalizes foreign IDs nor exempts arbitrary prefixed digests.
All other kit/source bytes, including the mapped Gate5 Work, Root, Host, Firewall and transport code, match this main. Static identity is not execution compatibility evidence.
A future authoring kit targeting current source must be a separately versioned and predeclared snapshot. Explicitly enumerate the accepted current source, import and resource closure; rederive signatures, source hashes and dependency inventory; review the impact and relevant predicates/examples; and freeze the result before claiming a cold trial. Do not edit historical kit hashes, silently overlay current modules on an old capture, change a historical source freeze to make a proof pass, or infer current execution from byte identity. A new domain can target current accepted interfaces with an explicit source basis and closure. The archived kit remains methodology, example and reference, not an evergreen installer.
docs/gate6_reference_v01/FINITE_COMPLETION.md, BOUNDARY_AND_FINDINGS.md and README.md make the source-view split explicit: ORIGINAL uses the old Testflix adapter and old compatibility module; REVIEWED uses the corrected Testflix adapter but old compatibility module; the current finite successor uses the repaired compatibility module. Older supplied workers run their exact historical source views, with no fallback to installed source and no repinning of the historical 810/811-source closures. Preparation-era admission/status text must be reconciled with accepted release and landing records rather than copied as current main status.
tools/run_gate6_football_reference_v01.py::prepare still compares recovered AUTHOR_VISIBLE/source dependencies byte-for-byte against installed counterparts and requires the exact reviewed 24-file candidate. The two differences above mean this historical preparation path cannot be assumed to be a generic fresh-domain runner on current source. The current Gate6 package deliberately materializes historical views for supplied checks. No fresh execution is implied by this description.
4.10. Accepted limits of the claims
The finite reviewed-reference trusted computing base includes exact source, runner, observer and examiner code, independently retained inputs, keys and readback. The evidence does not establish safety for arbitrary hostile Python, isolation from an operating-system administrator, general crash/race guarantees, production security, global federation/discovery/PKI, real venue or payment effects, or universal planning.
Peer transport is signed inherited framed pipes, not encrypted Internet-ready federation. Recorded bounds are one hop, four pointers, eight records, 16 lineage references, a 16 KiB pointer, 64 KiB body, 256 KiB frame and depth 12. Per-task bounds are two payload attempts, eight metadata/STATUS observations, two accepted bodies and 128 KiB. The finite protocol permits at most 32 requests and requires signed clean CLOSE; malformed or truncated frames are not clean EOF. The original stalled-IPC ceiling is 60 seconds.
PublisherBudget independently reserves attempts before release and disclosure bytes before body transmission. Restart reopens persistent counters. This is finite single-writer state, not a database lock.
Recorded signatures establish linked bytes and test-key identity, not factual truth. Supplied replay adds no current Root decision, Work execution or authority. Native lineage history and retained live-model captures are not fresh provider calls. Authoring-budget amendments and failed paths remain part of the record. Author test files were text-reviewed; their presence does not mean they were automatically executed. Provider prose can remain wrong even when the bounded typed parameters consumed downstream are correct. Installation, functional acceptance, publication and completion of the full program remain distinct claims.
05. ENGINEERING WORKFLOW, SOURCE ADMISSION, REVIEW, AND RECOVERY
======================================================================
05. ENGINEERING WORKFLOW, SOURCE ADMISSION, REVIEW, AND RECOVERY
======================================================================
This chapter describes a finite engineering workflow. It does not start a job,
authorize a repository change, install a review context, or grant runtime
permission. Resolve the actual task, source basis, available tools, and existing
state before applying any procedure.
5.1. Roles and the limits of each role
Task owner
Defines the intended outcome, material requirements, accepted scope, and any
authorized publication or external action. Accepts results within their stated
limits. Repository authorization and runtime Root authority are different
facts: a project instruction to develop or publish code is not a current
ActionCommitPacket for an application effect.
Work assistant
Maintains the technical context, reads the applicable inputs, turns the task
into a finite executable instruction, reviews returned artifacts, and reports
what is established and what remains. The assistant supplies a complete next
instruction when that work is already authorized. It does not infer success
from an executor summary or manufacture authority from candidate metadata.
Local execution agent
Works in the actual execution environment, inspects its repository and process
state, implements the allowed change, performs the required focused checks,
and returns the exact artifacts and evidence. It preserves unrelated work and
uses the authorized landing procedure when landing is part of the task. A
remote Work assistant does not automatically share this agent's filesystem,
credentials, browser session, or ability to inspect live processes.
Reviewer
Checks a declared obligation against independently obtained inputs and oracles.
Review may be performed by the Work assistant or a separately assigned reviewer;
state which arrangement was used. A new role label, another model call, or a
helper that emits an acceptance-shaped record does not establish independence.
Document the source, context, tool, and oracle separation actually provided.
Keep authoring, examination and runtime semantic participation distinct:
- A domain author produces a candidate implementation.
- An examiner evaluates that candidate using its declared independent inputs.
It may be deterministic code; it is not necessarily another model.
- A runtime semantic model provides proposals consumed by actual application
Work. Its output is untrusted evidence or a proposal, not a permission grant.
An executor must not complete an author's implementation behind the scenes
when the experiment claims independent authorship. A reviewer must not treat
an answer key derived from candidate-owned values as an independent oracle.
5.2. One finite work increment
A work increment, sometimes called a hop, has a known output and a stopping
boundary:
task instruction -> bounded implementation or preparation -> exact return
artifacts -> independent review -> decision and the next authorized task
Before starting, identify:
- The useful result and the exact deliverable.
- The input source closure and actual current basis.
- The files or external surfaces that may change.
- The predicate that accepts the result.
- The checks required to evaluate that predicate.
- The remaining boundary after completion.
A typical new-domain path consists of a domain contract and finite matrix,
implementation and examination, independent review, and authorized landing.
This is a workflow shape, not a promise of a fixed number of increments for
every task. Forecast known work separately from a reserve for concrete risks.
Do not create a universal framework, a new service, or a separate preliminary
project unless the requested outcome requires it.
Keep failures classified:
- A contract defect calls for a narrow contract decision and, when useful, an
independent generic probe of the changed interface.
- A candidate defect calls for precise feedback within the declared budget.
- A harness, reader, or serializer defect calls for a repair with provenance;
incomplete evidence must not be relabeled as domain success.
- A presentation defect calls for a bounded layout or wording correction.
- An infrastructure interruption calls for state inspection and continuation
from the last completed phase.
Before the last permitted correction in a measured authoring experiment,
collect the available finite diagnosis across independent safe branches. Do
not hide several independent findings behind an unnecessary fail-fast loop
and then attribute the exhausted correction budget solely to the author.
Independent checks may continue only where the contract permits them. A
dependent live or effect phase must not continue after its required gate fails.
Do not import the correction limit of a historical experiment into ordinary
development. For a new measured trial, declare the initial submission budget,
corrections, feedback, tools, hidden inputs, model, cost limits, and stopping
conditions before execution. Record any accepted protocol change explicitly.
Later success does not make the historical first attempt successful.
5.3. Architecture boundaries remain binding during engineering
Development convenience, a source-admission result, and an attractive report
cannot bypass the runtime authority topology.
- Root remains the sole local final and commit authority for its domain.
- Orchestrator proposes a bounded route; it does not accept its own proposal.
- BSEP remains the semantic membrane. The Semantic Architect proposes
structure and obligations. Runtime owns the materialized execution topology.
- Actors, child cells, executors, Work, Post V&V, and GT return proposals or
evidence. DRS, AVF, ratings, receipts, ledgers, cryptography, and replay are
not Root.
- Independent Roots make independent decisions. There is no implied SuperRoot
or transfer of authority between owners.
- A consequential effect requires a current Root-created scoped
ActionCommitPacket and the exclusive Effect Firewall or bounded Corridor.
A helper must not perform the same effect through a second route.
- Selected Work must actually execute, and the required later review or
consumer must use its result. Changing a label is not proof of consumption.
- Memory can inform a new decision. It does not preserve consent, permission,
finality, or proof that an action occurred.
- Observation time, receipt time, review time, and the authorized window are
different facts. A late response does not refresh an old observation.
- Expired, revoked, replaced, or consumed action authority does not revive
through replay, retry, context copying, or a new presentation.
Keep these distinctions explicit:
ForeignSignatureValid != LocalAuthorityAccepted
SourceAdmission != RuntimePermission
SourceAdmission != SemanticClosure
LocalAdmission != RemotePublication
ByteIdentity != TruthOfAnExternalObservation
A hash binds bytes; it is not automatically a provider signature, physical
attestation, independent reviewer, or local consent. Supported safe-derived
replay must not be described as full replay of an unsupported native schema.
5.4. Scope of the canonical upstream R1 procedure
R1 is the canonical upstream repository's reviewed source-transition policy.
It checks exact transitions against externally supplied review bindings and
real Git state. It is not a generic authorization service for every fork,
source export, or downstream application.
For work on the canonical upstream repository, use the applicable admitted
procedure, actual Git history, independently reviewed inputs, and authorized
operator. Do not substitute a convenient candidate verifier or an old example.
A downstream project must establish its own authorized repository policy,
review process, and runtime permissions. Copying an upstream review context,
installation bridge, owner-specific path, fixture, or accepted commit does
not authorize downstream changes. The upstream verifier intentionally binds
its named repository and history. A failure on an unrelated fork is not a
reason to remove its checks or claim upstream admission.
Public source exports can be read and used within their documented licensing
and execution boundaries. They do not contain every local Git object or
external review input required for canonical upstream landing. A ZIP of source
files is not a complete Git checkout and does not establish local admission.
The operative upstream files for a source-transition task are:
- AGENTS.md
- specs/current_architecture_lock_v01.md
- docs/repository_transition_admission_v01.md
- tools/reviewed_repository_transition_v01.py
- tools/check_active_architecture_authority_v01.py
- tests/test_repository_release_spine_v01.py
- tests/test_reviewed_repository_transition_v01.py
Read the current applicable source and contract. A historical proposal label
or archived instruction is evidence about its original phase, not a command
to repeat a completed landing. Conversely, source files alone do not prove
the current execution environment's clean status or installed context.
5.5. Raw identities, external pins, and context resolution
The upstream four-file enforcement identity map includes the admission
contract, policy helper, public architecture guard, and actual import-time
release binder. Each raw identity includes regular-file type, exact Git mode,
byte size, and SHA-256. The policy ID is the SHA-256 of the canonical map.
Focused policy tests are separate maintenance postimages, not a fifth entry
in that four-file map.
All raw bytes participate: constants, tables, functions, self-literals, and
comments are not omitted to avoid a hash cycle. An external bridge can bind
old and new identities without allowing the candidate to approve itself.
Expected manifest hashes, policy identities, and installation bridge pins
come from frozen independent review and the applicable owner instruction.
They are not inferred from candidate metadata, remote HEAD, the manifest's
own claim, a fixture, or a previous successful test. A context is an
operator-supplied trust input. It is neither a digital signature nor proof
that a separate reviewer exists. Platform approval is a separate boundary.
The canonical resolver accepts:
- An explicitly supplied absolute context path.
- The HEDGEHOG_REPOSITORY_REVIEW_CONTEXT_V01 environment path.
- reviewed_repository_transition_v01/context.json inside the actual Git
directory obtained with git rev-parse --absolute-git-dir.
A linked worktree has its own actual Git directory. When several context
sources exist, they must contain identical records; arbitrary priority must
not hide a conflict. Symlinks, malformed records, duplicate JSON keys, unknown
fields, and conflicting sources refuse. Guard and binder do not create,
repair, refresh, or invent context.
Ordinary proposals are inspected with previously admitted enforcement code.
The candidate must not choose an alternative verifier. The public guard loads
the policy helper beside that trusted guard, checks current invariants, and
resolves context. Without context, only the exact supported historical objects
can use the historical route; missing required review context is a refusal.
The actual release binder has import-time behavior. The binding named
G2E4_V0310_ACTIVE_BYTES_BINDING is computed by its real binding function; named
consumers include test_g37_binder_has_final_precedence_v01 and
test_g37_current_registration_pins_actual_release_files_v01. A script that
prints PASS does not replace these consumers. Do not import a large suite
merely to discover its structure or without understanding its import-time
work.
Invocation shape only; replace all paths after independent review and do not
run this as an unfilled instruction:
HEDGEHOG_REPOSITORY_REVIEW_CONTEXT_V01=/absolute/review/context.json \
/absolute/python /trusted/admitted/tools/check_active_architecture_authority_v01.py \
--root /absolute/candidate
This example does not distribute a review context, supply a pin, or authorize
execution. A downstream repository uses its own applicable authorized policy.
5.6. Closed records and transition kinds
Do not conflate these identities:
- accepted_basis: the immutable historical accepted implementation.
- base: the independently reviewed immediate basis of this transition.
- previous: the prior admitted commit, tree, policy, and manifest binding.
- finalized: the actual commit and tree already created for this transition.
- installation.tip: the actual installed maintenance commit and tree; it need
not equal the latest documentation commit or current main tip.
The manifest contains a complete ledger of safe relative paths, action A or
M, and preimage/postimage identities. Paths are unique and case-fold unique,
and cannot name Git internals. Additions have a null preimage. The supported
actions do not create a general delete or rename lane. A future commit hash
does not appear in the manifest.
The reviewed_repository_transition_v01 format has closed fields. Its manifest
binds schema, repository, policy_id, kind, accepted_basis, base, previous,
governed_namespaces, commit_message, and ledger. Its context binds schema,
repository, policy_id, kind, accepted_basis, base, previous, manifest, state,
finalized, and purpose. Use the actual contract for every structural record.
The reviewed_repository_policy_installation_v01 context adds installation:
an absolute bridge path, an independently expected raw SHA-256, and a nullable
actual installed tip. Do not silently use a context from an older schema as
the template for a newer admitted installation.
DOCUMENTATION
Adds exact new versioned inert capsules under
docs/showcase/<name>_v<number>. Every added file is ledger-bound, prior capsules
remain immutable, and code copies are inert data rather than executable
entrypoints. No runtime, guard, policy, test, or authority-metadata edit is
authorized by this kind.
The ordinary optional README change is exactly one bare English relative link
line. For example:
[Human label](docs/showcase/example_v01/README.md)
The line must end with the ordinary newline required by the actual predicate.
No bullet prefix, suffix explanation, unrelated whitespace, or other README
edit is included. Removing exactly that inserted line must recover the old
README byte for byte. Read the real predicate before writing the task.
ENGINEERING
Is a separately reviewed exact transition from the latest admitted tip. It
does not modify the enforcement chain, governance controls, prior capsules,
README navigation, Git configuration, or prohibited dot paths. Documentary
authority does not authorize engineering. Changes to protected registration
invariants require a separately reviewed maintenance instruction.
POLICY_INSTALLATION
Installs an exact independently reviewed enforcement successor. It is not a
general exemption from ordinary transition rules. The old verifier first
validates the admitted predecessor. Selection of the exact new verifier then
requires independent review and a separate applicable owner instruction.
Candidate tests do not provide that authorization. The historical BOOTSTRAP
route is a specific bounded proposal, not a reusable installation shortcut.
The finite upstream G6B and Human Entry maintenance bridges contain exact
publication exceptions. They bind a specific predecessor, complete old/new
enforcement, allowed maintenance postimages, registry identities, publication
ledger, README pair, message, and actual installed maintenance tip. Their
ability to replace one frozen README in one exact publication must not be
generalized to another README, namespace, payload, or later transition. The
public contract and source define these exceptions; historical examples do
not authorize a new one.
5.7. Current-state checks and preservation
Before editing, inspect the actual repository path, HEAD, full porcelain
status, branch, origin, index, applicable context, and task-owned processes.
Do not manufacture the expected state by resetting, stashing, deleting, or
silently including unrelated work. Keep an exact allowlist and intended A/M
delta. Every unlisted source, schema, contract, test, evidence, historical
capsule, public asset, and user change remains protected as applicable.
The canonical checks cover current branch/origin, operation markers, index
flags, parent/tree/ancestry, exact status and actions, complete current file
identities, protected neighbors, and the complete index stage/blob/mode map.
The proposal must be wholly unstaged or wholly staged in its exact supported
form. Partial staging and mixed states refuse.
Inventory consists of tracked files plus nonignored untracked files. Existing
and declared governed capsule namespaces are recursively checked, including
ignored files. This is not a claim to inventory every unrelated ignored
environment or cache. Immutable Git-object inventory may be cached; mutable
filesystem, context, index, metadata, and operation checks remain current.
Git mode and local POSIX access permissions are distinct. A regular repository
file may retain POSIX 0400, 0600, or 0644 while its Git identity is 100644;
POSIX 0755 maps to Git 100755. Other modes, including group/world-writable and
special-bit modes, refuse under this contract. The verifier never chmods
files. Repository permission compatibility does not relax external context,
detached manifest, or index mode rules. Do not normalize permissions with a
recursive chmod or rewrite preserved files for cosmetic consistency.
Run git diff --check and review the exact changed-path list before handoff.
Use the smallest deterministic checks sufficient for the task and its required
gates. Do not silently replace a focused check with a full historical suite,
install dependencies beyond scope, or call a provider to validate packaging.
Record exact commands, return codes, durations, failures, skips, and unavailable
checks. Source admission does not confer functional closure, runtime Root
permission, production certification, or publication.
5.8. Landing, M/P separation, and interrupted-operation recovery
M denotes a separately reviewed maintenance or policy-installation commit.
P denotes a separately checked publication commit on the actual accepted M.
Not every ordinary engineering change or documentation capsule requires a new
M. Use M/P only where the accepted exact procedure calls for it.
Freeze P's payload before landing. After the real M exists, deterministically
bind P's manifest and context to the actual M commit and tree. Independently
check that binding before applying P. A disposable rehearsal's M identity is
not an owner identity, and a predicted future commit is not a valid substitute.
The full landing sequence for each applicable transition is:
1. Verify the admitted predecessor, inputs, external pins, actual state, and
applicable authorization; perform any additional required operator checks.
2. Apply only the exact accepted payload.
3. Run the real public guard and binder for the complete unstaged proposal.
4. Stage only the exact allowed paths.
5. Run the real public guard and binder for the complete staged proposal.
6. Create the one exact commit if that commit does not already exist.
7. Run the real public guard and binder on the committed state.
8. Finalize the actual context with the authorized exact operator.
9. Run the real public guard and binder on the finalized state.
10. Perform the authorized ordinary push.
11. Independently read back the actual remote commit, parent, tree, inventory,
modes/object IDs, and changed bytes required by the accepted procedure.
For M/P, finish and verify M to the boundary required by the accepted operator,
obtain the required independent review of P's actual binding, and then perform
P's full sequence. Do not omit committed or finalized checks. Rehearsals do not
replace production phase checks, and local verification is not remote readback.
PREPARED requires null finalized data. It permits only the exact complete
unstaged proposal, complete staged proposal, or immediate sole-parent commit
on base. Before commit, origin/main equals base; for the exact committed child
it may equal base or that child. FINALIZED additionally binds the actual commit
and tree, refusing a different equal-tree child. A local verifier's
remote_publication=NOT_CHECKED is neither a successful remote readback nor
evidence that an independently verified push did not occur.
The authorized finalizer preserves the previous raw context under a unique
recovery name, writes a temporary ordinary file in the actual Git directory,
flushes it, and uses atomic rename as specified. If interrupted after commit
but before finalization, inspect that exact child using the same PREPARED
context. PREPARED_COMMITTED recovery finalizes the existing commit. It does
not create a second commit or regenerate review authorization.
If a push or transport fails after commit, inspect local state, remote state,
task-owned processes, and receipts before continuing. Resume only the missing
phase. Do not use reset, amend, force-push, or blind replay of an effect. Missing
context does not prove completion. A DISPOSABLE_TEST_ONLY record must never
be installed as the canonical owner's review context.
Preserve recovery materials with the reviewed return: prepared and finalized
contexts, actual manifests, external pins, bridges, independently checked
binding records, old raw contexts, command receipts, process state, and the
final checkpoint. A binding review may identify the PREPARED context, while
the final record identifies FINALIZED bytes; these hashes legitimately differ.
Relocation of absolute bindings requires the prescribed independent
reconstruction. Do not silently replace paths and treat the changed record
as the previously reviewed bytes.
5.9. Independent review of a returned artifact
Read the artifact and form source-based findings before using another
assistant's recommendations as a conclusion. Review of a summary alone is
not independent review of the underlying result. An executor's manifest and
PASS label are claims to be checked.
A sufficient bounded review normally follows this order:
1. Locate the actual input. A path in a message does not establish access from
another environment. Use the delivered attachment, retained artifact, or
exact public asset. If one required input is absent, identify that input.
2. Check the independently expected byte size and SHA-256. Before extraction,
inspect absolute paths, parent traversal, duplicate and case-fold-colliding
names, types, symlinks/hardlinks, modes, and reasonable counts/sizes. Check
complete gzip EOF/CRC or ZIP directory/CRC as appropriate. Do not execute
archive contents to inspect them.
3. Verify complete manifest coverage and each size/hash/mode. Apply its actual
self-inclusion or exclusion contract. Distinguish repository payload,
evidence, and transfer-only files. An internal manifest is not its own
independent trust anchor.
4. Read the task, final report, checkpoints, exact diff/postimages, command
receipts, and failures. Determine the actual source basis and imported
code. Establish that claimed checks cover the final bytes rather than an
earlier revision.
5. Reconstruct each material causal relationship: original input -> producer
-> review -> published bytes -> checked import/currentness -> consumer ->
Root/action -> receipt/readback. Derive expected values independently.
Not every local domain needs an external exchange; inspect the relationships
actually required by its contract.
6. Classify static analysis, saved-data checks, fresh native or mock execution,
retained captured LIVE evidence, fresh provider calls, authorship trials,
source admission, and publication separately. Preserve each result's scope.
7. Verify protected sources and modes, and whether task-owned processes ended
or were explicitly recorded for continuation. A saved environment receipt
is not a fresh observation of that environment.
8. Return a concrete decision: accepted within a named scope, an exact finding,
or a block on an exact missing input. Prepare the next finite instruction
when the next step is already authorized and its inputs are ready.
Not every review needs execution. Byte and causal review can be sufficient.
Repeat a run only to resolve a concrete remaining risk or satisfy a required
gate. Reviewing saved model responses does not inherently require another
paid model. Reviewing an archive does not inherently require application
runtime, hardware execution, or a fresh live demonstration.
Audit evidence is a collection of actual source closures, raw responses,
receipts, stdout/stderr, manifests, pins, causal links, failure records, replay,
and publication results. Do not invent a universal audit file. A frozen
PENDING or NOT_PERFORMED label describes its creation phase; later evidence
may complete that phase without editing the historical asset.
Finding record:
finding_id: <stable identifier>
obligation_or_claim: <exact statement>
source_paths_and_identities: <files and identities>
expected: <independently derived requirement>
observed: <actual value or absence>
phase: <where the discrepancy arose>
affected_results: <specific scope>
preserved_results: <unaffected established scope>
classification: <candidate / contract / examiner / harness / packaging /
tooling / missing evidence>
minimal_correction: <bounded change>
correction_check: <focused verification>
historical_status_retained: <unchanged prior result>
authorization_boundary: <new permission needed, if any, with its reason>
Explain the useful result, the unconfirmed part, the concrete cause, and the
bounded next check. A serializer or font problem is not automatically an
architecture defect. Do not promise a quick or final fix without understanding
the remaining dependency.
5.10. Template: a finite implementation task
This is a drafting template. Resolve every placeholder and remove inapplicable
sections before issuing a real task. Do not send an unfilled form as an
executable instruction.
BEGIN IMPLEMENTATION TEMPLATE
TASK: <unique task ID and finite deliverable>
Read the complete instruction. Inspect existing checkpoints and task-owned
processes first. Resume incomplete work from saved state; do not duplicate
completed or still-running operations.
Goal and authority
Implement <precise domain slice> using the accepted Radiolaria interfaces and
establish <named obligations>. This is <ordinary reviewed development OR a
separately declared independent-author trial>. The authorized scope is
<scope>. Repository landing, publication, provider calls, and real effects are
<explicit authorization state and boundary for each>. Existing authorization
that actually applies remains valid; do not infer unrelated authorization.
Inputs and basis
Repository or execution root: <actual path>.
Expected origin and branch: <verified values>.
Expected base commit and tree: <independently verified values>.
Input artifacts: <exact names, byte sizes, external SHA-256 values, locations>.
Applicable instructions: <AGENTS, current lock, authority classifications,
named contracts, focused source paths>.
Read actual source signatures. If current state differs, report the exact
difference and reconcile before affected mutation. Do not reset, stash, or
erase work to manufacture the expected state.
Isolation and preservation
Use <authorized separate clone/worktree or other bounded environment>.
Allowed paths and expected A/M delta: <exact list>.
Protected sources, evidence, modes, contexts, assets, and user work: <list>.
Do not edit the examiner or kernel contract to make a candidate pass. Report
a concrete contract contradiction with reproducing inputs and source paths.
WHAT
Intent and useful result: <complete specification>.
Actors, owners, input schema, normalization: <specification>.
Hard constraints and changes requiring consent: <specification>.
Deterministic selection and tie-breaking: <specification>.
Outputs, currentness, side effects, privacy: <specification>.
Budgets, timeouts, failure and reconciliation behavior: <specification>.
Do not expose a private examination answer as part of the public task.
HOW and source closure
Reuse <reviewed interfaces and admitted source closure>.
For external evidence exchange, keep producer E, published evidence B, and
independently derived consumer context C distinct. Derive expected consumer
inputs from independent original input/body/status/policy/time. A purely local
domain does not need an artificial peer exchange. Bind actual native Work,
required Root decisions, and actual receipt/readback where applicable.
Prior admission or foreign approval does not grant current local authority.
Finite acceptance
Freeze <case matrix> before execution. Each row states the expected behavior,
independent oracle, evidence kind, and relevant safe successful neighbor.
Run <focused checks and required gates> only.
Distinguish fresh native/mock execution, altered evidence, retained supplied
results, captured provider evidence, fresh model calls, and static checks.
Provider budget, if authorized: <provider/model/effort/service, role, calls,
tokens/cost, retries/corrections, stop conditions>. Otherwise make no new
provider calls. Never fabricate output in a LIVE-labeled lane.
Completion and recovery
Preserve exact inputs, phase, failures, and partial successes if blocked.
Inspect receipt and current state before repeating an uncertain effect.
Stop the dependent phase; continue independent safe work only as the contract
permits. Do not mutate the canonical repository unless this task authorizes
the exact accepted transition. Where authorized, use the accepted operator,
external pins, and all required R1 phase checks. Downstream projects use their
own authorized policy, not a copied upstream context.
Return
Deliver one verified archive to <actual delivery directory or channel>.
Include final report, exact candidate/postimages and delta, source/dependency
identities, finite matrix, commands and return codes/stdout/stderr/timing,
preserved failures, evidence relationships, preservation/modes, process and
resume state, and the full safe-path/size/SHA-256/mode manifest.
Reopen and verify the delivered copy. Report status, scope, limits, location,
bytes, SHA-256, and the next remaining boundary. Do not start a new stage
outside the authorized task.
END IMPLEMENTATION TEMPLATE
5.11. Template: independent review and exact landing
BEGIN REVIEW TEMPLATE
TASK: <review ID>
Review <artifact and external identity> against <task/contract/source basis>.
This is <static byte and causal review / specified authorized execution>.
Allowed inputs and tools: <list>. No additional runtime/provider/effect work
is authorized unless explicitly listed here.
First verify artifact safety, full manifest coverage, source identities,
actual final bytes, command provenance, and failures. Then evaluate
<finite obligations> using <independent inputs/oracles>. Separate recorded
claims from newly observed results. Inspect the relevant actual consumer,
currentness, Root, and effect boundaries. Do not execute archived code merely
to read it or accept candidate-owned expected values.
For each finding use the record in section 5.9. Preserve unaffected accepted
results and historical failure states. Return <decision artifact>, exact
source references, checks performed, unavailable inputs, and the one next
authorized step or concrete decision needed. No repository mutation or
publication is authorized by this review instruction unless stated separately.
END REVIEW TEMPLATE
BEGIN CANONICAL UPSTREAM LANDING TEMPLATE
TASK: <landing ID; exact transition kind>
Apply only <accepted transition> under <applicable owner authorization>.
This task is landing of reviewed bytes, not new development.
Accepted inputs
Archive: <name, bytes, independently expected SHA-256, actual location>.
Base: <actual commit/tree/origin/branch and required clean state>.
Operator and helper identities: <paths and external raw pins>.
Policy, manifest, context, bridge, and records identities: <independent pins>.
Allowed A/M ledger: <exact paths and pre/post identities>.
Protected neighbors, modes, capsules, releases, and assets: <scope>.
Review decision and applicable owner authorization: <exact references>.
Required rehearsal: <specified disposable procedure, or NOT_REQUIRED>.
Inspect actual repository, context, index, operations, and task-owned
processes. Verify the admitted predecessor and all input/helper pins before
using them. Refuse a changed basis or unrelated dirty work; do not erase it.
A disposable context must not become the owner's context.
Execute the accepted sequence, including its additional required checks:
apply -> unstaged guard/binder -> exact stage -> staged guard/binder ->
one exact commit -> committed guard/binder -> finalize -> finalized
guard/binder -> authorized normal push -> independent remote readback.
For M/P, bind P only after actual M exists. Obtain <required independent
review> of that actual binding before P apply. Preserve the frozen P payload.
Do not substitute a rehearsal commit or a helper's self-generated approval.
Recovery
If an exact committed child already exists, inspect and finalize that child
under the same approved identities. Do not create another commit, reset,
amend, force-push, or regenerate approval. If publication status is uncertain,
inspect actual remote state and receipts before continuing. Preserve old raw
contexts through the accepted atomic replacement procedure.
Verify actual remote commit/parent/tree, complete required inventory and
modes/object IDs, and every changed byte. Preserve all unrelated sources and
assets. Report final context, local admission, remote publication, protected
state, completed or remaining processes, exact commands, and limitations.
Deliver <verified return archive and manifest> to <actual destination>.
END CANONICAL UPSTREAM LANDING TEMPLATE
Do not issue this landing template with placeholders. Do not adapt it to a
fork by copying upstream owner context. A downstream landing task must name
that project's independently authorized policy and evidence requirements.
5.12. Template additions for an independent-author experiment
Apply these additions only when authorship independence is part of the
measured claim, not to every implementation task.
- Freeze separately versioned public HOW, WHAT, OUTPUT requirements, allowed
source closure, and private examination/oracle material.
- Specify the provider/model, resource and cost budget, initial submission,
corrections, feedback format, stop conditions, and permitted tools.
- Establish the actual exposure boundary. A configuration field or an empty
environments list is not itself a proven filesystem or tool allowlist.
- Give the author only the permitted corpus. Exclude examiner answers,
hidden cases, earlier candidates, solution history, and any handoff that
contains those answers.
- Record fresh context, provider/model identity, actual tool exposure, every
submission, every feedback item, and all attempts, including failures.
- The managing executor does not secretly write the implementation for the
author. Any permitted intervention must be defined and recorded.
- Use a cheap generic probe before expensive authoring only when it resolves
a concrete changed interface or uncertain exposure boundary.
- If the claimed isolation cannot be established, do not label the run blind
or independent. Identify the specific missing fact and an available
permitted alternative. Do not turn setup into unbounded preliminary work.
Any change to the measured protocol must be explicitly recorded and accepted.
Historical attempts and their original status remain part of the evidence.
5.13. Delivery, review conclusion, and live checkpoint templates
Executor launch text, to be filled with a real task file and destination:
Read <absolute path or accessible artifact URL of the complete task file> in
full. Complete that bounded instruction, starting by checking saved state and
task-owned processes. Do not repeat completed operations. Exact inputs,
constraints, and applicable authorizations are in the file. Return the
verified archive to <actual destination> and give a concise English result.
Work review conclusion:
Result: <accepted in named scope / exact finding / exact missing input>.
Checked: <bytes, source relations, saved traces, fresh execution; only done>.
Established: <useful outcome>.
Limits: <material boundaries supported by evidence>.
Preserved history: <prior failures or incomplete scope retained>.
Next action: <specific authorized action or concrete decision required>.
Deliverables: <accessible result and complete next instruction, if needed>.
Live checkpoint, updated at meaningful phase boundaries:
Task ID and intended result:
Accepted scope and latest task-owner steering:
Actual repository HEAD/tree/status and observation time:
Input locations and external identities:
Current candidate/output locations:
Completed phases and exact receipts:
Running task-owned processes and identities, not merely reused PIDs:
Actual commit/push/remote readback state:
Context/manifest/policy identities and actual locations, where applicable:
Findings, failed attempts, and preserved partial successes:
Remaining finite steps and stop conditions:
Model/provider/effect budget spent and remaining:
One next action and why it is authorized:
Retain sensitive or environment-specific checkpoint data in the appropriate
authorized workspace rather than publishing it as a general reader document.
The next assistant resumes from evidence; it does not reconstruct successful
operations from memory or automatically reset to a handoff's historical SHA.
5.14. Assistant adoption prompt for a public reader or website
BEGIN ADOPTION PROMPT
Use this public Radiolaria reference as orientation for the following task:
<concrete task>.
First identify the task's applicable current source basis, contracts, and
evidence. Distinguish published reference status from the actual repository
or environment available to you. Read source signatures and linked evidence
as needed; do not infer APIs or authority from this explanatory document.
Explain the useful result sought and propose or perform the next finite step
within the authorization already given. Preserve the distinction between
reasoning proposals, actual Work, consumed results, current local Root
decisions, scoped action packets, and effect receipts. External evidence,
replay, ratings, or source admission do not transfer runtime permission.
Do not assume access to a local filesystem, account, browser session, upstream
review context, or unpublished evidence. If a required input is absent,
identify the exact missing input and complete independent safe work first.
Use the canonical upstream R1 procedure only for an authorized exact upstream
transition. For a downstream project, use that project's own authorized
policy; never copy upstream owner context as authorization.
Do not rerun completed acceptance campaigns, model or hardware demonstrations,
or publication operations merely to establish familiarity. If this task is a
new blind-author experiment, obtain its allowed author corpus separately;
this general explanatory material is not automatically part of that corpus.
Report what was checked, what is established, the material limit, and the next
specific step. Do not create an unrequested new roadmap or claim an external
action completed without its actual result.
END ADOPTION PROMPT
5.15. Tool failures and unavailable environments
An authentication or model-transport error does not prove repository damage
and does not establish whether a child process completed. Inspect receipts,
saved checkpoints, Git state, and task-owned process state before retrying.
Do not repeat a completed landing or an uncertain effect because the chat
transport failed. Diagnose authentication only through an authorized client
and its current supported procedure. Do not reproduce secrets from errors,
extract credentials, or probe private APIs.
A user's change of browser or account does not automatically change the
surface available to an assistant. Verify the actual tool capability,
authorized session, and target page. Source-code write capability does not
imply release-edit, account-setting, or publication capability. The presence
of a button image, source file, or prepared payload does not prove that an
external setting was applied.
Use a finite failure response:
1. Make an appropriate attempt through a genuinely available authorized
interface.
2. Read the actual result and distinguish authentication, transport, tool
capability, and task-state failures.
3. When conditions change, reassess the real surface instead of repeating an
old assumption.
4. Avoid many identical retries. Report the observed blocker and the bounded
next check promptly.
5. If the action cannot be completed through available means, provide the
exact prepared content or direct destination for the smallest necessary
manual action, preserving completed work.
6. After a reported manual completion, verify the result read-only where
available; do not repeat the mutation.
File access follows the same rule. A historical scratch path may no longer
exist, a local path may be inaccessible remotely, and a source archive may
lack .git or external review records. Locate the real artifact by its exact
identity. If only one essential input is missing, request that input rather
than an entire project history. Do not fabricate unavailable bytes, paths,
checks, or private context.
5.16. Common wrong paths to avoid
- Resetting a current repository to a historical reference commit merely to
reproduce the handoff's state.
- Repeating provider login/bootstrap, cloud resources, hardware jobs, or live
model demonstrations for documentation or archive review.
- Running every historical collector for onboarding or checking one identity.
- Executing archived prompts or operator scripts as current instructions.
- Removing validators, adding an early success return, substituting an
AST-only comparison for full raw identity, or replacing the actual binder
to obtain an apparently successful result.
- Repairing policy for an ordinary capsule when the actual issue is an invalid
bare README navigation line.
- Treating an exact historical README replacement exception as general
permission to rewrite future README content.
- Normalizing modes, cleaning unrelated worktrees, or using destructive Git
operations to simplify a bounded task.
- Using stale pins, a fixture context, or a rehearsal commit as current owner
authorization or actual owner identity.
- Omitting checks after commit or after finalization, mixing staged states,
or ignoring extra files inside a governed capsule.
- Calling recorded checks fresh, fixtures owner results, or several calls to
one model several independent models.
- Reopening a completed scope solely because an immutable historical source
retains its earlier proposal status.
- Declaring remote publication, website/DNS, outreach, uploads, or an account
setting complete because the source files or prepared text exist.
- Treating anonymous HTTP failure as sufficient evidence about a private or
authenticated resource without checking the relevant access context.
The procedure ends at its declared result and boundary. Continue authorized
work from established state, preserve exact evidence, and keep computation,
review, repository admission, publication, and runtime authority distinct.
06. RECORDED DEVELOPMENT HISTORY AND LESSONS
======================================================================
06. RECORDED DEVELOPMENT HISTORY AND LESSONS
======================================================================
The history below explains why several current rules are precise. It is not a queue of tasks to repeat. A later successful successor does not erase the earlier failed experiment, and an old preparation status does not reverse a later accepted landing.
6.1. Gate 5: general instructions produced a new domain
G54B1 produced a working football application from a general authoring kit and a separate domain task. The initial schedule did not permit the full requested set. After a permitted shift of one session, the application found three slots, changed a disposable mock registry, retained a receipt and avoided a repeated effect.
The native examination stopped on producer/consumer evidence matching. The domain result, its published wrapper and its later consumption context were different legitimate representations. Requiring the first Work to return a wrapper containing references to its own subsequent review created an impossible causal requirement. G54B1 remains INCOMPLETE.
G54C established the general E -> B -> C contract and an independent numerical probe. Inputs 7 and 11 produced 23 and 35; the consuming Work produced 31 and 47. Nine negative controls examined the relationship. The change repaired a general evidence contract rather than selecting a preferred football answer.
G54C2/C3 clarified the trusted boundary. A Host and observer sharing a process do not by themselves require a new action service. The reviewed-reference code and observation assumptions must be stated. Protection against arbitrary hostile in-process Python was not established.
G54C4 established readiness for a finite reviewed-reference examination: five native scenarios and 16 events, plus 22 altered-evidence controls. The controls were not 22 newly executed native attacks. The earlier candidate passed the selected scenarios unchanged.
G54D ran a new cold author with an initial submission and three corrections. The final native matrix passed 15 of 16 cases. For REVOKED, the candidate safely stopped use without a new booking effect, but did not execute the separate local Root refusal required by that profile. The missing decision was not merely a packaging omission. Earlier fail-fast feedback had not reached this defect before the last permitted correction. G54D remains INCOMPLETE.
G54D1 explicitly extended that protocol by one fourth correction in the restored authoring lineage. It was not a new cold experiment or success inside the original correction limit. Native acceptance passed 16 of 16 cases, 56 events and seven isolated mock executions. Three live Gemini proposals were actually consumed. Two supplied checks produced identical results without additional calls or effects.
G55 admitted the exact engineering result:
64f2b38b0d671e4907bae8649fff8f3d2ee0751c
212 additions; 4,367 existing files preserved.
G56L published the reference presentation/evidence entry and closed the accepted Gate 5 Reference scope:
b6d11f5fcb619a10e877bbaaf6770881be315f16
399 added showcase files and one README link line; 4,578 other files preserved.
The transferable lessons are:
- Validate the general handoff contract before spending a new authoring budget.
- Distinguish an implementation error, an examiner error and missing evidence.
- Give the author complete available diagnostics within the remaining correction budget; do not repeatedly hide independent defects behind the first failure.
- Repair the applicable refusal behavior with a real decision, not a fabricated report field.
- Preserve protocol amendments and unsuccessful predecessors explicitly.
6.2. Gate 6A: finite composition and two narrow source repairs
G6A0 defined a 73-row finite matrix and early probes. A1, A2 and A3 established direct boundaries, portable composition and a frozen runner. Their 116, 135 and 175 test selections describe different stages; they are not independent counts to add together as proof of one version.
G6A4 completed 10 of 11 profiles. Living and Conformance passed; football passed 16 cases and 32 supplied checks. The live comparison used four Gemini generations and four countTokens operations, with 4,493 reported tokens and no retries. G35 Testflix refused a request projection containing tuples where the external schema required arrays. The failure was retained, and the frozen run was not patched in place.
G6A4R repaired the public request projection through plain_value_v01 while excluding private request fields. Five regression tests, fresh DIRECT 175 tests/525 phases and a complete G35 run passed. Nine retained profiles underwent separate compatibility review. The original G6A4 remains INCOMPLETE.
G6A5/A5R still lacked four complete obligations: N1-05, N4-04, N4-05 and N4-07. N4-05 reproduced a false positive in which 13 digits inside a generated provenance hash were interpreted as a card number. Valid Root and writeback geometry had passed. A separate serializer error prevented complete N1-05 evidence from being saved; that incomplete record was not labelled PASS.
G6A5R1 added exact recognition of the full generated identifier form:
\A(?:request|trace):drsmeaning_v01:[0-9a-f]{64}\Z
The change applies at two local compatibility conversion points. It does not remove the general secret classifier, broaden Root authority, or accept arbitrary prefixed text. All four obligations passed, along with 77 focused tests, identical clean-reader outputs and the required disposable R1 recovery procedure.
The four obligations establish different relationships:
N1-05: a high AVF score does not create action permission; independent builders and a lawful actual Root/mock neighbor were checked.
N4-04: real cold/warm execution reduced domain/Host Work from 1 to 0 while retaining an actual read and current Root decisions. This is qualified reuse, not a measured wall-clock speedup; two Root decisions remained.
N4-05: an immutable policy successor changes current lineage. An older record can remain inside its TTL and still be ineligible as the current policy. SUMMARY_ONLY does not open payload. This is not a global currentness service.
N4-07: source-bound inspection of four historical LIVE projections and their actual consumption, together with eight controlled checks. This was not a new live canary experiment.
G6A6 admitted and published the technical package:
commit 2e965ecb18e545e428380eb8e9aa5a7037a388be
tree 59a43cc0da470886804aab6c9af87fe17dfbd685
8,356 published files checked; 4,976 protected neighbors preserved.
The installed pure verification used 13 workers and two G5 children, completed in 126.26 seconds and matched the accepted canonical result. It made no new model calls.
The matrix has 69 satisfied rows and four approved deferments: N1-18, N1-19, N4-12 and N4-13, concerning the Factory/NeedleFactory and Marennya/UP programmes. Do not mark these four PASS or automatically schedule them when onboarding. The accepted package has a finite mixed source/evidence basis; it is not 11 fresh profiles run against one final source freeze.
6.3. Publication and reader examination
The B4 publication is a 25-slide editable presentation with a corresponding searchable PDF, a 58-page whitepaper and an eight-page Geometry Atlas comprising an overview and seven views. The accepted export uses Nimbus Sans in place of Arial for the defined profile; separate font programs are not distributed. The reference files do not need reformatting merely because a new assistant has arrived.
The first B5 attempt made no reader turns because the required tool-access boundary was not established. Settings and an empty environments list were insufficient evidence of the actual exposed tool set. This failed preparation remains recorded.
The accepted B5 continuation used two fresh contexts of the same gpt-6-astra/ultra model profile, not two different models. Route A had no tools; route B used a bounded read-only broker. The independent review accepted A at 16/16 and B at 20/20 with no critical confusions. There was one initial assignment per reader and no retries. The recorded durations were 155.64 and 304.88 seconds. B performed 59 operations: 42 reads, six listings and 11 searches.
All 59 broker responses were reconstructed from pinned inputs; the hidden rubric was withheld. The transport adjustment retained the assignment, model profile and effective access limits. Its documented trust scope depends on saved exact-version source, configuration, derivation/loading records and observable events. It does not claim remote platform attestation or prove the absence of all prior model knowledge. It is not permission to bypass a platform's tool or access policy in a later experiment.
HUMAN_NOT_RUN remains explicit. The recorded reader inputs are immutable. A later enlarged explanatory Reader or full-source export is not automatically a newly reader-tested document. See docs/showcase/radiolaria_os_v01/release/B5_ACCEPTANCE.json and related source-bound records.
The publication transition used a maintenance M followed by the reviewed publication P:
M 2201c5a970c6a118c312ba213e113dcbcd4b2ccd
P 800f38c3407b3f4ec93dc943c977ffeecb614881
A later Human Entry V02 transition used:
M a9bb644064e887cb859201d052962e3aa2ac1479
P 62716374021e273b46878a824af39c1a94593cd7
These are public historical identities, not reusable authorization for a new transition. The eight release assets remained unchanged by the later entry improvement.
6.4. How to inspect accepted evidence without rerunning everything
Read the finite reference documentation before choosing a verification mode:
docs/gate6_reference_v01/README.md
docs/gate6_reference_v01/TECHNICAL_REFERENCE.md
docs/gate6_reference_v01/BOUNDARY_AND_FINDINGS.md
docs/gate6_reference_v01/FINITE_COMPLETION.md
demo/verify_gate6_reference_v01.py defaults to a pure source-bound inspection mode. Its expected successful package status is PASS_FINITE_MIXED_BASIS_REFERENCE_PACKAGE. Without an independently supplied expected manifest pin, self-consistency is not independent acceptance.
The supplied mode runs the declared checking workers against preserved evidence and exact source views. Original, reviewed and current source views are different and must not be mixed inside one interpreter or silently replaced with installed modules. A supplied check does not fall back to a fresh producer when evidence is missing.
Inspection of saved files, execution of saved-data verifiers, fresh deterministic reference execution and optional live experimentation are separate operations. Use only the level required by the task and its actual authorization. A new domain's verification is selected for its changes; the complete historical campaign is not the default development loop.
07. PUBLIC RELEASE, READER ROUTES AND RIGHTS
====================================================================== 07. PUBLIC RELEASE, READER ROUTES AND RIGHTS ====================================================================== 7.1. Choose the appropriate public material RADIOLARIA_PRESENTATION.pptx and RADIOLARIA_PRESENTATION.pdf The primary 25-slide overview. The PPTX contains editable text, shapes and arrows; the PDF retains searchable text. RADIOLARIA_WHITEPAPER.pdf The 58-page connected technical explanation, including definitions, examples, implementation routes and evidence limitations. RADIOLARIA_ATLAS.pdf The eight-page Geometry Atlas: overview plus seven complementary views. READ_RADIOLARIA.xml The explanatory machine-readable entry, 62 blocks and approximately 1.49 MB. Suitable for first orientation with an assistant that can accept that input size. It is not the full source tree. RADIOLARIA_REVIEW_BUNDLE.zip The preserved reader/review package with its historical input provenance. Do not rename it conceptually into the complete current repository or assume that it includes later navigation changes. RADIOLARIA_FULL_SOURCE.xml The complete source export, approximately 807.5 MB. Use file tools, indexes and bounded searches. Do not promise that an ordinary model context can consume the entire export as one prompt. RADIOLARIA_FULL_SOURCE_BUNDLE.zip Approximately 339.3 MB and 8,757 files, including the declared snapshot and packaging indexes. It is not a .git directory or the entire Git history. Four loose historical font files are explicitly excluded. The later main branch is intentionally different from this frozen release snapshot. Use GitHub's README as the human entry, the Release as the download entry, and the version DOI for citation of this exact published release. The all-versions DOI is a version-family reference; it does not identify a different file set for this snapshot. 7.2. Link and archive integrity The Zenodo metadata identifies the same release version and links to the GitHub Release using isIdenticalTo. The GitHub Release description links back to the Zenodo record and lists the version DOI. The eight files were fully streamed and compared by exact name, size and SHA-256. Their inventory is included in Section 09. For a later archive check, download bytes through an ordinary streaming client rather than passing large assets through a model. Distinguish a metadata comparison from a full byte comparison. Record which was actually performed. Do not create a new record, version, tag or rebuilt asset merely to check an existing archive. For an offline reader package, preserve the relative path structure and the dependencies required by links. A PDF copied away from its references may still open but no longer provide useful evidence navigation. Any personal index or backup packaging is separate from the frozen release assets; it does not authorize changing those assets. 7.3. Rights and accurate attribution The repository's LICENSE and COMMERCIAL-LICENSING.md remain the source for code licensing. Code is AGPL-3.0-only; alternative terms require a separate written agreement. This handoff is not such an agreement and does not infer a commercial grant. The accepted RIGHTS.md assigns CC BY 4.0 to the specified new explanatory author text and original diagrams. It excludes code, verbatim code excerpts, historical source/evidence and third-party material. The mixed full-source ZIP/XML does not receive a blanket CC BY licence. Follow the actual attribution and notices rather than applying one label to every file. Image permissions and font embedding qualifications have their own defined scope and provenance. Public use authorization for identified images is not a certification of every possible rights chain or permission for every future image. Loose font programs are not included in the release assets. A website adaptation of this handoff should retain its snapshot, explanatory status and source links. It should not claim that the document is a new accepted runtime contract, a security certificate or a replacement for project licensing. 7.4. Public explanations and future work Present the useful computation first, then its mechanisms and evidence. Keep Work, decision, effect, receipt and history distinct. Preserve the difference between actual model contribution, a controlled owner input, a mock registry change and a real-world transaction. Regulatory mapping can discuss traceability, scoped authorization, human-oversight integration and inspectable failure history. It should separate a real requirement, the engineering mechanism, the evidence available and the integrator's remaining obligations. Root does not automatically mean a human. The reference kernel is not automatically an EU AI Act or ISO certification. Check current primary regulatory sources for a new legal or compliance claim; do not treat a historical chapter as updated legal advice. Website deployment, domain/DNS configuration, outreach, new integrations and future research are separate work scopes. A repository homepage field does not prove a deployed site, and a prepared email does not authorize sending it. New domains use the method in this handoff but earn their own acceptance; historical football success is not acceptance of a different application.
08. UNDERSTANDING CHECKS AND CONTINUITY
====================================================================== 08. UNDERSTANDING CHECKS AND CONTINUITY ====================================================================== These questions are useful for an assistant learning the system. They are not a hidden test and should not be mixed into a new blind-author input without declaring that exposure. 1. What useful work does Radiolaria support beyond refusal? It assembles and continues typed work around an intention, combines different forms of computation, consumes actual results, reuses qualified experience, retains unaffected work and performs explicitly permitted consequences. 2. Does each nested Work need a new Root? No. Computational nesting and ownership boundaries are different. A child remains bounded; independent Roots correspond to actual owners, not every level of a graph. 3. Does a foreign signature authorize local use or action? No. It binds bytes and the recorded signer identity. Current local policy and Root decide admissibility, use and consequences. 4. Why are E, B and C different objects? E is the producer result. B is assembled after the producer and its relevant review, preserving E and evidence references. C adds later verified consumption context. Exact projections and causal joins matter; equality of all three raw representations is not required. 5. Why is a candidate's declared input not an independent oracle? A substituted input could validate itself. Expected input must be reconstructed from independent original data and the verified body, status, policy and time. 6. Why did a safe REVOKED outcome fail the G54D profile? The profile required a separate actual local Root refusal, and that branch had not invoked it. The authorized D1 continuation repaired it. This does not create a universal rule that every early type rejection must call Root. 7. What follows an effect with an uncertain reported outcome? Preserve the actual stage, error, possible receipt and state/readback. Do not label the operation a clean no-effect refusal or repeat the mutation blindly. 8. What does repeating a successful booking mean? The application returns checked history/readback without another effect. A terminally consumed native packet does not grant a second execution. 9. Can replay recreate Host authority? No. Saved evidence can establish recorded relationships; it does not restore current Host state, live source capture or permission. 10. Can a frozen kit's source be silently replaced with main? No. A new experiment needs an explicit versioned source closure. Historical evidence remains on its recorded source basis. The two known Gate 6 differences are not a reason to rewrite Gate 5 history. 11. What does cold/warm Work 1 -> 0 establish? A reduction in specific domain/Host dispatches through actual qualified reuse. It does not establish a wall-clock speedup, absence of Root decisions or training of model weights. 12. Can an LLM participate in execution? Yes, as a bounded semantic participant whose actual output is consumed. Its proposal does not create Root authority or permission for an effect. The runtime owns the execution topology. 13. Why does main differ from the release target? A later independently admitted Human Entry transition changed navigation and illustrations. The eight reference assets remain pinned to their publication commit. 14. Can a successful source review authorize a current action? No. Source review, native capability admission and current Root/Host action authorization are separate decisions. 15. What must be checked before resuming an interrupted landing? Actual Git/context state, process identities, existing commits, receipts and remote publication. Resume the missing phase; do not create a second commit because the assistant connection was interrupted. 16. Is the entire programme complete because the reference release is public? No. Technical acceptance, finite reference closure, publication, integrations and approved research deferments are different claims. This distinction is not a reason to reopen completed obligations. 17. Does a local pure domain need an external DRS exchange or a mock effect? No. Select the applicable interfaces and acceptance rows. The E/B/C exchange contract and action-specific tests apply when those features are in the task's declared scope. 18. What does an evidence hash prove? The identity of the bytes covered by that hash. It does not establish semantic truth, provider origin, current applicability, authorization or independent review by itself. Continuity record for an active task: - Task identifier, goal, accepted scope and latest steering. - Actual repository identity and the time it was observed. - Exact input artifacts and external expected hashes. - Candidate/output locations within the current authorized environment. - Completed phases, failed attempts, receipts and remaining finite work. - Task-owned running processes and unambiguous process identities. - Existing commit, finalization, push and remote readback state. - Applicable source, policy, manifest and context identities. - Model/provider/effect budgets used and remaining, if applicable. - The next concrete action and why it remains authorized. Write this record at meaningful boundaries. Do not copy a private operator context into a public artifact. A public handoff needs reproducible source identities and methods; an active operator checkpoint may also need environment-specific bindings that should stay with the authorized operator. When preparing a successor handoff, put the latest verified state first. Preserve the reasons for important decisions, exact source routes, unresolved findings and evidence limits. Avoid dumping an entire chat or reviving old PENDING statements as current blockers. One coherent entry should make the next task easier without claiming to replace its actual inputs.
09. SOURCE IDENTITIES AND PUBLIC ASSET INVENTORY
====================================================================== 09. SOURCE IDENTITIES AND PUBLIC ASSET INVENTORY ====================================================================== 9.1. The eight frozen release assets Each size is the exact raw-file byte count. SHA-256 applies to the downloaded file, not to a reconstructed or unpacked representation. These identities preserve the published reference snapshot; they do not authorize replacing an asset or assigning the same tag to different bytes. RADIOLARIA_ATLAS.pdf bytes: 36663 SHA-256: 835acce5837b8c8138e07ae3657d011ab454d0446a708fbdc2a1d164c8aaa9b6 download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_ATLAS.pdf RADIOLARIA_FULL_SOURCE.xml bytes: 807537837 SHA-256: 01c19c93ee6d2707ac43db62b3e0868ce009c18225adb0602acda59990b04bb9 download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_FULL_SOURCE.xml RADIOLARIA_FULL_SOURCE_BUNDLE.zip bytes: 339347838 SHA-256: 6a176f910cdb9737f8417128911f80a6d0590397f806cfc6fc189ae400135f7c download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_FULL_SOURCE_BUNDLE.zip RADIOLARIA_PRESENTATION.pdf bytes: 616975 SHA-256: 74bcce61dc67da68ec749a94d1587b1b2c47e795c72c33ac555dc68de6c55fc9 download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_PRESENTATION.pdf RADIOLARIA_PRESENTATION.pptx bytes: 1728230 SHA-256: 01e9482b304929d7bae63e01479dd0365e172ba4e5bd3d625f4b25b0262dd48f download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_PRESENTATION.pptx RADIOLARIA_REVIEW_BUNDLE.zip bytes: 12895598 SHA-256: 027f694fdd06cf8d686bfbd42bad47a29a8bbdec8418cd74c380fa6cc8b12a29 download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_REVIEW_BUNDLE.zip RADIOLARIA_WHITEPAPER.pdf bytes: 302367 SHA-256: d47f1e06ca597f80b6e3bf8846f5a91c57dbc474ef72ffa4e4907a84fb1722d2 download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/RADIOLARIA_WHITEPAPER.pdf READ_RADIOLARIA.xml bytes: 1488319 SHA-256: b3fc2c160eabf5d8e3ff342c7d1da1e6558dd5d50ea0576196b5e71a9459d033 download: https://github.com/AAkhtanin/hedgehog-os/releases/download/radiolaria-os-reference-g6b6f-v01/READ_RADIOLARIA.xml 9.2. Pinned source map All paths below were checked against the reference main tree. A source hash identifies its bytes; it does not promote an explanatory file into a normative contract. For a newer task basis, compare the actual current source and applicable contracts rather than restoring this snapshot automatically. To open a relative path at this exact snapshot, append it to: https://github.com/AAkhtanin/hedgehog-os/blob/62716374021e273b46878a824af39c1a94593cd7/ For raw text, append the same path to: https://raw.githubusercontent.com/AAkhtanin/hedgehog-os/62716374021e273b46878a824af39c1a94593cd7/ AGENTS.md bytes=24492; sha256=861181e63ee2fc9329e8541ff7ac89fc776a4b94489a9fe7a471ae179835db41 README.md bytes=18979; sha256=c0e304f82c90bf6ebccd3741f19751fc68ecfb9c59cc44fbc6db4ce0b5607609 LICENSE bytes=34523; sha256=0d96a4ff68ad6d4b6f1f30f713b18d5184912ba8dd389f86aa7710db079abcb0 COMMERCIAL-LICENSING.md bytes=960; sha256=50628fca63d697535669adfa1526d45fa7465e2f22a45707cc69f6fa56b023ee specs/current_architecture_lock_v01.md bytes=33255; sha256=0e122424aa2cd7f6b3156596f7038d4b46a5d9fc9b4c0962b565730ea5164682 specs/document_authority_index_v01.json bytes=115870; sha256=346c9bb839abc5c24f4bf55a02d46f6dae45ec87e9008ae48918aaa6a8b91d48 release/successor_context_manifest_v01.json bytes=125589; sha256=7a4ea0ae6d68b54f4ca29d5c7adc2c39fc73b0248a9269c65c0f1439984d9980 release/retired_architecture_inventory_v01.json bytes=6112; sha256=1f8fcc297ecc284dc6b69d51e38d4f849bd5bd794d14c523a6f5769dc4d0aee9 release/current_schema_surface_v01.json bytes=10166; sha256=a410bb8dc729b93b9c3d8b0d57a0d06f951e69ceaeb7cd1f5ea8e4c5a3d613cf docs/repository_transition_admission_v01.md bytes=15922; sha256=edadcf15e34a565a25f1ee575b13654c0f9d5bb6ba606ba5d5266ff1d4fa60a6 tools/reviewed_repository_transition_v01.py bytes=54993; sha256=932ac578d6e430ce15b8fc980d50babc0e69ebb3dfbf76c5d271c688fb1f13b7 tools/check_active_architecture_authority_v01.py bytes=851085; sha256=01a261aeb3c36bd1e079b02a315133b2b90f5d778544418b4af8491b83c57738 tests/test_repository_release_spine_v01.py bytes=221693; sha256=b36539a6a59658cb061a2c658963481231951ccdc7e17e25ead00175dc8132b0 tests/test_reviewed_repository_transition_v01.py bytes=27651; sha256=72f1fa69fc8d478926f9a1f5ad7b682388f1ba996ae54ceb949bf14a78ae2c16 docs/common_action_and_dynamic_composition_contract_v01.md bytes=89306; sha256=7f747d14065c11e6194fce94aabcd107693e7d27cd75c5558c37cef0a0a3092d docs/common_action_and_dynamic_composition_checkpoint_v01.md bytes=24546; sha256=fd708a885e0635e9f82bb8baad212f9cbd6db363888a8ae380893a718201e5e9 docs/drs_semantic_address_space_reuse_certificate_g2_b_preflight_v01.md bytes=104493; sha256=0b72a7794c0eee9d86d7745a304f7c4341d3e2c8b7ffece3665f16e38643a094 docs/gate3_outcome_feedback_contract_v01.md bytes=44833; sha256=abd975e70d3b92f14b283dc0fe961a50c9d0d12db4e907f682d2dc9d6fcf3862 docs/gate4_reference_contract_v01.md bytes=9603; sha256=2de0e065adcea5df395fe8a5b62088b9c9d13ff3f91275ad993cf51b53b205ec docs/gate5_reference_contract_v01.md bytes=5436; sha256=138f77869fed786a219d5ad0ac6313d8a82fafd33e975f4f32057371a58960b5 docs/gate5_reference_checkpoint_v01.md bytes=4083; sha256=a4df63f63d2839c37e4c2a7449bc7f8089fc543649af7eab0457cb47e6fc9d75 docs/gate6_reference_v01/README.md bytes=3444; sha256=c68370b3534baa57bed90531de710eef446e8404d5e2a76d52d6f7cdf5aa5618 docs/gate6_reference_v01/BOUNDARY_AND_FINDINGS.md bytes=6065; sha256=87e071e967752dee35178cc3d97ebb0cbe0cea7a5e9c37e66ca537dd122a59b1 docs/gate6_reference_v01/FINITE_COMPLETION.md bytes=4020; sha256=2d56e250dc735db9b78227e6ecd84ea3713344b3fa748b029a4723372808eced docs/gate6_reference_v01/TECHNICAL_REFERENCE.md bytes=4799; sha256=520a302b3a99b62958e3e6f9553cb2f3a1fb02544183508f317d42cc400658b9 docs/gate6_reference_v01/contract_index.json bytes=140195; sha256=0e0bd99f9384d5687bf4700743526d3b7655be5c5f362ef50a1bc2e09ff89444 hedgehog/kernel/abi_v01.py bytes=43017; sha256=cc1a660a3cdceb5c30728b27016e99a752561969d744744d55b18ce0af398d31 hedgehog/kernel/work_composition_v01.py bytes=89168; sha256=1f068c0d1156a312f36a7e55b7cbaf0e3e84503168b3305e10420bd8e3d3cbf0 hedgehog/kernel/root_decision_v01.py bytes=43352; sha256=cae47026ba5c8a52ca25b2774e7ac8e92eb4e312c2eb5d1776c3a16ee276409c hedgehog/kernel/effect_firewall_v01.py bytes=128809; sha256=46a453e0ef98565f0de59ab062b307eff6549a716973268a034592a5f98463bd hedgehog/kernel/transition_registry_v01.py bytes=142980; sha256=16d2d3c31004bccc3064c74290bad4a7303669b63c428cebf6be8763574ae0ca hedgehog/work_execution_host_v01.py bytes=52262; sha256=56b7ba85e8612df517d566418fa60a2512bcf9f6707877618a3d846010e6071b hedgehog/context_packets.py bytes=57747; sha256=080c69dfe24206549fbf27ae65141105c6e4b4c802b05296f1d90026823e9f32 hedgehog/capability_memory_binding_v01.py bytes=31977; sha256=6e4f2ca3be12a1e9a93d7b59318d9a6ba90ea6f43bbd9bccc81d8bdefe1b1b79 hedgehog/outcome_calibration_v01.py bytes=57704; sha256=ce4a29c7dfe0f536b53fce8435ac584509681bf1cef4511d46ac0bfd48e05079 hedgehog/gate4_pressure_budget_v01.py bytes=9665; sha256=768094bea0cc012046f6c1c514ccbf108e340932f86324d76c59cad0d6680d27 hedgehog/gate4_strategy_reference_v01.py bytes=6667; sha256=26d840964b017a508d15e8e9e039928dc9c9847aa186fb4940ea2a5b49b9de06 hedgehog/drs_g2b_compatibility_v01.py bytes=48485; sha256=5d4fb04cbf2e0819a0df5da17df724a0441747d2573fff924086dc9cfabaece1 hedgehog/domains/testflix/outcome_feedback_adapter_v01.py bytes=10595; sha256=c979f15b8dc093a9924060efc5db05917ccf20c86d55b1a50b738ac4fa25bde6 hedgehog/external_drs/gate5_exchange_v01.py bytes=26991; sha256=dfc22ae7beb39be05e7433bac7ceff9f2704e81c2638d0fe4727b2aa3eb36b07 hedgehog/external_drs/gate5_lifecycle_v01.py bytes=29240; sha256=627fa5171fbaad181c2d4e1c5e8047cc3b8b5957601a317049b05c980ae3de6f hedgehog/external_drs/gate5_native_v01.py bytes=20725; sha256=bba423e434d9fb9d5695b649191e9f31a2743b7bbe9327c80c0ce8b34162e3ec hedgehog/external_drs/gate5_supplied_v01.py bytes=10358; sha256=5cbb0e4851d65496e2097147b57a4bf285d36bdca9752f75db01694efe755f69 docs/gate5_authoring_kit_v01/START_HERE.md bytes=3627; sha256=c974139935b6e7596dd32ebb060761365ba3f2d284119d9e6d6c0fe8f9f425ba docs/gate5_authoring_kit_v01/HOW_TO_BUILD.md bytes=22495; sha256=092617adb77cd5f8b99bb9ac50fcfa97c3e8f3975ab48248678770c69477cb21 docs/gate5_authoring_kit_v01/SHARED_EVIDENCE_CONTRACT.md bytes=3649; sha256=f8bd6f6b3be330a5ee816b76ddb5581d2cf39d43a28f3712ae85c722656c2886 docs/gate5_authoring_kit_v01/EXECUTION_CONTRACT.md bytes=4252; sha256=77359dd95b91c3c8c6aa6645d7757276aa7adbaf23104703847136be0b23fd2c docs/gate5_authoring_kit_v01/PUBLIC_CLARIFICATION_G54D1_V01.txt bytes=1899; sha256=24f25ae2c9b1a9003172296def66c429b260ac0bdbf70641b782d330e767e080 docs/gate5_authoring_kit_v01/NATIVE_INTERFACE_MAP.md bytes=78644; sha256=9141ec43d5f50a62c6bfa4b4ebded7a89d9c14c89e76d1842d4df96de7c64575 docs/gate5_authoring_kit_v01/CANDIDATE_OUTPUT.md bytes=8546; sha256=89176fc7cc395cfb02407a6e2e6be87b3579740b403e2a06acadbe3c475cdd73 docs/gate5_authoring_kit_v01/WHAT_TO_BUILD.md bytes=8066; sha256=6e1a591c4369a9f37e5ffa727099c4c2ad0594895fd32e63b28f24f3538fd39f docs/gate5_authoring_kit_v01/FOOTBALL_PROFILE.md bytes=8263; sha256=01f5afd9cab361bef65a456ff612c84d33d9bb39e56ab5719f200fcf0de5ec53 docs/gate5_authoring_kit_v01/DEPENDENCIES.json bytes=1107; sha256=bf2527bce5d2d86b4e9918edf1f7f5911efd9b8b022bcb94771efff6a6ed6c48 docs/gate5_authoring_kit_v01/AUTHOR_SYSTEM.txt bytes=884; sha256=c485e9b00274020e86560f742aaea6c3f2167b8e29bd889623a4e138f2864738 docs/gate5_authoring_kit_v01/AUTHOR_TOOLS.json bytes=4103; sha256=dfdca2e497f7b8f1f0456470fdf34f3a98e210ce8abb3e2c05577940118ceb76 docs/gate5_reference_evidence_v01/member_index.json bytes=6550307; sha256=b3a313c5a2b6555c611e5efeeb9d10ffabe2d2e9af499f08f273ff34aa1c17c3 demo/verify_gate5_reference_v01.py bytes=9287; sha256=2852dda26988804f665181374ae9d2a44d96d395808f038707db4f02b7b63c22 demo/verify_gate6_reference_v01.py bytes=25211; sha256=2390e56ed1a886e9338ee14b49c41542583dc0ea573cdf4d399b08df2b0e1608 tools/run_gate6_football_reference_v01.py bytes=20399; sha256=7886cc9e9378766cf0741e13ff4f736b701c8ab60d808696722c5398cbe535af docs/showcase/radiolaria_os_v01/README.md bytes=5304; sha256=fa414a9b215f765ea0af3084e04a2846d6be6cd23269745ab577d0a5607948e1 docs/showcase/radiolaria_os_v01/RIGHTS.md bytes=2127; sha256=5fb2bd13733cac279dbed1313c1c79f486a5deb6103d5b8385d0e1f2f3016fec docs/showcase/radiolaria_os_v01/FONT_NOTICES.md bytes=7945; sha256=7be7ab3bad9fc876ed285f7aaf5dd1f3cbd9b0c40fe7a49b985a84f4fce99952 docs/showcase/radiolaria_os_v01/release/B5_ACCEPTANCE.json bytes=5628; sha256=ebf8f357a47ca3bb19802b220812b5dfd5e3719d4e26baec29ce1a29e8229983 docs/showcase/radiolaria_os_v01/release/IMPLEMENTATION_ROUTES.md bytes=6264; sha256=460373be1e021e6bb202aa6496d91f0a37d60f862fc09ce8225114a2df661ebc docs/showcase/radiolaria_os_v01/release/IMPLEMENTATION_ROUTES.json bytes=161738; sha256=1c10a0231116b34b4d9978f94dcc74ce23f4f1e3eaf028f82871eda6a9300ef5 docs/showcase/radiolaria_os_v01/reader/DEEP_SOURCE_INDEX.json bytes=471226; sha256=d1847a0fa57d01204b736604e871296923cebebe5721b2e12b68da151eed879e docs/showcase/radiolaria_os_v01/references/experience/source_ledger.md bytes=26401; sha256=9bfcf11a69a480adf2256769c45151c8f50ab68d20f66d554e2ca18cc52d5c30 docs/showcase/radiolaria_os_v01/atlas/README.md bytes=959; sha256=e23b4157ac06749499b7cd2803ea9ab975e849bed540ad5e84b18c4c2805660e docs/showcase/radiolaria_os_v01/atlas/OVERVIEW.md bytes=1400; sha256=7569a9ed081a2675d7a7f73447feb23e06e58acd2f40254f34dee45e5e143e29 docs/showcase/radiolaria_os_v01/atlas/ENTITIES.md bytes=18497; sha256=18b09f89e3f424ac0f8accb2652984620b4af9a545b557088e57cb59477ef151 docs/showcase/radiolaria_os_v01/atlas/views.json bytes=29792; sha256=5fb8345ee38a105f4e8a570f77f37eceed554e43823c53a5052fb0f05c7ec0ca docs/showcase/github_entry_v01/README.md bytes=1384; sha256=e33551761b6af7212b46d412dded61c68f0d36e7b3904c8e97640387b4c1f044 docs/showcase/github_entry_v01/PROVENANCE.json bytes=3373; sha256=1b3439055fb136aef2a78788cde5b538ac966d33fee418fc25809ce7880e3035 docs/showcase/radiolaria_os_v01/chapters/01_intent.md bytes=5026; sha256=d998d2596c9aea6e4188a1b31054827ba19bdd0cd7b218f3dcdbc2b32d27a18b docs/showcase/radiolaria_os_v01/chapters/02_geometry.md bytes=8226; sha256=605546c0c925a2e0227ce4bcee1c6df21ae96c17abadeb567a08700e7de78777 docs/showcase/radiolaria_os_v01/chapters/03_time_memory.md bytes=5821; sha256=1269173e193ed03b95b79c3de7d6808e8fa5e72665cc05da4798c86526c7c3be docs/showcase/radiolaria_os_v01/chapters/04_compute.md bytes=4029; sha256=2f529d440a476f85aee4e65f1b015d8039b06d953f47676890e23062c5d043b4 docs/showcase/radiolaria_os_v01/chapters/05_experience.md bytes=11280; sha256=5c361836ff5d4d5afe6bfcb3b1c51185f918e465ed4c69e266a809b25d8c70a6 docs/showcase/radiolaria_os_v01/chapters/06_method.md bytes=4325; sha256=03babc1005bf9eeb57f05aff82c380e3ba24e13597b2630a49263ce512cd2bf4 docs/showcase/radiolaria_os_v01/chapters/07_football.md bytes=5485; sha256=7565123162a213ceba58e94f1803d472c7138b2f0fd4dd7a9dc114fa935f87b1 docs/showcase/radiolaria_os_v01/chapters/08_wedding.md bytes=14960; sha256=309edae34fae20ddc7c247f8ceaba94334721cc0ee183b8e7ab5a35f3b96bdfd docs/showcase/radiolaria_os_v01/chapters/09_workspace.md bytes=12814; sha256=92bf2d71b9c8b718f90f61650b95bd087a55da72eb1205ebc3b9370b6e30b6e4 docs/showcase/radiolaria_os_v01/chapters/10_cross_case.md bytes=20578; sha256=bc7b3eb787f9345fc365acb5ffa62a34098990398c57cbfa4c175b9347e03764 docs/showcase/radiolaria_os_v01/chapters/11_integration.md bytes=3808; sha256=f2a2b28640e3a294f9f46e034601a2958e25d45b2018a734636989f821657a78 docs/showcase/radiolaria_os_v01/chapters/12_context.md bytes=8345; sha256=254c1983b03788a2102be0cf091ab465038019e4a03774874c596f03df7e492e docs/showcase/radiolaria_os_v01/chapters/13_verification.md bytes=4736; sha256=5c929271e9e8c50eaa771facc3a2153959be95186613ec3595f28312f7a8a31f docs/showcase/radiolaria_os_v01/chapters/14_identity.md bytes=4050; sha256=ee7d282e9f291b3eebc25ec2388e28ff995b257b2c67cdafd2a9375694972069 docs/showcase/radiolaria_os_v01/atlas/GEO-1.md bytes=1631; sha256=ef62aa1b7e4ab4477166bf25450314a400ec036293fdc15f8dbac8847c43e09b docs/showcase/radiolaria_os_v01/atlas/GEO-2.md bytes=1513; sha256=45379cae0235ac3fc82885468d88c2e3f57fe910aebb1586eb2f521c4c6acd96 docs/showcase/radiolaria_os_v01/atlas/GEO-3.md bytes=1629; sha256=b343de3b7297d82fa96aa1e5bcece571cb0587f0e2fa0977b70da380eb2ff87f docs/showcase/radiolaria_os_v01/atlas/GEO-4.md bytes=1039; sha256=2e0d34f9edb679eb8a18f3bb767ca33d6f0f0f55773504d6e4e20031f8c740c9 docs/showcase/radiolaria_os_v01/atlas/GEO-5.md bytes=1782; sha256=0847627d3b2ef4f05f01fdb4c5c3c2a471ff8e566c7f4239708b38178036262b docs/showcase/radiolaria_os_v01/atlas/GEO-6.md bytes=1287; sha256=6391eb923e098fe210f982e2f0d85027c8d4f48b1951b51767f651b34d13c7e9 docs/showcase/radiolaria_os_v01/atlas/GEO-7.md bytes=1202; sha256=4c9dce74792b900c19fef56481fdccafe9aafd81e7c84bfc5d672da4672a86a7 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/__init__.py bytes=74; sha256=e68644b065f0506d9c4a1b022ea7099cdeb209d5b712c566744946748569404d hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/domain.py bytes=13992; sha256=a98b895d5c6d59ee72d64f6beef97567e320eef4bfc986d8513ae7888abf3662 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/effect.py bytes=13939; sha256=f06c394747f8746dae844e787c760f07c782e9eb1dad918a6864aa2fa7803be7 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/memory.py bytes=8783; sha256=a6a3c1eebe8a068d945a4b249ce3d1dacba8c52ceca580fecbcbb3e7df209686 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/native.py bytes=5217; sha256=abfb69a8445614db75562f59a818b36d6a9c4f559960b5dc5a42f89d743a4707 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/policy.py bytes=5518; sha256=f9edbd5b8d2975496663d62e798d298c32c19c831fcfb55e7ae8ba776f585821 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/profile.py bytes=2764; sha256=4e3e36c6ba87c7156afdf739c0f86a2965379fe88ac3668b7f76b6f56978d8bc hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/publication.py bytes=10594; sha256=5df538e6d0c4ddc50cf2f836e23b41da268f147d2ddda1ec518a7468f3b0be61 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/run.py bytes=35816; sha256=0d56d79c2c0cdf4a9d410a4bf39331b5ad9e9fa8f7a9fbc2b1eabfc974f98f12 hedgehog/domains/football_pitch_booking/pack_v01/candidate_domain/storage.py bytes=915; sha256=85b2ab1b045f48eb816838a988999c4174c5a3f9bb26fb05da59e1a56b194e1b This list is a navigation and identity aid, not an instruction to import historical modules or execute an archived operator. The general kit is a frozen historical profile; its known differences from current source are explained in Section 04. 9.3. What was retained and what was generalized Retained: the accepted architecture, domain-development method, exact public interface names, producer/publisher/consumer contract, examination principles, action and refusal distinctions, source-version lessons, repository admission/recovery method, finite task templates, recorded result boundaries, reader routes and public identities. Generalized or omitted: individual names, conversational preferences, personal or commercial intentions, workstation locations, personal delivery folders, private return-archive dependencies and the original operator's live admission context. A new operator must use its actual authorized environment and review inputs. Public source and release URLs are retained because they identify the material being discussed. The adaptation was checked for English-only ASCII text, source-path resolution, preservation of the embedded source bytes, public/private separation and substantive architecture/domain/recovery consistency. This editorial check did not execute project code, run new readers, call runtime providers, change repository state or publish a website.
10. EMBEDDED GENERAL AUTHORING KIT DOCUMENTS
======================================================================
10. EMBEDDED GENERAL AUTHORING KIT DOCUMENTS
======================================================================
The five English documents below are included verbatim so that an assistant can read the original HOW rather than rely only on a summary. Each block has its public repository path, exact byte count and raw SHA-256. They remain frozen Gate 5 v01 documents present in the reference main tree. Instructions inside them describe their historical profile and do not automatically authorize a new task.
Their historical source basis is commit 76a6952c9d0f31c976d2501c5641457a20de153a plus the reviewed G52 proposal and listed G53 extension. Current accepted source differs in the two explicitly identified Gate 6 repairs. A new current-source kit must use a separately declared and reviewed closure; do not rewrite these old bytes or relabel a historical run as having used the repaired source.
START_HERE and EXECUTION_CONTRACT preserve profile-specific constraints, including the measured authoring and two-peer setup. Local pure-domain development does not acquire an obligation to reproduce that entire experiment. The clarification document preserves the exact profile requirement behind the historical fourth correction. It does not create a universal requirement that every rejected parse invoke Root.
This full public handoff includes explanatory results and review lessons. Keep it separate from a new blind author's inputs unless that exposure is explicitly part of the declared protocol.
BEGIN VERBATIM SOURCE | docs/gate5_authoring_kit_v01/START_HERE.md | bytes=3627 | sha256=c974139935b6e7596dd32ebb060761365ba3f2d284119d9e6d6c0fe8f9f425ba
# Domain Authoring Kit G54D V01
Revised public kit frozen before one new trial, not an acceptance record.
Read SHARED_EVIDENCE_CONTRACT.md and the separately marked FOOTBALL_PROFILE.md.
The source snapshot/public ABI and the original numerical/MOCK examples are unchanged.
The added numeric_context.py and numeric_work.py demonstrate only integer evidence
consumption. The private conformance runner and football examiner are not author inputs.
This is HOW to build one bounded, reviewed extension, not a task solution or an
installer. Read HOW_TO_BUILD.md, NATIVE_INTERFACE_MAP.md, CANDIDATE_OUTPUT.md and
the numeric/MOCK examples. The separate TASK_VISIBLE brief defines WHAT to build.
The approved source inventory is a read-only interface reference, not permission
to rewrite it. All public signatures are extracted from the shipped exact source.
Basis: accepted main76a6952c9d0f31c976d2501c5641457a20de153a plus the reviewed G52
external proposal and the explicitly listed G53 profile extension. The latter is
an external review proposal, not admitted owner source. The actual LICENSE is
provided unchanged: GNU Affero General Public License version3. No legal or
commercial entitlement is inferred from this reference package.
Use the existing configured Python/dependencies in a trusted reviewer environment.
Do not install packages, execute downloaded code, access secrets, provision a
sandbox, edit guards or manufacture source-admission context. A candidate remains
inert until independent source/import/I/O review AND an approved restricted runner.
The supervisor's current generic smoke is not a cold-author examination.
Set portable variables outside this kit; no personal paths are required:
```sh
export PYTHONDONTWRITEBYTECODE=1
export PYTHONPATH="$KIT/source"
"$PYTHON" -B "$KIT/example/run_numeric.py" "$NEW_OUTPUT"
"$PYTHON" -B "$KIT/example/check_effect.py" "$NEW_MOCK_OUTPUT"
"$PYTHON" -B "$KIT/example/reconcile_receipt.py" "$NEW_MOCK_OUTPUT"
"$PYTHON" -B "$KIT/candidate_contract.py" "$CANDIDATE"
```
The first command runs two actual native calculations with a changed input and a
refusal. The next changes only a disposable local two-record MOCK snapshot via
the native effect boundary, tests refusal paths, then reads its saved receipt.
These commands must not be used as an execution authorization for unknown code.
They create new output directories; preserve failures rather than overwriting.
The source snapshot is an AST-derived local import closure plus the actual Work
schema resource. An initial dry run identified that missing resource; it is now
included and the complete example succeeded from the snapshot alone. No .git,
owner context, secrets, target implementation or examiner fixture is included.
Static source may expose more functions than this kit supports. Only the mapped
public seams and stated finite donor profiles are claimed, not every imported
module or arbitrary callback behavior.
Numeric source contains no allocation, scheduling or booking logic. The MOCK donor
is the attributed G50 two-record example; its original IDs are retained for source
attribution, not authority. The G52 adapters are declared calibration donors:
their high-level role names and operation modes are not generic placeholders.
The author-visible manifest includes the public task but excludes private task values, examiner cases and supervisor
evidence. The complete return archive is for the supervisor/reviewer only; never
give that whole archive to the author. Freeze changes require a new kit/trial
version. Source admission and native capability admission are separate decisions.
END VERBATIM SOURCE | docs/gate5_authoring_kit_v01/START_HERE.md
BEGIN VERBATIM SOURCE | docs/gate5_authoring_kit_v01/HOW_TO_BUILD.md | bytes=22495 | sha256=092617adb77cd5f8b99bb9ac50fcfa97c3e8f3975ab48248678770c69477cb21
# HOW to Build a Bounded Domain: G54D V01
## 1. Architecture and Ownership
The public runtime is a chain of proposals, checks, local decisions and separately
owned effects. User/event intake belongs to a local Root. Semantic actors propose;
they do not own topology or finality. Runtime materializes Work. A result proposal,
Post V&V record or GT advisory remains evidence. Root decides locally. A current
Root-created scoped ActionCommitPacket is needed for consequential effects; the
Effect Firewall/bounded Corridor is the exclusive effect owner. Host binds current
source state, capability code and dispatch. External DRS transports evidence, not
Host handles, Root state or permission. There is no SuperRoot.
Use accepted modules; do not copy kernel logic into your domain. The shipped source
is read-only. An adapter may define a new closed body profile, pure domain functions,
typed parameter mapping, a finite MOCK executor and source-bound validators. It
cannot change ABI, Root rules, currentness, effect kinds, admission, crypto laws,
guard/binder policy, existing schema enums or arbitrary executable-plugin rules.
When a required operation lacks an honest public seam, report the exact symbol,
input and refusal. Do not emulate an ACCEPT or invent an undocumented function.
Three distinct decisions are required: independent source review for new code;
native capability admission for a concrete local executor/validators; current local
Root/Host authorization for an action. Neither passing own tests nor a foreign
signature supplies the other decisions. Candidate state is DRAFT, CANDIDATE_READY,
QUARANTINED_FOR_REVIEW, APPROVED_FOR_DECLARED_SANDBOX, then a separately recorded
LOCAL_REFERENCE_ADMITTED or REJECTED. Revocation only applies to a real admission.
## 2. Source Surface and Limits
Use NATIVE_INTERFACE_MAP.md/json for exact positional/keyword signatures, dataclass
fields, file paths, line numbers and source hashes. Required imports are included
in source/. `hedgehog.kernel.work_composition_v01` reads
`schemas/work_composition_v01.schema.json`; this resource is part of the inventory.
The dependency record describes the measured interpreter and installed packages,
not a request to upgrade them. Examples use only existing dependencies and stdlib.
The reference transport is two peer processes over inherited, length-framed pipes.
It is signed, not encrypted, internet-ready or isolated from an OS administrator.
Each peer owns its key capability, state, local policy, Root and Work. The trusted
launcher distributes synthetic inputs and separately pinned PUBLIC keys, not an
offer, answer or cross-Root capability. Different peer roots never share finality.
No arbitrary URL/path from payload is opened. There is no HTTP redirect mechanism.
Bounds: one automatic hop; four pointers; eight records per opened bundle; sixteen
lineage references; pointer16KiB; body64KiB; frame256KiB; JSON depth12. Each declared
task has at most two payload attempts, eight metadata/status observations, two
accepted bodies and128KiB cumulative accepted body bytes. The finite G52 sessions
have at most32 requests, clean boundary EOF and signed CLOSE. A malformed/truncated
frame is not clean EOF. Existing stalled IPC read ceiling is60 seconds, a finite
donor operational bound; changing it requires an explicit versioned profile.
## 3. Build Actual PURE Work
The concrete example is `example/run_numeric.py::calculate`. Its executor computes
3*x+2, while a distinct validator checks typed output and the equation. Input7 and
input11 produce23 and35, respectively. No expected numeric result enters bootstrap.
Boolean input is refused before a Work is constructed. A missing required input
also produces a real Root refusal, not an offset/default answer.
1. Define module-level executor and input/output validators, not anonymous closures
created from text. An executor receives a BoundCapabilityInvocationV01 and
returns a tuple of ActionEffectParameterRecordV01 from the public builder.
2. The input/output schema is expressed using exact names and value_type strings
(INTEGER, TEXT, REFERENCE, BOOLEAN as appropriate). Optionality, consequential
flags and allowed values are native fields, not free-form hints.
3. `mocks.admit_pure_operation_v01` observes local code and admits the exact PURE
definition through Host. That donor is finite controlled reference code, not
permission for downloaded code or an OS sandbox.
4. Make `work.WorkItemV01(work_id, definition_id, owning_root_id,
inputs, resource_refs, depends_on, guard, review_obligation_id)`.
Literal bindings come from `donor.work_literal_v01`. Binding an earlier actual
Work output uses the public WorkOutputBindingV01/WorkInputBindingV01 shapes in the map;
do not replace it with a caption or hardcoded fixture value.
5. `donor.prepare_work_program_v01` returns `(program, common, host)`. It constructs
the semantic proposal, candidate, topology and controlled source/Host, then
`advance_work_program_v01(program, **common, host_map={root_id: host})` invokes
the actual local executor. Call the supplied program-result validator with the
same live values; then `work_program_result_to_artifact_v01` produces the artifact.
6. Preserve invocation, admission snapshot, typed inputs/output, artifact, status,
source references and actual host.work_attempts. A nominal function called
"Work" or a summary JSON alone is not evidence of native performance.
The donor's Work source clock is controlled logical1014, not current Unix time.
Its preparatory action stages use1010..1014. Do not substitute wall time for one
axis or claim its logical envelope is today's observation. Domain
source envelopes and current external reviews use explicitly recorded UTC. A real
external input's validity is checked separately at the domain boundary. Later live
deployment needs its own supported current source, not an unbounded logical fixture.
`gate5_native_v01.native_work` offers only source/mean/corrected calibration modes.
It is a declared donor, not a function that accepts a new domain mode magically.
Use the lower public Work/capability pattern for new functions, without changing
that donor or forcing data into its calibration fields. All strings that reference
IDs are canonical identifiers, not permission tokens.
## 4. Local Root Review
`gate5_native_v01.root_review(root, transaction, candidate, subject, checks,
claim_value, now, predicate=..., window=None)` is a working controlled adapter.
It constructs public semantic request/evidence/contribution/trust/synthesis/packet
records, then invokes the actual Root kernel and validates its result. Inspect its
source and mapped underlying public builders when adapting to new semantics.
`checks` must be actual independently computed boolean predicates. Supplying True
because a foreign record says so is not validation. `claim_value` must carry exact
consumed inputs, Work output/ref and dependent source identity. The source-bound
Root input/result is retained using `root_plain`; JSON copies cannot be revived as
new grants. Local acceptance for reading a foreign body is not effect permission.
For the controlled adapter, `now` is the recorded integer UTC observation and
`window` is a half-open validity interval. The helper compares the actual wall clock
with that interval. All checks plus temporal validity are required for ACCEPT.
Preserve refusals with their real reason and phase; never rename a failed call as
PASS because setup or teardown completed. Source review, publication review,
release/access review, retrieval review, import acceptance and final result review
are distinct subjects. Link their exact IDs rather than duplicating one conclusion.
## 5. Persisted Pointer Lookup and Descent
LocalDRS is storage, not an authority service. `write_semantic_record` stores a
SemanticDRSRecordInput; `resolve_semantic_candidates` accepts SemanticResolveQuery
and returns candidates requiring local review. Reconstruct a new LocalDRS instance
to exercise persisted reads. A process-global Python map is not restart evidence.
The address layer carries only safe metadata: publisher, source/revision, declared
meaning/schema, body hash/length, time/scope and an operator endpoint alias/object.
Do not put source observations, full schedule/body, private notes or complete native
trace in a public summary. Do not put hashes into natural-language summaries merely
to evade a privacy filter; content_fingerprint binds exact metadata separately.
For native descent, build SemanticAddressV01, DRSTimeEnvelopeV01 and authority
envelope through their public builders. Start as CONNECTOR_OBSERVATION, not accepted
memory. A later locally reviewed context can refer to its own Root input/result
while keeping foreign identities in source refs. Existing MemoryPointer storage
classes are LOCAL_*; external metadata is not a new remote enum or a fake local
file pointer. The explicit provider bridge opens the exact permitted remote body.
The working numerical donor `pointer_descent_v01(pointer, folder, source_end)`
shows the complete public query/evaluate/retrieval-plan/Root/descent-request/
execute_local_memory_descent sequence. It is calibrated to its declared Root/domain
labels, not a general new-domain entrypoint. Reuse that sequence in a reviewed thin
domain adapter with its own current identities and exact fields from the interface
map; do not suppress its source checks or rename foreign owning_root_id as local.
Source end bounds the context. A historical query labels original source/time and
may have a new local READ review, but no new remote body fetch or computation.
Completely pure supplied verification makes no new Root decisions either.
## 6. Canonical Wire and Identity Order
`gate5_contracts_v01` supplies decode/canonical/sha/shape, bounded arithmetic,
time validation, public commitment signatures and the closed carrier field sets.
The machine interface index includes FIELDS, IDS, bounds and exact callable shapes.
`decode` rejects duplicate keys, floats, NaN/Inf, noncanonical bytes, excessive
depth/size and unsupported JSON scalars. Integers use an explicit signed bound;
bool is not int. References/labels are bounded ASCII. Do not lossy-normalize an ID
into a path. Use allowlisted logical IDs, not caller paths. IDs hash the canonical
body excluding their own ID field. Lists keep their order; sets must be normalized
deliberately, never sorted because a comparison failed.
The shared contract is **DomainResult E -> PublishedEvidence B -> ConsumptionContext C**.
Read SHARED_EVIDENCE_CONTRACT.md for independent expected-input derivation.
E is the actual native semantic output, not a future signed carrier containing its
own Work/review identity. A reviewed domain profile defines payload_projection(B)
and a deterministic context mapping. Signature validity is not semantic equality.
Validate E's actual selected values and no-result status against original inputs
independently of the consumer answer. A correct consumer cannot replace a missing
or invalid E. Join the authenticated release decision reference to the actual
release Root result, not only another signed field. Historical E keeps its
original computation and use time; later current review is separately checked.
Construct dependencies in this order:
1. Independently recorded source inputs -> native Work returning E -> source review.
2. Closed B with payload_projection(B)==E and original source record/revision/Work/
review/time/lineage fields. Bind the original inputs independently, not by copying
a candidate's proof.inputs into the expected-input argument.
3. Content manifest covering `body.json` length/hash and original source refs.
4. Pointer tying body and manifest; its publication/release records are separate.
5. Domain-separated POINTER/STATUS/RELEASE commitment+signature over exact bytes.
6. Import candidate retaining foreign IDs plus local request, verification facts,
dependency/policy fingerprint and status observation. Immutable candidate first,
separate immutable disposition/review later; do not introduce a hash cycle.
7. After public signed bundle/status validation at the independently recorded local
use time, construct C from that B, authenticated STATUS and local request/policy.
Native consumer Work must consume exactly C. Its output and subsequent local
Root review bind the original request, signed B and actual consumed result.
Keep the signed originals; a summary or C itself cannot authenticate its origin.
Manifest never contains its own hash, pointer or signature. Pointer includes the
prior source review, not a future review that includes the pointer. SignatureValid
proves only the pinned test key and linked bytes, not arithmetic or policy truth.
Keysets are accepted from independent local trust input, never from the same body
advertising itself. `sign`/`verify` call the existing Ed25519 commitment primitives;
no custom cryptography, no private key export or key hash in logs. Production PKI,
enrollment, encryption and administrative isolation are not claimed.
The generic numeric signed-profile subtest is explicitly a component crypto check,
not a peer release or grant. Full peer execution is the separately source-bound
recorded G52 donor, not repeated for prose or recast as fresh new-domain execution.
## 7. Add One Reviewed Body Profile
The G52 default is intentionally closed to BoundedCalibrationSummaryV01. Changing
only a schema_id in JSON cannot turn it into another type. The G53 correction adds
`ReviewedBodyProfileV01` and keyword-only `profile=` on `validate`, `check_pointer`
and `check_bundle`. Old calls without profile remain strict calibration calls;
the frozen calibration profile cannot be replaced through this seam.
After independent source review, trusted LOCAL deployment code may instantiate one
profile with exact schema_id, closed body_fields, closed scope_fields, local policy
acceptance field and three module-level local functions:
- validate_body(body): check all domain types, ranges, equations and closed meaning.
- validate_pointer(pointer): enforce safe metadata-only summary; no full payload.
- validate_binding(body, policy, pointer): check domain semantics against the local
task, units/recipient policy and coherent source projection as applicable.
These callbacks raise a specific ValueError on failure; they do not return a trusted
boolean from the wire. They are part of the reviewed TCB, not proof that arbitrary
caller code is safe. Profile construction is not source admission. The record is
not serialized and there is no module-name loader or network plugin registry.
Explicit profile selection occurs from separately reviewed local source/policy,
not from a payload-provided import path or self-selected validator. Unknown type
or schema mismatches fail. The example's `numeric_profile.py` is a concrete working
non-target profile with independent arithmetic validation.
Common required body fields remain version, source_record_ref, source_revision,
source_work_ref, source_review_ref, time_envelope, ttl_base and source_lineage_refs.
Generic validation still enforces bounds, exact identity, provenance, signatures,
request/release binding, freshness/status/conflict and local policy. Schema-specific
fields are NOT smuggled into calibration observation/correction fields. Use
`validate(kind, value, profile=PROFILE)` before canonicalizing/signing. The old
`contract(kind,value)`/BoundedContractV01 remains the strict calibration carrier;
it has no implicit profile. Do not use it to silently accept a new schema.
The G52 high-level source/requester lifecycle is a calibration donor. A new domain
owns its body producer and consumer adapter using the same PipeChannel,
PublisherBudget, signature/status/temporal/current-review primitives. It does not
rewrite transport, Root/ABI or invent federation. ImportStore's calibration
convenience `validate_bundle` is not profile-aware; call public check_bundle with
the explicit reviewed profile, then preserve its authenticated state/dedup rules
in the domain adapter. Do not monkeypatch BODY_SCHEMA, functions or global policy.
New schema/adapter changes are part of the exact candidate diff for source review.
## 8. Currentness, Refusals and Accounting
Use the public DRS TimeEnvelope builder and original axes. End is
min(valid_to, pt_created_at+ttl_seconds, source_observed_at+local_max_source_age).
Use is valid only for valid_from<=t<End and source_observed_at<=t. Ingestion or
verification time cannot renew source life. Status independently requires
checked_at<=t<min(valid_until,checked_at+60), matching request/subject/nonce,
pointer/stream/publisher and effective_at<=t. Actual refresh does not rewrite an
old signed timestamp. Store authentic terminal entries before current-use refusal.
Persist highwater by publisher,pointer_id,stream_id. A lower revision is rollback;
same revision/different immutable entry hash is equivocation; same entry with a
new freshness envelope is not conflict. REVOKED/SUPERSEDED do not become ACTIVE
after replay/restart. Distinct pointers' streams are independent. Status unavailable
is CURRENT_STATUS_UNKNOWN, not a permanent peer blacklist. Historical source/result
bytes survive. Changed local policy/dependencies require a new appropriate review.
Use `PublisherBudget(path, contexts)` on the source side independently of B's
`Budget`. Contexts are operator-established before execution, binding requester,
object, purpose, use and source revision. Reserve authenticated attempts before
release review; reserve allowed disclosure bytes before constructing/sending body.
Count denied responses separately from bytes released. Never split an exhausted
task just to reset counters. Helper reconstruction/restart reopens the same state.
The source profile is finite single-writer, not a cross-process database lock.
The receiver checks frame size before reading, then typed body size/integrity and
cumulative accepted-body budget. Wire bytes and semantic body bytes are distinct.
Keep reciprocal bytes and attempted-operation evidence. After refusal, show a
lawful current input succeeds. Dedup key is publisher/source_record/revision/hash;
same identity with different bytes is quarantine, not best-score selection.
No duplicate import earns independent-source or history/reputation credit.
## 9. MOCK Effects and Recovery
The exact G50 donor is `mock_effect.py` plus `check_effect.py`, with a minimal
import-only shim to the reviewed native modules. It changes two disposable local
records, not real money/device state. Functions are observed by
`observe_local_capability_code_v01`; build business semantics, field definitions
and native capability definition, then `admit_local_capability_v01` with the exact
validators/executor. Only MOCK_CONSEQUENTIAL is used. Do not add an effect kind.
Resolve typed inputs first. `build_native_work_action_v01` derives a canonical
packet from admitted business semantics, root/transaction/business object and
input values. Its donor expects device_ref plus declared bindings and controlled
1010..1014 source time. `authorize_and_prepare_action_v01` performs the real Root
review and lifecycle preparation; `host_for_prepared_action_v01` binds the live
trusted source. Dispatch requires current revision and exact evaluation context.
The native firewall checks candidate/packet/owner/inputs/code before the callback.
The MOCK executor binds a logical key from canonical.idempotency_identity, not a
model string, to exact payload hash and packet. It atomically replaces one local
snapshot containing both records and key map, then returns typed output. The
native registry owns the actual EvidenceReceipt and consumed lifecycle. Save them
before optional rendering. Check actual state readback, not just "COMMITTED" text.
The donor tests stale Host revision/time, foreign owner, changed payload and exact
duplicate of the terminal packet. A duplicate dispatch REFUSES; native Host does
not automatically return a second successful receipt. Do not retry CONSUMED.
`reconcile_receipt.py` demonstrates application-level saved MOCK receipt/readback:
same key+same payload+matching stored snapshot returns the old receipt as HISTORY;
changed payload refuses; unknown/missing/mismatched state yields
UNKNOWN_REQUIRES_READBACK and dispatch=False. This does not reconstruct a Host,
grant or prove hostile-storage/power-loss recovery. In uncertain real execution,
stop retry, inspect supported independent state/readback and reconcile under a new
current review if needed. Compensation is NOT_APPLICABLE for the disposable
all-or-none one-file mock, with no cross-owner rollback or real-world transaction
promise. A domain must state its own exact failure policy, not claim universal
atomicity because os.replace appeared in a sample.
## 10. Candidate, Tests and Handoff
Fill every Authoring Passport field in CANDIDATE_OUTPUT.md. Own tests cover a genuine
positive, altered input, refusal and a lawful continuation. Unknown body fields,
bad arithmetic, source/time/scope mismatch and removed producer output must not
fall back to an expected caption. Keep profile validation independent of the
producer's success flag. Include a source impact/dependency list and exact limits.
The package checker checks data/path shape only; it never executes the candidate
or awards acceptance. A complete independent source/import/I/O review precedes
the first candidate execution. The current environment has not been certified as
the required cold-execution boundary. A plain directory, venv, PURE admission or
static code scan is not OS isolation for arbitrary Python. No sandbox experiments
or tool escalation are instructions to the author. Await the separately reviewed
runner. Existing native guards stay active in trusted execution infrastructure.
For future trial: initial response plus at most three specific test-feedback
corrections. Architecture hints are recorded assistance. A kit defect versions the
kit and preserves the failed trial; at most one additional trial without a new
scope decision. Report observable inputs/outputs/tool accesses, not hidden reasoning.
Never include credentials, old chat, examiner tests or target answers in author
context. Runtime semantic-model calls, author-model calls, native Work and transport
counts are separate. Native/result verification is not installation or Gate closure.
END VERBATIM SOURCE | docs/gate5_authoring_kit_v01/HOW_TO_BUILD.md
BEGIN VERBATIM SOURCE | docs/gate5_authoring_kit_v01/SHARED_EVIDENCE_CONTRACT.md | bytes=3649 | sha256=f8bd6f6b3be330a5ee816b76ddb5581d2cf39d43a28f3712ae85c722656c2886
# Shared Evidence Contract G54C V01
## Causal Objects
1. DomainResult E is extracted from a real typed native producer result whose
invocation is checked against independently held original source inputs.
2. PublishedEvidence B is built afterwards. A closed, reviewed domain profile
defines payload_projection(B)==E and any original-input relations. B retains
original source record/revision, actual Work artifact and source review IDs,
original time axes/lineage and domain request binding. Independent supplied
checks validate the real source review, not just an ID-shaped string.
3. ConsumptionContext C is built only after exact canonical B/manifest/pointer,
publisher signature, release, request/audience/use, current signed STATUS,
local policy, half-open time, conflict and history checks pass. Its closed
domain mapping uses verified B/STATUS and independently held policy/request/time.
It may add derived facts unavailable when E was produced. It must not rewrite B.
Source review, publication review, release review, retrieval/import review and
consumer result review have different subjects. No future identity is an input
to the Work that creates it. Retain signed originals outside C. A public-key
signature authenticates bytes and the independently pinned test key, not semantic
truth, current local authorization or a consequential effect.
## Independent Verification
The examiner receives original source inputs and clock/policy/bootstrap observations
from a trusted runner, outside candidate output. It validates producer admission,
invocation and result, derives E, checks actual source review/refs and independently
compares the published payload to E. It validates exact publication bytes and
request-bound STATUS through the existing public component primitives. It then
derives expected C without reading candidate consumer inputs, and passes that C to
the public supplied native validator. It compares consumer output to the reported
result and subsequent Root claim. Candidate input/output agreement alone is not
this proof. A re-signed wrong payload, valid stale/wrong STATUS or a genuine Work
that consumed the wrong C must fail the appropriate independent predicate.
Original highwater/conflict and current policy belong to the local observer.
Unknown/unavailable current STATUS is not ACTIVE. checked_at <= use_time <
min(valid_until, checked_at+60); source life and request deadline independently
bound use. Verification/ingestion does not renew historical evidence. A new
independent supplied check creates no new Root decision, Work or permission.
## Finite Profiles
Names such as offer_json and offer_sha256 are profile fields, not kernel laws.
See FOOTBALL_PROFILE.md for the task-specific mapping. The non-football numerical
example uses B.computed as E, binds B.sample to the actual source INTEGER input,
and consumes C through one typed PURE capability. It computes
B.computed + B.sample + C.source_projection.request_revision; it does not select
resources, expose an allocation algorithm or provide a target answer.
The numeric mapping is exactly implemented in example/numeric_context.py. It
requires the existing numeric ReviewedBodyProfileV01, actual public bundle/status
checks, independently pinned local policy and use time. It is a signed component
example with peer_io=0, not a second peer exchange, release authority or effect.
Native Work uses the existing controlled logical 1014 source clock; external
evidence/reviews have separately recorded UTC times. No Host authority is revived
from serialized proof, and no protected source/API is changed by this contract.
END VERBATIM SOURCE | docs/gate5_authoring_kit_v01/SHARED_EVIDENCE_CONTRACT.md
BEGIN VERBATIM SOURCE | docs/gate5_authoring_kit_v01/EXECUTION_CONTRACT.md | bytes=4252 | sha256=77359dd95b91c3c8c6aa6645d7757276aa7adbaf23104703847136be0b23fd2c
# G54 Precreated Peer Contract V01
This clarification is frozen before any author request. It replaces only process
startup, not domain/Root/transport rules. No target solution is supplied.
Implement `run_peer_v01` with EXACT keyword arguments documented in
CANDIDATE_OUTPUT.md. Import must have no Work, Root, transport or effect side effects.
`role` is `publisher` (source owner) or `requester` (local consumer). Each peer has
its own process and fresh native Root/key instances, its own Path `state_root`
and `output_root`. It must not inspect another role's filesystem or private state.
The operator supplies that role's JSON as `input_data`; it contains:
```json
{"version":"g54.peer_input.v01","role":"publisher","events":[],"policy":{}}
```
`events` is a finite ordered list (at most8). Every event has `event_id` and
`request` using the exact normalized public fields in CANDIDATE_OUTPUT.md. Publisher
events additionally have `inventory`, the current source-owned normalized inventory.
Requester events never receive inventory, an expected offer, or a solution. A
`semantic_proposal` and its `semantic_capture` may be attached to an event; define
their closed domain shape and validate/actually consume them, without letting the
proposal invent consent. `policy` is explicit operator-owned local policy. Document
your closed keys for audience/release/refusal controls and booking approval.
The independent runner can supply changed current inventory on a later event.
Original request history and prior state must remain preserved. Process events
in order; return after this finite list, with no infinite service loop.
`channel` is the actual existing
`hedgehog.external_drs.gate5_exchange_v01.PipeChannel` instance. Use its documented
`send(kind,value)`, `receive(allow_eof=False)` and `counters()` methods; messages have
the existing closed framing/kinds, max256KiB, canonical JSON and peer timeout.
Use existing Budget/PublisherBudget and contract/profile signature validators;
do not reimplement the external DRS protocol or hide I/O inside PURE executors.
`bootstrap` is a separate preassigned inherited PipeChannel for initial PUBLIC
keyset exchange only, bound to these two reviewed peer processes by trusted setup.
Generate each private signer inside its owning process, send only public projection
over bootstrap, verify expected peer Root identity from operator policy, pin that
keyset for subsequent data-channel validation. Never install keys from a body or
data-channel pointer. Bootstrap is test setup, not Internet PKI or an authority grant.
Write one event directory `output_root / event_id` with report.json and necessary
native/signature/review/receipt/source evidence. Event IDs must be safe local names.
Preserve requested objects for the private examiner: offer_body.json,
venue_work.json, requester_work.json and requester_review.json where applicable.
The file format for native Work is the exact generic producer/consumer format
shown in HOW_TO_BUILD and CANDIDATE_OUTPUT. A role may write only its own state/output
directories. Returned dict/None is completion data, never a success certificate.
The trusted parent separately records OS wait status and native function events;
candidate-written counters cannot replace this observation. Unknown failures must
remain observable, not silently turned into expected domain refusals.
Runtime: pinned Python3.14 Linux arm64; jsonschema4.26.0/cryptography48.0.0; no package
installation. PYTHONPATH contains exact kit/source, generic examples and candidate.
No network, subprocesses/threads, new imports from unapproved paths, dynamic code,
monkeypatching, protected-source writes or direct effect outside native Firewall.
Read-only code; own disposable outputs;512MiB aggregate memory, one CPU,16PIDs,
30/35 CPU seconds per worker,90second wall,16MiB/file and64MiB output filesystem.
Reviewed source checks must work on the immutable candidate module files.
This is finite reviewed Python, not a hostile-code or OS-administrator isolation
claim. The candidate must not introspect/disable the trusted observer or use other
process file descriptors. The reviewer checks source and native artifacts; a JSON
claim named TRUSTED is not accepted as independent telemetry.
END VERBATIM SOURCE | docs/gate5_authoring_kit_v01/EXECUTION_CONTRACT.md
BEGIN VERBATIM SOURCE | docs/gate5_authoring_kit_v01/PUBLIC_CLARIFICATION_G54D1_V01.txt | bytes=1899 | sha256=24f25ae2c9b1a9003172296def66c429b260ac0bdbf70641b782d330e767e080
The owner has explicitly authorized one additional correction of your existing
G54D implementation. This is correction 4 and the fifth submission of the same
development lineage. The original initial-plus-three trial remains INCOMPLETE.
Preserve the working implementation; do not rebuild the domain or change the
supplied runtime/source. Only you may write the candidate repair.
The final review found that a valid request receiving authenticated non-active
STATUS ends in an honest REFUSED report without an actual local Root refusal
decision. In particular status_not_active enters the known ValueError handling
path in requester_event. Native observation saw the preceding metadata/retrieval
decisions, but no decision for this refusal. There was no booking execution or
registry mutation. The last feedback before your third correction did not report
this missing decision; it reported the publisher result and source-reference
problems, which you addressed.
Public clarification for this profile: when a valid request is rejected after
verified non-active STATUS, preserve the authentic terminal state, stop current
use and effects, obtain a genuine local Root refusal and bind it to the original
request, actual reason and corresponding source/status evidence. The completed
REFUSED result must have that actual basis. A previously accepted metadata review
or a fabricated serialized decision is not this refusal. Do not revive expired
permission or import the publisher's authority. Unknown errors must stay visible;
possible post-effect uncertainty must not be converted into a clean no-effect
refusal. Preserve independent roles, history, receipt/state behavior and all
existing native checks. Make only the related candidate changes and necessary
local tests/documentation. Submit one corrected complete pack through the inert
broker. Do not execute it or claim acceptance from syntax.
END VERBATIM SOURCE | docs/gate5_authoring_kit_v01/PUBLIC_CLARIFICATION_G54D1_V01.txt
END OF PUBLIC WORK ASSISTANT HANDOFF
Next action: relate the concrete task to the applicable source contracts and proceed within its authorized scope. Reading this file does not start a model run, effect, repository transition, release or recovery operation.