An outside AI built the domain. It could not admit itself.

A human-defined construction method, HOW, and a separate Football task, WHAT, were given to an outside gpt-6-astra author. It returned a program that computes offers, checks another owner's evidence, supports separately permitted mock booking and keeps checkable history.

The candidate went through independent examination and corrections before owner acceptance and source admission. This is a concrete transfer of construction knowledge: another author can return an implementation for review, while acceptance remains the owner's responsibility.

Accepted G54D1: 16/16 native cases, 56 events and seven isolated mock executor starts. Those are acceptance results for the authored program. The three Gemini runtime proposals are a separate experiment and a different model role.

HOW and WHAT gave the author different kinds of knowledge

Before: the core and its supported interfaces already existed. The outside author received a construction kit, an approved read-only interface snapshot and generic examples. Change: a separate Football task specified the business constraints and evidence obligations. After: the author produced a candidate implementation for independent examination and later source admission.

The generic numerical example explains 3 × x + 2; a separate mock example explains bounded effect and receipt behavior. Neither gives the author the winning football slots. The kit's machine interface map supplies concrete signatures and source identities, not an unrestricted right to call every function in a source tree.

The author could edit a candidate; it could not install it

The broker exposed seven bounded operations: read and search approved source, read candidate files, write or patch candidate text, inspect text syntax, and submit a frozen package. A patch had to name the expected whole-file hash and a unique literal replacement. The broker generated the package manifest.

Source review preceded bounded candidate execution. The trusted runner created publisher and requester processes with their own inputs; the private examiner checked independent observations. The author did not receive examiner code, private fixtures or a ready-made repair patch. This is a declared access surface for a reviewed Python reference profile; it is not a universal sandbox or a claim about the model's pretraining.

The examiner was corrected too

The history contains two different failures that should remain distinct.

G54B1 reached useful two-owner computation and a mock effect, but stopped at an impossible causal demand. The old examiner required the complete published body at both Work boundaries even though the instructions required that body to be assembled after producer completion and review. The producer could not already know its future review identity.

G54C corrected the general evidence contract: E is the computed result; B adds publication and review context; C adds the recipient's checked current-use context. The requirement is preservation of the actual result and independently reconstructed causal links. In Football, B.offer = E and C.body = B. The contract was tested with a separate numerical profile before a new cold author trial. B1 remains INCOMPLETE; the correction does not rewrite that result.

G54D then reached 15 of 16 native cases. A revoked source correctly stopped use and booking, but one exception branch omitted the required genuine local Root refusal. An earlier ACCEPT decision was not that missing refusal. This was a real candidate obligation, distinct from B1's examiner defect. The original initial-plus-three correction budget ended with INCOMPLETE.

One explicit fourth correction completed the missing decision

The public clarification and protocol amendment allowed one further correction, index 4 and submission 5. Resuming the old ephemeral thread was unavailable. The process therefore declared RESTORED_CONTINUATION: the same public kit, the author's own final source, earlier delivered feedback and the narrow new diagnostic were supplied to a new turn. Identical hidden model state was not claimed.

G54D1's repaired branch records actual BLOCKED_FAIL_CLOSED Root decisions bound to the revoked source and original requests. The refusal binding record gives observer call sequences 29 and 66, application reason status_not_active, separate request hashes and source/status references. These are observed calls joined to evidence, not a decorative refusal label written into a report.

The same corrected candidate then passed all sixteen native scenarios, including useful variation, permission failures, source revocation, history and continuation. Seven isolated mock executor starts belong to distinct controlled paths; they are not seven repeats of the principal booking.

Exact source bytes connect the author to the accepted candidate

The authorship reconstruction starts with 22 exact G54D author files and applies the recorded edits. It reconstructs 23 author files plus the coordinator-generated manifest, for a 24-file package. The production delta is in candidate_domain/run.py; the author also adds a refusal test and updates explanatory files. Tests read as source are not counted as separately executed tests.

Selected exact fields from docs/showcase/gate5_reference_v01/source/g54d1/AUTHOR_TRIAL/attempt_04/authorship_reconstruction.json, publication ref 800f38c3407b3f4ec93dc943c977ffeecb614881:

/status = "EXACT_SEED_PLUS_AUTHOR_EDITS"
/correction_index = 4
/submission_number = 5
/submitted_packs = 1
/supervisor_implementation_edits = false
/coordinator_manifest = "EXACTLY_RECONSTRUCTED"

These recorded source relationships support attribution. The supervisor supplies task material, feedback and review machinery; it does not silently implement the candidate correction and credit it to the model.

Acceptance, runtime participation and admission stay separate

The primary NATIVE_ACCEPTED.json records /case_count: 16, /event_count: 56 and /mock_executor_starts: 7. Its status is a source-bound native acceptance prerequisite, not a claim that every later publication step was already complete.

The separate live series consumes three genuine gemini-2.5-flash proposals for EXACT, SHIFT and VERIFY, followed by five native events and one further mock executor start. All proposals were precollected; the VERIFY model did not observe the receipt created later. Controlled owner input supplied confirmation and repetition. This distinguishes authoring a new implementation from interpreting a runtime request.

G55L subsequently admitted and published the reviewed engineering bytes through a separate owner process. The authoring dossier, A11 records the engineering commit 64f2b38b0d671e4907bae8649fff8f3d2ee0751c. Historical preparation documents retain their earlier PENDING fields. Source admission does not turn a manifest into permission for today's business effect.

This is one bounded transfer result. It establishes neither a universal time or cost of integration nor an integrated automatic Capability Synthesis Factory. Restored continuation also differs from a new blind-author trial.

Inspect the selected evidence

An external author produced the accepted continuation

G54D1 records an external gpt-6-astra author and 16 native cases, 56 events, 7 isolated mock starts under an explicit amendment.

Record: G54D1 correction4 submission5. Class: historical acceptance. Scope: G54D1 correction4 submission5; historical published evidence read as data, no new core execution or authority.

View source — G54D1 correction4 submission5

Recorded edits reconstruct the submitted source

Exact seed plus author edits reconstructs the accepted candidate, with a separately generated coordinator manifest.

Record: G54D1 author source reconstruction. Class: historical source reconstruction. Scope: G54D1 author source reconstruction; historical published evidence read as data, no new core execution or authority.

View source — G54D1 author source reconstruction

The final correction obtained the missing local decision

The revoked-source events now contain actual native refusal calls, causally bound to the authenticated status and original request.

Record: G54D1 p7_candidate_7071661f. Class: historical consumption verification. Scope: G54D1 p7_candidate_7071661f; historical published evidence read as data, no new core execution or authority.

View source — G54D1 p7_candidate_7071661f

Original authoring dossier · Presentation PDF · Technical appendix PDF · Original independent_authoring.xml (57,068 bytes). The Reader index names its separate common source corpus; the small Gate 5 entry document is only the starting pointer. The Football focus XML introduces what the resulting program does, and this page explains its construction history.

This is the same kernel, extended by another author

Creating the implementation and accepting it are separate, just as Wedding's checked result and save are separate. The authored program then uses External DRS without taking ownership of another Root's decisions. A new body of code joins supported relationships; it does not gain general authority from the fact that it works.

Growth distinguishes this creation path from obtaining an already existing pure capability. Workshop explains the practical human, architect, executor, reviewer and owner roles; the public work handoff provides deeper source-bound context. Continue to one kernel and Geometry for the common relationships. Reading these materials does not start a build or execute the core.

Sources and scope

This page is a website explanation, updated 2026-10-02. Its explanation does not create a new execution result or change the status of its sources.