Follow a claim into its sources.

Architecture

Seven views of the same activity — Geometry Atlas connects each GEO-1…GEO-7 source to concrete episodes and the one-kernel map. It is distinct from the Incident Atlas.

The accepted architecture lock defines the common relationships and their authority. Publication chapters and diagrams help explain them; they do not create new runtime interfaces. A conceptual arrow is not evidence that every depicted relationship ran together in one experiment.

The Geometry Atlas overview holds one activity in view while separating its concerns:

Mathematics

Each numerical profile answers a particular question. The experience chapter defines integer advisory values, exact rounding, eligibility and finite budget allocation. Its fixed-point scale Q = 10^9 is dimensionless; a Work dispatch unit is neither a token nor money, time or a probability of safety. Hard eligibility precedes ranking.

The Wedding mathematics section defines guest/table indicators, mandatory-condition penalty H, and the objective Q(x) = P × H(x) + selected_soft_cost(x). Here Q(x) is an optimisation objective, distinct from the fixed-point scale. A solver or QPU supplies candidates; original-condition checks determine whether they satisfy this finite problem. The recorded hardware result establishes no quantum advantage.

Calibration is another profile: an exact rational correction in abstract input units. Qualified reuse reduces specific observed Work calls while retaining a current read and Root decisions; it does not by itself measure a wall-clock, energy or monetary saving. These explicit meanings make the formulas useful without turning them into universal performance claims.

Evidence and verification

Keep five evidence classes separate: accepted contract, recorded execution, independent checking within a declared trust scope, architectural inference, and proposed future work. A signature binds bytes within its trust scheme; it does not establish every external fact or grant present permission.

Start with a concrete claim in Workspace, then follow its record ID and field into the selected source. The Paper Source Key resolves editorial labels, primary paths, historical refs and selectors. Five Implementation Routes connects Root/effect, child topology, memory/time, external evidence and authorship to contracts, code and recorded checks.

Reading sources, checking saved artifacts, starting a new deterministic run, making a new live provider call, and independently admitting code are different activities. The verification chapter describes those routes. This website performs no runtime campaign, and opening it authorises none.

Incident-to-Proof Atlas

A refusal is only half the test. Read the full map of twenty cards and five experience consumers, with selected lawful continuations and offline Readers.

The Incident-to-Proof Atlas asks where finite risk/control cases reach a boundary and how lawful work can continue. It is a different publication from the seven-view Geometry Atlas.

Its twenty cards cover five actual consumers and seven effective history samples. Direct causal evidence, a shared domain proof and a control performed after consumption have different scope. The catalog combines reported incidents, research and requirements; it does not claim twenty independent learners or reproduced external attacks. Exact case routes preserve those distinctions. Supported saved replay does not recreate live authority.

Reference today

Version and role travel with the material. These are related sources, not one byte-identical snapshot:

Historical PENDING, NOT_RUN and unpublished-proposal captions remain in frozen documents. Read them with their date and later Release/handoff record; neither erase them nor use them as today's publication status. The four approved technical deferments remain separate from achieved reference scope. Downloads identifies the actual published files.

Limits and trust

The finite reference assumes its declared reviewed source, runner, observer and independently retained inputs. It does not establish general isolation for arbitrary hostile code, production security, universal planning, global federation or a real banking/booking deployment. Those operational responsibilities belong to the next integration's specific design.

A limitation does not excuse a demonstrated violation of an in-scope contract. Equally, an uninspected source is not evidence that a mechanism is absent. The historical scope document preserves its preparation-era status, while the method chapter explains how to assess a bounded claim and its lawful neighbour.

Continue with one demonstration, use an evaluation prompt, or follow the Build route for a concrete domain task.

Trace the result, not just the answer

Exact bytes and hashes identify a material; a seal binds the relationships declared by its scheme. A signature is checked against its identity, key, source and scope. These checks do not establish every real-world assertion or grant a new action. The external E → B → C story follows producer, published body and current local consumer; Wedding follows a measured candidate through validation and consumption.

Stored verification checks recorded relationships. A native rerun executes its declared program again; a mock action, fresh live model call and hardware task are different operations. Opening a Reader performs none of them.

A better score cannot buy a forbidden action

Experience follows finite G4 allocation: cap 9, spent 1, reserve 1, distributable 7; actual Work shifts from 4/3 to 3/4. AVF is advisory within eligible choices. Hard conditions and current owner permission remain outside that trade. N1-05 records a score of 1.0 without effect authority; full AVF/StrongGT are separate deferred mathematical work, not a reason to erase the finite working path.

From governance commitments to executable controls

Engineering interpretation · 3 October 2026

Safety commitments need controls that people can inspect: what may run, who may authorise it, and what evidence shows the outcome. A policy states the obligation; an execution system must make the relevant checks concrete.

Radiolaria provides a technical basis for that work: local decision authority, scoped action permissions, cryptographic evidence and verification of recorded execution. The Incident-to-Proof Atlas connects defined failure cases to refusal boundaries and useful continuations. Source routes let reviewers inspect the mechanisms and their stated trust assumptions.

The 29 September 2026 White House Accord on Super Intelligence calls for internal controls, internal review, independent external assessment and board oversight. It is a voluntary commitment, not a certification standard. Radiolaria offers mechanisms and recorded evidence relevant to selected execution controls; it does not replace independent auditors, corporate oversight or a deployment-specific compliance assessment.

Sources and scope

This page is a website explanation, updated 2026-10-03. Its explanation does not create a new execution result or change the status of its sources.