Evidence and claims

This page describes the classic campaign contract/claim model. Minimal studies apply the same separation of execution and scientific acceptance through experiment receipts, source-bound methods review and prospective repair commitments; see minimal research. A minimal study’s answer completion policy can finish on a reviewed root falsification; the repair policy requires an independently supported repair. Do not apply classic instrument-claim APIs or completion rules to every workflow.

Evidence is prospective

A decisive claim requires a contract registered before its evidence-generating action. The contract states what is observed, how outcomes map to dispositions, what uncertainty matters, what makes the observation inconclusive, and which exact program commands may generate evidence.

Agent-authored files and artifacts written while code is still being designed remain workbench records. They cannot be promoted retroactively. An execution-generated artifact is only eligible after the process succeeds and, for detached work, its authenticated job receipt is reconciled as succeeded. Failed, cancelled, or operationally unknown jobs remain non-evidence even when they left a plausible partial output.

Every contract also declares its purpose. A claim_decision contract tests the claim’s predicted outcome and may support or falsify it. A terminal_record contract records a bounded failure to decide—for example, an unavailable instrument, an unrealized physical antecedent, or uncertainty that cannot be reduced within the campaign. Completing that record can justify instrument_limited or unresolved; it cannot justify supported or falsified. A scientific claim can register a terminal record only after a tracked prospective attempt under a claim-decision contract, so an agent cannot skip the scientific test and commission only an excuse for stopping. If the scientific prediction changes, the agent must register a successor claim rather than rewrite the contract to make failure satisfy it.

Instruments are claims too

A simulator or analyzer must establish representation, physics controls, boundaries, diagnostics, and numerical-regime behavior under a commissioned contract. The supported instrument claim qualifies later bound executions; a successful process exit alone does not.

Claim closure

Supported and falsified dispositions require sufficient, validation-passing, provenance-tracked evidence from one selected claim_decision contract version. When a contract is amended, earlier observations remain visible but cannot decide the new version; fresh evidence must be generated after that version was registered. Unresolved and instrument_limited require a complete prospective terminal_record, so stopping honestly remains auditable without converting a blocked experiment into support.

The root hypothesis is immutable. A narrower operational child can close under its own evidence contract without silently changing or overclosing the root.

Keep the hypothesis tree scientific

A scientific child is a proposition that changes, narrows, repairs, or competes with a physical prediction and can become the next falsification target. An established formula, observable definition, numerical method, or implementation cross-check normally belongs in the parent’s evidence contract rather than in a new refines node. If an estimator or diagnostic genuinely needs an independent audited disposition, it is a diagnostic_of validation claim. This keeps the hypothesis tree from becoming a list of analysis steps.

Finite grid samples do not by themselves establish a universal statement over a continuous interval. Claims using terms such as “throughout,” “every,” or strict monotonicity require an analytic argument, a validated enclosure, or an explicit resolution-bounded scope.

Counterexample, repair, and adjudication

A qualified counterexample closes the exact scientific claim it violates. The Scientist role then registers a repairs child that cites the motivating counterexample, explains how the replacement accommodates it, states the minimal semantic change, and identifies a future falsification condition. The old counterexample is motivation, not evidence for the repair. The Falsifier role must register a fresh contract and attack the replacement again.

When a meaningful search finds no counterexample, the acting model cannot declare its own evidence sufficient. A separate Judge context reviews the prospective contract, provenance, validation, uncertainty, coverage, and bounded artifact excerpts. An insufficient judgment returns concrete gaps and the falsification loop continues while wall time remains. A sufficient judgment means that the record is complete enough for the Judge’s explicit scientific disposition; it does not mean “supported.” The deterministic gate then checks that the disposition matches the contract purpose. A complete claim_decision case may close as supported, while a complete blocker record closes as instrument_limited or unresolved.

A falsified scientific claim is not a finished frontier by itself. It requires a minimal repairs child that accommodates the counterexample and becomes the next falsification target. Wall-time exhaustion may bound an open investigation, but it never becomes support.

Auditability is not infallibility

The ledger preserves exactly what rule the agent chose and how it applied that rule. A scientist can still choose an underpowered uncertainty design or overstate the interpretation. The independently reviewable record makes that error visible and prevents it from being confused with missing provenance.

Historical campaign artifacts are immutable. Later software may continue to read legacy adjudications that lack an explicit scientific disposition, but it must label them as legacy records and must not infer support in the user interface. A correction is appended as a separate audit record with hashes of the original artifacts; the original ledger and report are never silently rewritten.

Use simjecture corrective-audit RUN --reviewer NAME --finding TEXT --corrected-interpretation TEXT --artifact mvp_report.json --artifact hypothesis_ledger.json to append that hash-chained record.