| name | research-proof-auditing |
| description | Use when auditing computational or formal proofs, writing or reviewing mathematical argument notes, checking evidence sufficiency, reviewing Sage/GAP/Lean/Aristotle verification, detecting proof fraud, or deciding whether a result is accepted. |
Research Proof Auditing
This skill is the canonical proof, evidence, fraud-detection, and audit-sufficiency authority for the research repo.
Canonical source
The source of truth is this skill plus references/proof-auditing.md.
Read references/proof-auditing.md before auditing computation scripts, formal proofs, proof claims, mathematical argument notes, mathematical verification output, adversarial audit evidence, or acceptance claims.
Core principle
Assertions with external sources are proof. Print statements are theater.
A computation that prints success proves nothing. A computation that asserts a sourced expected value proves the relevant claim when the asserted predicate is mathematically adequate for the task.
Argument-shape gate
Mathematical prose must expose the dependency chain. Reject notes that replace
construction with naming, cite authority instead of stating theorem hypotheses, inflate
immediate consequences into literature claims, or use vague relational language where a
map, embedding, quotient, pullback, or birational comparison is required.
Immediate facts should be called immediate. The audit burden belongs on the construction
of the objects and maps that make those facts relevant.
Standardness calibration
Before accepting a "standard" claim, classify it:
- immediate consequence of a presentation;
- trivial first-principles derivation from definitions;
- standard foundational theorem with explicit hypotheses;
- niche research theorem requiring a precise citation and hypotheses;
- project-specific claim requiring construction or computation.
Do not let difficulty drift upward. If a fact is immediate from rank, signature, a Gram
matrix, a standard decomposition, or a textbook construction, state it directly and move
the proof burden to the upstream construction that supplies the input.
Do not let progress accounting drift either. A short derivation from definitions is not
a milestone comparable to a niche theorem that identifies the correct object, map, or
moduli comparison. Do the short derivation in place; reserve audit weight for the claims
that smuggle in assumptions.
Required audit stance
- Check the exact
GOAL.md obligation before judging evidence.
- Check the claim's visible status in the card, workstream report, and living paper.
Reject polished prose whose margin notes or report links do not expose provenance,
contentiousness, computation dependence, or human-review needs.
- Reject substitute computations, sampled evidence, bounded search without a proof of exhaustiveness, and prose claims presented as certificates.
- Reject authority chains, conclusion-smuggling names, and "standard" claims that do not
state the standard theorem, hypotheses, and exact role in the argument.
- Reject public mathematical surfaces that return raw nonmathematical Sage base types
such as
Parent or Element unless the surface is a true deep base-category bridge.
Most audited code should know the mathematical category of the returned object.
- When a broad Sage-facing type is genuinely required, require an explicit mathematical
or interop alias such as
SageCategoryObject or SageElement instead of leaking
naked Sage infrastructure names into public proof, spec, or API surfaces.
- Reject mathematical research findings that depend on external or web sources but were
left only in chat. Require a durable mathematical report memory with source links
before treating the research as reusable audit input.
- Treat prior session claims, markdown summaries, and agent self-reports as unverified until backed by a passing script or formal proof.
- Require every assertion's expected value to trace to
GOAL.md, literature, independent computation, or a cited derivation.
- Reject proof computations with zero assertions, self-computed expected values, hardcoded boolean verification, print-statement theater, or exception-swallowing verification.
- Formal proofs must have precise theorem statements, no
sorry, and successful project build/check evidence.
- Treat reviewer consensus as evidence, not proof. If review appears to converge by
weakening, obscuring, or restyling the claim instead of fixing it, flag false
consensus. If review cycles do not reduce a precise disputed assertion, stop the loop
and require escalation.
Load with
- Load
research-orchestration when the audit is part of state-machine execution or acceptance.
- Load
research-code-style when audit findings concern code style, assertions, exceptions, constructors, or mathematical API design.
- Load
research-math-boundary when audit findings concern the trusted mathematical base, backend ownership, Sage/GAP/Julia routing, or exact computation semantics.
- Load
category-spec-audit when auditing category-spec plans, cards, smoke work, or implementations.
Stop conditions
Stop acceptance and route follow-up when the evidence proves a cheaper proxy instead of the requested claim, when exactness is replaced by numerics or sampling, when theorem hypotheses are unchecked, when witnesses are missing, or when the target theorem has drifted.