Method

Automate the inspection. Bind the criticism to evidence.

Not every audit requires an executable verifier. When a claim can be independently reconstructed from public artifacts, the verifier turns the evidence chain into an executable check rather than an automated truth machine. The same bounded logic also applies when the decisive work is source tracing, construct analysis, metric decomposition, or claim lineage.

Core evidence chain

One case, one bounded dependency path.

Source artifact
→
Candidate defect
→
Independent verification
→
Minimal repair
→
Result change
→
Claim consequence
Rule: discovering a defect is not the same as establishing its consequence. The dependency from source to published claim must be traced rather than assumed.
Evidence depth

M0 → V4

M0 machine candidate · M1 source located · V2 independently verified · V3 reproduced · V4 consequence verified.

Impact depth

A0 → A5

Cosmetic only → technical defect → reproducibility impaired → specific result affected → explicit published claim affected → core substantive conclusion affected.

These axes stay separate. A highly verified defect can still have low substantive impact. A dramatic narrative is not a substitute for demonstrated claim dependency.
Progressive verification

Spend depth in proportion to the load carried by the claim.

Layer 1

Quick Verify

Fingerprint the input artifact and independently recompute the focal consequence from row-level or draw-level data.

Layer 2

Deep Verify

Move farther upstream: reconstruct matrices, transformations, solves, or other model-specific intermediate objects and compare downstream output.

Layer 3+

Source & Provenance

Pin the exact archived source with hashes and anchors, then disclose the provenance path required to regenerate compact audit artifacts.

AI-assisted, evidence-verified

AI can find candidates. It cannot certify its own accusation.

AI may assist retrieval, document comparison, code-smell detection, dependency tracing, and procedure drafting. A public finding requires source-level or executable verification.

Never automate the accusation. Automate the inspection. Never trust the model output. Verify the evidence. Never infer maximal damage. Measure the consequence. Never hide uncertainty. Bound the verdict.
Human adjudication

The interpretive layer remains explicit.

Construct validity, mechanism, theory–evidence fit, alternative explanations, substantive interpretation, fairness, and the boundary of what follows are not collapsed into a single automated score.

Quantitative boundary: calculations can be reconstructed when public artifacts permit it. Questions that require specialist statistical judgment are identified as requiring that expertise rather than silently treated as solved.