The defining property of a first-principles reviewer is fresh context — and the useful insight is
that this is a control, not a personality. It is not "a second model with different opinions"; it
is a reviewer deliberately deprived of the narrative that would let them restate rather than re-test.
research-method names both halves:
§8: a reviewer given prior findings will restate rather than re-test them unless instructed to
refute. Review instructions MUST direct the reviewer to attack.
§9.4: point a refuter at ONE claim — "review this" and "refute this" are different instruments.
§7.2: Linear Sequential Unmasking — order of exposure as a control.
Proposed: a refutation mode that hands out one claim and its DIRECT evidence, with the surrounding
vault withheld. Something like peira refute <id>, emitting a packet-shaped brief containing the
claim, its warrant, its immediate supporting nodes and their grades — and nothing else. No sibling
claims, no narrative, no conclusions reached elsewhere.
Two properties worth building in:
- It must be scoped to one claim. A refuter handed a document reviews the document; a refuter
handed a claim attacks the claim.
- The instruction must say refute, not review. That belongs in the emitted brief itself, not in
whatever a caller happens to type.
The guard against a permanent contrarian is measurable, not temperamental
A refuter that refutes everything carries no information. That is not a matter of tone — it is
§4.4 of the grading spec:
A detector firing on ~100% or ~0% of inputs is describing itself. Investigate the instrument
before reporting the result.
So a refuter's rate is a diagnostic of the refuter. Worth reporting alongside its findings, the same
way a detector's base rate is reported alongside its precision.
Related: #8 (instrument validity is the same idea applied to tools rather than reviewers).
The defining property of a first-principles reviewer is fresh context — and the useful insight is
that this is a control, not a personality. It is not "a second model with different opinions"; it
is a reviewer deliberately deprived of the narrative that would let them restate rather than re-test.
research-methodnames both halves:Proposed: a refutation mode that hands out one claim and its DIRECT evidence, with the surrounding
vault withheld. Something like
peira refute <id>, emitting a packet-shaped brief containing theclaim, its warrant, its immediate supporting nodes and their grades — and nothing else. No sibling
claims, no narrative, no conclusions reached elsewhere.
Two properties worth building in:
handed a claim attacks the claim.
whatever a caller happens to type.
The guard against a permanent contrarian is measurable, not temperamental
A refuter that refutes everything carries no information. That is not a matter of tone — it is
§4.4 of the grading spec:
So a refuter's rate is a diagnostic of the refuter. Worth reporting alongside its findings, the same
way a detector's base rate is reported alongside its precision.
Related: #8 (instrument validity is the same idea applied to tools rather than reviewers).