Minerval

← evals

EvalIs a graph well built?

Structural scorecard

Counts taken from a run's graph, no model involved. Each one serves a section of the rubric.

on every scored run (1)free

What it checks

Counts a program takes from a run’s graph, with no model involved: how many claims came out of how many words, how long their wording is, how often the Matcher merged, how deep the decomposition goes, how many verdicts have reasoning behind them. Each count serves one section of the rubric and is a symptom of a failure that section names.

The numbers only mean something against earlier runs: a count that moves beyond the How much a number varies between runs when nothing changed. A difference smaller than that is not a result.evals guide after a change is a lead, and its direction says which agent to look at.

Detail

The metricshover a name for its standard

Letters are rubric sections. Hover a metric for the standard it serves; click for the section. Judge rows come from the judged scorecard.

metriclatest (2026-08-09)across runs
A · Extract all substantive claims, faithfully and charitably, across every claim type. Questions, commands, meta-text, stipulative definitions, and hedged non-assertions are not claims.rubric33.16single sample
B · The canonical form is the shortest neutral statement of the proposition as it is actually debated: about fifteen words, rarely more than twenty-five, terse and frame-independent, self-contained, and meaning-preserving. State the proposition at the precision the discourse debates it; do not sharpen it with parameters the author never committed to, and do not mint placeholders for missing ones — the vague proposition is the claim.rubric25single sample
B · The canonical form is the shortest neutral statement of the proposition as it is actually debated: about fifteen words, rarely more than twenty-five, terse and frame-independent, self-contained, and meaning-preserving. State the proposition at the precision the discourse debates it; do not sharpen it with parameters the author never committed to, and do not mint placeholders for missing ones — the vague proposition is the claim.rubric8%single sample
C · Two formulations are the same claim when they turn on the same considerations: nothing could count as evidence or argument bearing on one without bearing equally on the other. Identical decomposition is a useful diagnostic, not the definition.rubric2.29single sample
D · Neutral decomposition: identify what a claim depends on, don't evaluate it. Decompose until reaching genuine bedrock (bedrock fact / contested empirical / value premise).rubric6single sample
D · Neutral decomposition: identify what a claim depends on, don't evaluate it. Decompose until reaching genuine bedrock (bedrock fact / contested empirical / value premise).rubric58%single sample
E · Redundancy should collapse and structure should be shared: claims that recur across posts become one node, common dependencies become shared subclaims, and related/opposing positions are connected, not siloed. The admin thinks at graph level — upstream, downstream, lateral.rubric6single sample
F · Use the six statuses honestly (Verified, Supported, Contested, Unsupported, Contradicted, Unknown); never round a genuinely contested claim up to verified or down to contradicted; every assessment carries a substantive reasoning trace; assessment is holistic judgment, not mechanical aggregation.rubric100%single sample
§21 · coherence violations0single sample
B · The canonical form is the shortest neutral statement of the proposition as it is actually debated: about fifteen words, rarely more than twenty-five, terse and frame-independent, self-contained, and meaning-preserving. State the proposition at the precision the discourse debates it; do not sharpen it with parameters the author never committed to, and do not mint placeholders for missing ones — the vague proposition is the claim.rubricn/anot on this run
B · The canonical form is the shortest neutral statement of the proposition as it is actually debated: about fifteen words, rarely more than twenty-five, terse and frame-independent, self-contained, and meaning-preserving. State the proposition at the precision the discourse debates it; do not sharpen it with parameters the author never committed to, and do not mint placeholders for missing ones — the vague proposition is the claim.rubricn/anot on this run
C · Two formulations are the same claim when they turn on the same considerations: nothing could count as evidence or argument bearing on one without bearing equally on the other. Identical decomposition is a useful diagnostic, not the definition.rubricn/anot on this run
importance · Use the six statuses honestly (Verified, Supported, Contested, Unsupported, Contradicted, Unknown); never round a genuinely contested claim up to verified or down to contradicted; every assessment carries a substantive reasoning trace; assessment is holistic judgment, not mechanical aggregation.rubric0.3single sample
judge · claim-bar pass rate69%single sample
judge · importance overrated share0%single sample
judge · readability (1–5)4.08single sample
judge · reasoning fit (1–5)3.77single sample
judge · impartiality (1–5)3.77single sample
judge · sycophancy sharen/anot on this run
judge · overhedged sharen/anot on this run
judge · overconfident sharen/anot on this run
judge · canonical-form miss sharen/anot on this run
judge · political bias sharen/anot on this run
How the numbers are computed

A pure function over the graph, unit tested: metrics.ts. Shared subclaims are counted once. The two coherence rules are checked mechanically: a claim cannot be verified while a claim it requires is contradicted, and two claims joined by a contradiction edge cannot both be verified.

Authorship metrics (how often the Matcher reworded the Extractor’s proposed The one neutral wording of a claim that every source's version maps to. About fifteen words.constitution) need the proposal stored on the instance, which graphs built before September 2026 lack: those rows read n/a.

What it cannot show

Whether any claim is well stated, well decomposed, or fairly judged. The counts say where to look; a wording of fifteen words can still be wrong in every way §3 cares about.

Run it
npm run corpus:score -- blackholes --no-judge     # after a run; writes scorecard.json