Scholar Contest

this skill

Stage 4. Attack the ontology. Four passes, serial by default: one judge, one candidate at a time, every verdict traceable to the evidence that settled it.

Read references/evidence-tiers.md before recording any number. The analyzer ships beside it at references/term-metrics.py; <analyzer> below is that path.

User arguments: $ARGUMENTS

Pass 1 — mechanical

Run the analyzer:

$ <analyzer> notes/ontology/<subject>/

It reports synonymy candidates, polysemy candidates, orphans, circular evidence, and fan-out per type.

THE ANALYZER PROPOSES. IT NEVER DECIDES.

A high similarity score is a question — are these one concept? — not an answer. Record the number in evidence/metrics.md with the command that produced it, per the evidence tiers.

Pass 2 — serial judgment

One candidate at a time. For each, write to evidence/contested.md:

• Claim — what the ontology currently asserts.
• Objection — the strongest case against it, argued properly rather than noted.
• Evidence — what settles it, cited or measured.
• Verdict — upheld, revised, or killed.

KILLS ARE RECORDED, NOT DISCARDED

A kill often carries more than a survivor, because it names a boundary the ontology drew in the wrong place. A candidate rejected for the wrong reason tells you the type above it is mis-specified.

Judge one candidate fully before starting the next. Batching them produces verdicts that argue with each other.

Pass 3 — completeness

Read sources.jsonl and ask what the study does not know:

• Which sources were never opened, and whether the vocabulary plausibly lives in one of them.
• Which terms rest on a single citation.
• Which types have exactly one member. A one-member type is usually a description wearing a type's clothes.
• Which terms have no citation that shows them in use, only in definition.

Pass 4 — stop condition

Stop when a round surfaces no new contested claim, or at --rounds (default 3). The convergence rule is the spike-ratchet skill's: stop when the design stops fighting back, not when a counter runs out.

The circularity check

Runs in every pass. A citation whose quoted text contains the term it supports proves the corpus uses the word, which was never in question, and hides whether the corpus uses it the way the definition claims.

This is not theoretical. Routing eval prompts in this repository contained their own skill's name, which made every rename candidate look expensive until the self-reference was noticed.

Panels

--panel hands one contested claim to independent adversarial participants via the weave-ask skill, recording their verdicts in evidence/contested.md beside the serial one. Where weave is not installed, say so and continue serially — the panel is an amplifier, not a dependency.

Rules

• Every analyzer finding has a recorded verdict before this stage hands off.
• A verdict cites or measures. "On reflection this seems fine" is not a verdict.
• Do not edit terms.jsonl to make a finding go away. Revise the claim and record why.
• A claim raised and not settled goes in evidence/contested.md as open, not in the brief.

Output

Open with a one-line hero (✓ <n> candidates judged, <n> killed, <n> open or ⚠ Blocked: <reason>), then exactly these sections:

1. ## Findings — what the analyzer proposed, by kind.
2. ## Verdicts — each candidate, its objection, and how it resolved.
3. ## Kills — what was killed and what the kill revealed.
4. ## Coverage — what the study does not know, from pass 3.

End with an ask-user-choice panel offering next steps (for example: another round, run render, widen the corpus, stop here) — skip the panel only in plan mode.

Portability notes

• ask-user-choice — present the listed options and wait for the user to pick one. Hosts with a structured multiple-choice tool (Claude Code's AskUserQuestion) should use it; otherwise print a numbered list and wait for a numbered reply. Never proceed on an assumed answer.
• $ARGUMENTS — the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.
• Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.