Menu
Classify each prose claim before writing to paper/paper.md. Enforce the controlled vocabulary and reject illegal proof language.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | evidence-contract |
| description | Classify each prose claim before writing to paper/paper.md. Enforce the controlled vocabulary and reject illegal proof language. |
The Writer agent and the Supervisor both invoke this skill before committing
any update to paper/paper.md. The contract prevents the manuscript from
outrunning the evidence.
Scan the draft section you are about to write. For every sentence that asserts a mathematical result, highlight it and tag it as a candidate claim. Examples of sentences to tag:
Use this table to assign a claim type to each tagged sentence:
| Claim type | Prose you are allowed to write | Evidence required |
|---|---|---|
proved | "We prove / establish / show that …" | A file at proofs/<claim-id>.md exists and project.yaml has claims[<claim-id>].status: proved |
partial-proof | "We give a partial proof of …" | proofs/<claim-id>.md exists with Status: partial-proof |
proof-sketch | "We sketch a proof that …" | proofs/<claim-id>.md exists with Status: proof-sketch |
simulation-supported | "Empirically, we observe …" | experiments/<exp-id>/summary.md exists and is linked |
hypothesis | "We conjecture …" | Entry in conjectures/conjectures.md |
writing-only-interpretation | "Intuitively …" / "We interpret this as …" | None, but the sentence must carry the inline tag <!-- interpretation --> |
For every proved-type claim:
project.yaml. Find claims[<claim-id>].status.proved → reject. The sentence must be rewritten to
match the actual status (e.g. downgrade to "we sketch a proof").proofs/<claim-id>.md. Confirm it exists and has a non-empty
## Argument section with zero open **G1.** gaps.partial-proof until the gaps are closed.For every simulation-supported claim:
experiments/<exp-id>/summary.md exists.figures/.hypothesis.Every proved or simulation-supported claim in paper.md must carry an
inline anchor immediately after the sentence:
… achieves a (1−1/e) approximation. <!-- claim:C1 proved -->
… we observe a threshold near 0.5. <!-- exp:exp-003 simulation-supported -->
Before finalizing the draft update, output a short pre-commit manifest for the Supervisor:
Evidence pre-commit check
─────────────────────────
proved C1 ✓ C3 ✓
proof-sketch C2 ✓
simulation exp-003 ✓
hypothesis C4 ✓
interpretations: 2 (tagged)
Rejected claims: none
If any claim was rejected, list it and explain why. The Supervisor must see
this manifest before the draft is appended to paper/paper.md.
status: proved in project.yaml and a gap-free proof
file exists.proved.writing-only-interpretation sentences must be tagged inline; do not
present them as results.Academic research assistant for literature reviews, paper analysis, and scholarly writing. Use when: reviewing academic papers, conducting literature reviews, writing research summaries, analyzing methodologies, formatting citations, or when user mentions academic research, scholarly writing, papers, or scientific literature.
Adversarial review of mathematical proofs, theorems, and formal arguments. Use when asked to review a proof, check a paper's correctness, verify a theorem, find bugs in mathematical reasoning, audit a derivation, or stress-test a formal argument. Implements iterative self-correction from Woodruff et al. (2026) and the Generator-Verifier-Reviser pattern from Aletheia (Feng et al., 2026).
Read, update, and promote claim statuses through the controlled vocabulary. Includes the project.yaml patch protocol and the conjectures.md update procedure.
Systematically test mathematical conjectures by searching for counterexamples, verifying small cases, and constructing proofs or refutations. Use when Raman proposes a conjecture, when a paper claims an unproven bound, when testing whether a theorem generalizes, or when asked to "find a counterexample," "test this conjecture," "verify this bound," or "check if this holds." Implements the simulation and counterexample search techniques from Woodruff et al. (2026) and the neuro-symbolic verification loop from Aletheia (Feng et al., 2026).
Generate research ideas by finding cross-disciplinary connections, retrieving obscure theorems from distant fields, and identifying analogies between different mathematical domains. Use when Raman asks for new research directions, when ingesting a paper that might connect to other fields, when looking for "bridges" between topics, or when asked to "find connections," "what other fields use this technique," or "cross-pollinate." Implements the cross-pollination techniques from Woodruff et al. (2026, Section 2.2, 4.x) and AI Behavioral Science's framework for studying AI-human strategic interactions (Jackson et al., 2025).
Extract, classify, and integrate PDF papers from the project inbox/ folder into the literature pipeline. Use this skill whenever PDFs are present in inbox/, when a user says "I dropped some papers in the inbox", or before any literature or ideation pass that should incorporate locally stored papers.