evidence-contract

name	evidence-contract
description	Classify each prose claim before writing to paper/paper.md. Enforce the controlled vocabulary and reject illegal proof language.

Skill — Evidence Contract

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.

Step 1 — Extract candidate claims

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:

• "We prove that the greedy algorithm achieves a (1−1/e) approximation."
• "Empirically, we observe that the threshold is near 0.5."
• "We conjecture that no polynomial-time algorithm beats ratio r."
• "Intuitively, this arises because spreading saturates early."

Step 2 — Classify each candidate claim

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 -->

Step 3 — Resolve each claim against the artifacts

For every proved-type claim:

1. Open project.yaml. Find claims[<claim-id>].status.
2. If status ≠ proved → reject. The sentence must be rewritten to match the actual status (e.g. downgrade to "we sketch a proof").
3. Open proofs/<claim-id>.md. Confirm it exists and has a non-empty ## Argument section with zero open **G1.** gaps.
4. If the proof file has open gaps → reject. Status cap is partial-proof until the gaps are closed.

For every simulation-supported claim:

1. Confirm experiments/<exp-id>/summary.md exists.
2. Confirm the linked figure file exists under figures/.
3. If either is missing → rewrite as hypothesis.

Step 4 — Inject evidence anchors

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 -->

Step 5 — Supervisor pre-commit check

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.

Hard rules (cannot be overridden by any playbook)

• "We prove / establish / show that" is illegal unless the matching claim ID has status: proved in project.yaml and a gap-free proof file exists.
• A simulation result can never promote a claim to proved.
• writing-only-interpretation sentences must be tagged inline; do not present them as results.
• If you cannot find the claim ID for a sentence, surface it to the Supervisor rather than guessing.