Verify Rotation Citation

RETIRED (batch-32 meta-phase, 2026-06-05). Pre-redirect orchestrator-era skill: it verifies a rotation_claim JSON for the old Critic role — a JSON-claim shape that no longer exists (the redirect replaced rotation_claims with CYCLE.md proposed-changes blocks). Read by NO live .claude/agents/* definition. The live citation-verification replacements are verify-citation-range + the tools/citecheck/ mechanical checker (wired into the live producer/critic/lowering-verifier specs). Kept under _retired/ as historical record (recoverable from git).

verify-rotation-citation

The Critic's per-claim verification workflow. Adversarial — finds errors, not agreement.

When to invoke

• Critic, once per rotation_claim produced by the Synthesizer in a cycle.

Inputs (prefetched by orchestrator)

• The rotation_claim JSON.
• The cited source ranges (text fetched ahead of time; do not re-fetch).
• Any tests cited or named in scaffolding/test-linkages/ for the source range.
• The prior-layer claim(s) (Li-1 forms the from_form is rotating from).
• The current concepts/ library state for any primitive referenced.

You do NOT see the Synthesizer's chain-of-thought. You do NOT see the in-flight cycle's scaffolding. Only the claims, the cited source, and the persistent project state.

Procedure

For each rotation_claim, evaluate sequentially:

1. Citation supports `from_form`? Read the cited source range; does it contain the structure from_form asserts? If not → citation_does_not_support.

2. Reduction mechanical? The justification's argument — does it skip steps or hand-wave? An L1→L2 algebraic argument that elides which kernel-fusion is being unfolded is suspect → rotation_chain_breaks.

3. Mutation pattern accurate? For L0→L1: does the cited source actually exhibit the mutation pattern claimed (in_place_overwrite vs accumulator vs alias_with_input vs scratch_buffer vs pure)? Look for the destination buffer's use: is it read first then written (accumulator) or just written (overwrite)? Is it the same buffer as one of the inputs (alias)? → mutation_pattern_mismatch.

4. Ownership accurate (L4 only)? For L3→L4: does the L4 form correctly distinguish sim state (evolves) / operator internal params (closure-held, constant during solve) / ephemeral intermediates (per-step, doesn't survive)? Misclassification of a closure-held matrix as sim state, or a per-step temp as state, → ownership_misclassified.

5. Missing cases? Does the source have error paths, special branches, edge conditions that the rotation doesn't address? An L1 claim that omits a branch is missing the branch's semantics → missing_case.

6. Load-bearing vs transparent classification? A numerical trick (non-associative reduction ordering, mixed precision, fast-math, deterministic-vs-atomic) marked transparent when it changes results under naïve reformulation → load_bearing_trick_classified_as_transparent. When in doubt, mark unclear and let the human triage.

7. TEST CONSISTENCY. Where tests exist (cited by Synthesizer, listed in scaffolding/test-linkages/, or discoverable via the find-tests-for-region skill), do the test inputs and assertions support the L1 mutation pattern and the L2/L3 algebraic claims? A test assertion contradicting a claim → citation_does_not_support (tests are L0-equivalent evidence).

8. Labored rotation? Is the rotation technically correct but obviously forced — special cases, exception branches, forced-fit transformations? → labored_rotation_push_back_candidate. Include a push_back_suggestion: which lower-layer change would eliminate the friction?

Verdict assembly

• If issues found are all unclear or labored_rotation_push_back_candidate → verdict = revise (claim is on the right track but needs refinement).
• If issues include citation_does_not_support, rotation_chain_breaks, mutation_pattern_mismatch, ownership_misclassified, missing_case, or load_bearing_trick_classified_as_transparent → verdict = revise or reject depending on severity.
• No issues → verdict = pass.

Cross-cycle pattern → lesson

If you notice a pattern across the cycle's claims — the Synthesizer consistently misclassifying mutation, two slices' L3 forms hinting at a missing shared primitive — write a one-sentence lesson for lessons.md. Lessons are cheap and disproportionately effective.

Output

A single JSON object validating against schemas/critic_verdict.json. Nothing outside the JSON. (The orchestrator handles lessons.md append and episodic.jsonl write.)

Friction → problems/

If the verdict schema's kind enum doesn't have a category for what you're seeing, or the Synthesizer's diff is structurally hard to evaluate (mixed concerns in one diff, claims that span multiple edges without separation), file as a problems/ entry.