Fable review — independent cross-model challenge of a plan/diagnosis

A common failure mode in chemistry-tool development is a confident-but-wrong premise — a diagnosis that reads as solid, is repeated across sessions, and turns out wrong on re-derivation. Automated correctness gates — a round-trip/canonicalizer check, a regression suite, an atom/mass-balance certificate — all verify code, not reasoning. This skill gets an independent review from a different model family before shipping, aimed at the diagnosis. The reviewer reads and reports; it does not implement. A different-family model reviewing a diagnosis is fine; one working as a parallel implementer on the same files is not (see your project's own rules on parallel work), because a fresh model has not read your project-specific invariants.

How to run

Spawn one reviewer via the Agent tool on a model family other than the session's: fable by default, opus when the session itself runs on Fable. A same-family reviewer shares the author's blind spots, and an alias for the session's own family resolves to the session's exact model, so it would not be a second opinion.

Agent(
  subagent_type: "claude",   # catch-all agent type
  model: "fable",            # alias, follows the newest Fable; "opus" if the session is on Fable
  description: "Fable review of <plan>",
  prompt: <the review brief below, with the doc path(s) filled in>,
)

Run it in the background (async) when the user wants to continue other work meanwhile; wait for it when the plan is blocking a commit. One reviewer, not a fan-out — an oversized fan-out of reviewers adds cost, not independence.

If your organization's model allowlist permits no version of the requested family, Claude Code runs the reviewer on the session's own model and shows a warning; check the model named on the reviewer's row in /tasks before counting the review as cross-model.

The review brief (fill in the paths)

You are an adversarial reviewer on a different model family from the author. Your job is to refute, not to agree. Read <DOC PATH> (and any files/commits it cites). This is <describe your chemistry tool, e.g. a deterministic structure→name namer, a property-prediction pipeline, a reaction-outcome predictor>; the dominant failure mode here is a confident-but-wrong premise that passes every gate because gates verify code, not reasoning.

Read <METHODOLOGY DOC PATH> for your project's standing invariants — at minimum: spy-before-you-code, verify-your-harness-actually-ran, VERIFIED-vs-ASSUMED tagging on load-bearing claims, rank-by-sole-blocker (not touched-count), and bounds that can all hold at once (bounded-goals). VERIFIED = cites the command and output, run id or quoted source that shows it; ASSUMED = does not. For every load-bearing claim, not only the first few:

1. Is it VERIFIED or ASSUMED? An ASSUMED claim may not justify deleting a guard or sizing a fix.
2. Could the measurement be right but the diagnosis wrong? Name a concrete alternative cause.
3. Is any reachability/yield/size number derived from static inspection or a warm-cache/batch run rather than a fresh-process measurement over the target rows?
4. Does any proposed fix touch a correctness-sensitive path without a full-gate/regression plan, or pair a coverage floor with a byte-identity/total-proof requirement (a classic unsatisfiable- goal trap — see the bounded-goals skill)?
5. Is a "central decision point" named as the fix site without a spy proving it is on the path?

Output: a numbered list of every concrete refutation or gap you find, each tagged [BLOCKER] / [RISK] / [NIT], each with the exact check that would settle it; the tags carry severity, so report low-confidence gaps too. If you find nothing, say so and name the single claim you are least able to verify. Return only this list: leave the plan unrestated and the code unchanged.

After the review

The reviewer reports everything it finds; the filtering is yours. Check each finding against the evidence before acting on it: accept it or refute it with a result, rather than agreeing by default or implementing it unexamined. A [BLOCKER] must be resolved (or explicitly refuted with evidence) before the plan ships. Record durable outcomes in the finding doc.