v3 Slice 2: Tracer bullet - one causal item end-to-end (generate, run, score, leaderboard)

[mark-allwyn]

Parent

#13

What to build

The thinnest complete path through the whole pipeline - one causal item flowing generate -> run -> score -> leaderboard - proving every layer integrates before any volume work.

• A seeded, deterministic generator for one causal bundle (B01) producing conjunctive items: each item carries several machine-checkable parts (e.g. minimal sufficient adjustment set as a node set, adjusted effect estimate as a number with tolerance, identifiability as a verdict). Same seed -> identical items.
• The dual-verification admission gate: the generator's graphical-solver answer is cross-checked against a seeded linear-Gaussian simulation + partial correlation; any disagreement rejects the item before it ships.
• Refusal-neutral cover stories only.
• The Anthropic provider (ported from legacy) returning a structured result: content, stop_reason, stop_details, input/output/thinking token counts, native config used. Claude 4.6+/Fable run with adaptive thinking, no temperature, max_tokens >= 16000.
• A minimal runner: execute pending (model x item) jobs, save raw responses atomically (write tmp, validate, replace), preserving multi-run history per item.
• An offline scorer: extract each tagged part, normalize per type, score each part, set item_correct = all parts correct (conjunctive); classify response status. Never re-calls the API.
• A console leaderboard: accuracy = correct / attempted, with refusal/invalid shown as separate columns, computed via the shared accuracy helper.

Acceptance criteria

• Mocked-provider test: a canned API response flows item -> raw -> scored -> leaderboard with correct conjunctive accuracy and status
• Same seed produces byte-identical B01 items; the dual-verification gate rejects an injected miskeyed item
• Live smoke: one real claude-fable-5 (or claude-haiku) call records stop_reason and thinking_tokens
• A refusal stop_reason classifies as refusal, not invalid; an unparseable part scores that part wrong without voiding the item
• Re-running score never re-calls the API; reruns append to per-item history
• pytest green

Blocked by

• v3 Slice 1: Scaffold + spine - package, registries, tagged-text extractor, status taxonomy #14