/bench-fixture-gen

bench-fixture-gen (benchmark-internal, NOT a user skill)

Generates the shared per-task fixture that every competing arm starts from: a realistic dataset plus an image asset base. Fairness rule: fixtures are generated ONCE per task by this preprocessor and sealed (sha256) before any arm runs; arms never generate their own.

Workflow

1. Author a fixture spec (`fixture-spec.json`):

• `task_id`
• `dataset`: { `global_key` (window global), `entities`: [{ name, count,

fields: [{name, kind, enum?, range?}], relations? }], `disclosure` }

• `images`: generate-task-assets item list ({file, prompt, style})

2. Dataset: `node scripts/generate-task-dataset.mjs --spec <spec> --out <dir>`

• grok-4.6 authors realistic records for the spec (headless, isolated);

the script VALIDATES structure (entity counts, field presence, enum membership, relation integrity), retries once on failure, then writes `data.json` + `data.js` (window global) + prints aggregate seeds. 3. Images: `node scripts/generate-task-assets.mjs --spec <images-spec> --out <assets-dir>`

• grok image_gen, paced, self-healing auth sync, per-file sha manifest.

4. Seal: record data.json sha + assets-manifest in the task fixture and in every prepared cell's `cell.json`.

Rules

• Deterministic where possible; all sample data labeled fictional in a

`disclosure` string carried inside the dataset itself.

• Never real brands/people; Korean-locale tasks use Korean-plausible

fictional names.

• Aggregate expectations for data-fidelity audits are DERIVED from the

sealed dataset at audit time, never hardcoded.

• KPI terms used in briefs must be DEFINED in the brief (lesson: the

"open cases" ambiguity in wholesale-2026-08-18).