bakery-rubric

name	bakery-rubric
description	Find ambitious bakeries producing from raw components on-site, resist par-bake and discovery bias, and return an evidence-grounded shortlist for a requested location.

Scratch-Bakery Scorecard & Travel Prompt (v8.14)

Find bakeries that are made from raw components on-site AND interesting to a frequent buyer. The hidden attribute is production ambition: serious bread, lamination, pâtisserie, chocolate, or locally relevant specialist craft rather than frozen or par-baked inventory finished for display.

Scratch production is the eligibility gate. Interestingness distinguishes survivors. Ratings are a quality floor, not the ranking signal. The goal is a trustworthy shortlist—not an objective winner and not paperwork that merely looks complete.

What this skill must resist

• Par-bake invisibility: photographs cannot distinguish frozen dough from scratch production.
• Prominence bias: broad popularity-neutral discovery protects obscure specialists.
• Map incompleteness: targeted quality, scratch, local-language, recent-opening, adjacent-category, and specialist searches recover prominent or unlisted misses.
• Format and register prejudice: chocolate shops, ethnic specialists, cottage bakers, preorder operations, and plain-spoken traditional bakeries are judged from production evidence, not labels or jargon.
• Missing-evidence collapse: absent process language is a research state, never a low craft score.
• Delegation drift: workers retrieve quotations and literal values; the primary orchestrator makes every judgment.

Version line

This skill and the restaurant skill share one version and bump together.

v8.14: both rubrics now preserve positive production evidence through provenance-labeled partial scores: unknown criteria remain unknown rather than becoming zero or blocking all scoring. Bakery partials may credit a narrowly proven component such as a house filling without transferring that evidence to dough, lamination, bread, or unrelated products. Existing complete-score calibration, the ≈4.3★ bakery quality gate, and the rating-unconfirmed tier are unchanged.

v8.13: canonical schema-validated 06-decisions.json is the Phase 6 handoff; optional generated Markdown is non-authoritative; completed runs may offer the separate local-first /interactive-results skill. The calibrated rubrics and rating floors are unchanged.

v8.12: shared travel-time scope, identity-readiness routing, and adaptive durable evidence dispatch. Restaurant-specific evidence, decision, validation, and diversity rules do not alter bakery scoring. The calibrated rubrics and rating floors are unchanged.

v8.11: bounded identity reconciliation for locally relevant item roundups; product × adjacent-format discovery; targeted falsification of user-reported omissions; and provenance-labeled aggregator-attributed provisional ratings. The calibrated rubric and rating floor are unchanged.

v8.10: operational contract hardening; synchronized canonical placeholders and unique leaf-batch outputs; worker-side identity-first direct-place rating and current-access collection; explicit acceptance/rendering gates; pytest-compatible contract validation; and rendered-export visual inspection.

v8.9: phased execution; just-in-time phase loading and visible completion gates; mandatory broad-survey plus adaptive multilingual targeted discovery; a canonical verbatim evidence-only worker prompt; semantic return acceptance; original-worker targeted repair when available; orchestrator-only scoring; and a post-score coverage audit. The calibrated rubric is unchanged.

v8.8 and earlier: model tiering; exhausted-unavailable; recursive fan-out; embedded evidence floor; boundary standardization; completion gating; popularity-neutral enumeration; market scarcity; execution; register independence; and category-conditional bakery craft. Their full unchanged definitions and calibration history are in phase-6-scoring.md.

Primary-orchestrator authority (MUST)

You are the sole orchestrator and judgment layer for the entire run.

• You MUST personally perform classification, positive disqualification, S/I/E/R handling, G/G′, scarcity, tiers, ties, occasions, final confidence, and rendering.
• You MUST NOT delegate scoring or ask a worker to “apply the rubric.” A worker verdict is invalid even when plausible.
• You MUST use phase-4-worker-prompt.md verbatim and may substitute only its declared placeholders.
• You MUST inspect every returned venue semantically. Row counts never establish evidence quality.
• You MUST message or resume the original worker for targeted repairs when the runtime supports it; use a fresh worker only under the Phase 5 fallback rules.
• You MUST NOT render recommendations before Phases 1–7 pass.

Just-in-time phase rule (CRITICAL)

Before executing each phase, you MUST read the named phase file or files in full immediately before doing that work. The summaries below are orientation only and are insufficient for execution. Do not read all phase files once at the beginning and rely on memory. Do not skip, combine, reorder, or silently narrow phases.

Before advancing, run the printed completion gate, mark every item ✅ or ❌ with concrete evidence, and show the concise checklist to the user. Any ❌ means remain in the current phase.

Phase map

1. Scope, run directory, and catchment Read ../reference/phase-1-scope-and-catchment.md. Before any research, create the unique <YYYY-MM-DD>-<location-slug> run directory and manifest; every later artifact stays there.
2. Candidate discovery Read ../reference/phase-2-candidate-discovery.md and discovery-reference.md. Run broad, targeted, visible-head, multilingual, adjacent-category, and marker-item discovery.
3. Discovery convergence Read ../reference/phase-3-discovery-convergence.md. Union, normalize, deduplicate, inspect gaps, and require a zero-new-candidate pass or a precise limitation.
4. Evidence research Read ../reference/phase-4-evidence-research.md and phase-4-worker-prompt.md. Dispatch the canonical prompt verbatim.
5. Evidence acceptance and repair Read ../reference/phase-5-evidence-acceptance.md, ../reference/shared-status-and-provenance.md, and phase-4-worker-prompt.md. Inspect every record and repair defects.
6. Bakery scoring and classification Read phase-6-scoring.md. Apply the current eligibility and partial-scoring policy yourself.
7. Coverage audit Read ../reference/phase-7-coverage-audit.md and discovery-reference.md. Newly found bakeries loop through Phases 4–6; repeat until the audit converges.
8. Rendering Read phase-8-rendering.md only after every earlier gate passes.

Global hard stops

• No single-source “complete census” claim.
• No targeted search result treated as qualification evidence.
• No missing evidence converted to a low score or disqualification.
• No worker-created DQ, score, tier, ranking, or confidence accepted.
• No product noun alone treated as process evidence.
• No final occasion matrix or ranked shortlist before the Phase 7 gate passes.
• No permission-seeking handoff of unfinished candidates.
• No durable run artifact outside the Phase 1 {RUN_DIR} and no reuse or overwrite of an earlier run directory.

Begin with Phase 1 for the location supplied by the user.