# Lab #02 — Playbook v2 (Deterministic reference query)
**Date locked:** 2026-07-22
**Status:** Candidate
## Hypothesis
Replacing Phase 2's model-in-head reference scoring with one local,
quality-graded deterministic query will reduce unsupported reference selection
and unknown-field promotion without adding another user checkpoint.
## Single variable from v1
Only reference discovery changes:
- **v1:** the host model reads fingerprints and scores candidates in-head;
- **v2:** the host runs the shipped local query helper, presents its ordered
candidates and quality labels, and keeps the existing user selection step.
All mandatory checkpoints, personas, agent pool, iteration cap, asset policy,
DESIGN.md variation rules, and handoff behavior remain v1.
## Query contract
1. Resolve one channel-local data root; never mix catalogs across channels.
2. Require fingerprints and the conservative reference-quality manifest.
3. Exact catalog id hints win, but never erase their quality status.
4. Otherwise rank semantic matches, then quality, then stable id order.
5. No recognized signal means `needs_clarification` and zero candidates.
6. `verified_v2` may supply evidence-qualified fields only.
7. `partial` and `legacy_snapshot` are inspiration/context only until reverify;
they cannot populate machine tokens as brand facts.
8. Unknown fields remain absent at the smallest unresolved boundary.
## Acceptance
- exact same input and data produce byte-equivalent candidate ordering;
- every returned id exists in the loaded fingerprint catalog;
- every candidate carries the manifest's exact quality status;
- no Partial/Legacy candidate receives token-promotion permission;
- an unrecognized vague request returns no “safe default” reference;
- EN, KO, JA, ZH-CN, and ZH-TW query fixtures pass;
- missing or stale quality data fails closed with an actionable install/doctor
instruction.
## Rollback
Rollback v2 if exact brand recall regresses, any ungraded reference is presented
as verified, any Partial/Legacy candidate is allowed to seed brand tokens, or
the same frozen query produces different ordering.