[Air000000]

Scope

Freeze an inert prospective comparison, not another prompt: v5.2.1/v5.3 prompts and strict validator remain unchanged before any new labels. Reuse frozen450 TRAIN snapshot/results and full ordered Top14; exclude the prior selection's recorded exclusions/family exclusions/selected30, then normalized-question/leading-document proxy against old cases. Deterministic SHA256 seed priority, within-batch same-question/leading-document dedup, exactly50, no smaller fallback/post-label balancing. Eligible114 before within-batch dedup; all selected50 have distinct leading documents. No new retrieval, gold-based selection or DEV.

Interpretation

Fresh relative to recorded refusal selections/visible family proxy, NOT untouched project data: all TRAIN previously used in retrieval/automated analysis; leading-document proxy does not guarantee independent semantic families. Proposed Qwen drafts plus current-assistant preprediction evidence review remain AI proxy, not independent expert gold. Report actual class support; no label/sample changes to reach the old minimum-five/class rule. No formal acceptance/production gate, generated-answer accuracy or unseen DEV claim regardless of results.

Proposed two-stage authorization (NOT executed)

1. Up to50 Singapore qwen3.6-plus-2026-04-02 annotation calls, max4096 output tokens, existing blind prompt. Assistant reviews questions/drafts/cited evidence with recorded depth, audits recoverable citation formatting without overwriting originals. Freeze all50 reviewed labels/counts/hashes before any classifier prediction. If defensible complete labels cannot be frozen, stop before classification.
2. Up to100 fresh paired Singapore qwen3.5-plus-2026-04-20 calls,50 per arm, unchanged payload/schema/validator, max2048 output tokens. Fixed question order and alternating arm order. No target labels, review reasons, IDs, retrieval metadata or secrets in classifier requests.

Both stages use temperature0/thinking off/JSON, SDK retries0; at most150 calls total, no paid retries. Proposed stage ledger caps annotation CNY10, classification CNY15, total CNY25, no transfer/recharge. Planning estimates annotation CNY7.6092975 + classification CNY10.848336 = CNY18.4576335, using official International rates checked2026-10-16 and fixed FX7.5; byte/framing proxy is not certified tokens or an invoice guarantee. New isolated durable checkpoints, pending before POST, raw/usage before parsing, exclusive lock; saved-raw recovery without new calls, no repeated completions, unknown charge blocks resume. Classifier schema errors remain terminal null outcomes. New explicit authorization required for both data/model/stage/budget scopes.

Artifacts and verification

• Reuse selection, full-context builders, annotation prompt and existing dependencies; no paid runner/client/checkpoint framework added.
• Bind source/protocol/request/prompt/input/packet/previous-selection hashes. Only public aggregate metadata/QIDs/fingerprints; real questions/source bodies/blank annotation packets remain under ignored data/.
• 317 offline tests with external-network guard and temporary DB passed; Ruff/in-memory compilation/diff checks passed.
• Selection, unique-leading-document50, private payloads and preflight reproduce exactly; previous v5.3 contract/journal hashes unchanged.
• Zero provider calls; no .env read, labels, predictions, DEV, relabeling, serving or primary dirty-checkout changes.