What the human review found
The submitter reviewed all 150 saved drafts: 54 good, 66 acceptable and 30 not acceptable. Notes highlighted generic festival wishes, awkward no-reply closings, abrupt care transitions and product questions left unanswered. These verdicts describe the saved v7.5/v7.6 outputs, not the current backend.
What changed after review
v7.7 requires every candidate to answer the same named-product question with the requested known facts. Greetings follow the occasion, with delivery timing shown separately to the advisor. No-reply cliches are checked, and irrelevant care guidance is no longer injected into every scenario. v7.7.1 corrected an overly narrow Mid-Autumn theme check.
The latest refinement / v7.8
Further feedback identified the unnatural word 松快 and unsolicited prices in product introductions. v7.8 excludes that word and omits product prices unless the current task explicitly asks about price or budget. A stored profile budget does not authorise quoting. An explicit request to omit prices wins; confirmed price-objection replies remain supported.
What has actually been verified
Six targeted cases produced eighteen rule-passing drafts, selected from five v7.7 runs and one v7.7.1 run. The initial Mid-Autumn false positive is retained. v7.8 passes 92 local engineering tests, but has no new real-model evaluation batch or human ratings. Neither result establishes an overall improvement in writing quality.
The next decisions
Generate targeted v7.8 outputs for greetings, product introductions, explicit price questions and price objections; review them independently without copying old ratings. Check unsupported assumptions, names and natural transitions. Then broaden unseen intent-routing tests, catalogue matching and latency observations. The static showcase is built; public access still needs deployment checks and screenshots will be added by the submitter.
Evidence boundaries
- Original batch: 50 cases / 150 human ratings.
- Targeted v7.7?v7.7.1: 6 cases / 18 rule-passing drafts.
- v7.8: 92 local tests; new output review pending.
- Real-model routing accuracy: not established.
- Showcase built; public access not yet verified.
After the human review
The new constraints prioritise answering the named product, occasion-specific wishes, separate delivery timing and natural care transitions. A Mid-Autumn synonym false positive was found and corrected during the targeted checks. Six cases produced eighteen rule-passing drafts in v7.7?v7.7.1; they have no new human ratings. The later v7.8 word-choice and pricing changes have local test coverage only; they are not represented by those eighteen drafts.