Ship and explain the reversible adapter release

From a fresh clone/directory, run the commands that prove held-out tone score ≥ 0.82 and factual hallucination rate ≤ 0.05 on 40 gold pairs. A second person plays support ops manager reviewing tone before a seasonal campaign and follows your script without coaching. The demo includes limitations: what LoRA adapter that keeps support replies in brand voice while facts stay in RAG still will not do.

Bundle: contract snippet, passing tests, metric snapshot for tone_pass_rate and unsupported_fact_rate, security negative probe, rollback note, owner name. Reference FT-PROMO-LEAK-12 as the drill you rehearsed. If any item is missing, the ship gate fails even if the happy path dazzles.

For Fine-tuning lab, name the human who gets paged, the review date for thresholds, and the condition that triggers reevaluation. Endpoint POST /v1/fine_tuning/jobs remains the production surface you operate for LoRA adapter that keeps support replies in brand voice while facts stay in RAG — not a slide. Keep FT-PROMO-LEAK-12 in the handoff template so the next owner inherits the drill.

After the peer demo, schedule the next threshold review date. File a short changelog that mentions FT-PROMO-LEAK-12 and the control that addresses it. Archive the evidence pack where your team already stores launch records. Resist rewriting everything “for real production” in one weekend — operate this slice until the metrics bore you, then widen. Final self-check: if telemetry vanished, would you still know to hold? If yes, you learned the operating posture this lane teaches.

Trim the demo script until every command is necessary. Record a peer signature line: name, date UTC, pass/fail. File known limitations as bullets, not apologies. Link the evidence pack from the README. Schedule the next game day on a calendar, even if it is solo. Archive the branch tag or release digest you actually shipped. In this chapter the product is LoRA adapter that keeps support replies in brand voice while facts stay in RAG, the human stakeholder is support ops manager reviewing tone before a seasonal campaign, and the incident id you design against is FT-PROMO-LEAK-12. Re-state the oracle in your notes — held-out tone score ≥ 0.82 and factual hallucination rate ≤ 0.05 on 40 gold pairs — and keep the invariant visible: weights never encode store hours or prices; those stay in retrieval docs. Track tone_pass_rate and unsupported_fact_rate as the scoreboard. Surface under change control: POST /v1/fine_tuning/jobs. If you only have forty minutes, finish the fixture for training set includes tomorrow's promo price; model invents it after promo ends before polishing UI. Promotion language stays ternary: promote, hold, or roll back based on evidence, not hope.