Debug invented citations and overconfident prose

Map failure classes for POST /v1/research/brief: retryable vs fatal vs needs-human. Retries need budgets, jitter, and idempotency rules aligned to claim without evidence → abstain or drop claim; no orphan sentences. The chapter’s signature failure — model invents kb://minutes-2099 citation — must take a deliberate branch, not a generic catch-all.

Implement the degrade/rollback/refuse behavior analyst compiling a brief on bike-share policy from a fixed corpus needs when CAP-RESEARCH-FAKE-ID-3 repeats. Prefer scoped controls (one flag, one weight, one tenant, one secret version) over fleet-wide restarts. Preserve evidence; do not delete logs to “clean the demo.”

After containment, check claim_citation_coverage and fabricated_id_rate moves in the safe direction and watch for retry amplification. Write the stop condition that ends the incident response for this lab.

Chaos note: inject only one fault class at a time and restore fixtures after. Watch for dual failures — dependency down and retry amplifier — which is how model invents kb://minutes-2099 citation becomes an outage. Customer communication templates (even if only for the drill) beat silence. If you queue deferred work, define poison-message handling. Budget documents should state the maximum extra spend allowed during retries. Close the loop by linking the containment action to a dashboard panel for claim_citation_coverage and fabricated_id_rate.

Draw a state diagram for degrade modes and put it in the repo as ASCII if needed. Cap concurrent retries across the process, not only per request. Ensure cancellation propagates to downstream HTTP clients. When failing closed, choose a user-visible message that does not leak internals. Practice the single command that flips the kill switch or weight to zero. After recovery, drain or inspect deferred work before declaring green. In this chapter the product is citation-first research assistant that maps every claim to evidence IDs, the human stakeholder is analyst compiling a brief on bike-share policy from a fixed corpus, and the incident id you design against is CAP-RESEARCH-FAKE-ID-3. Re-state the oracle in your notes — brief on fixture corpus: ≥ 0.95 claims cited; fabricated ID rate 0 — and keep the invariant visible: claim without evidence → abstain or drop claim; no orphan sentences. Track claim_citation_coverage and fabricated_id_rate as the scoreboard. Surface under change control: POST /v1/research/brief. If you only have forty minutes, finish the fixture for model invents kb://minutes-2099 citation before polishing UI. Promotion language stays ternary: promote, hold, or roll back based on evidence, not hope.