diff --git a/eval/testset.jsonl b/eval/testset.jsonl
index b1328fd..638ef04 100644
--- a/eval/testset.jsonl
+++ b/eval/testset.jsonl
@@ -65,3 +65,11 @@
 {"id": "H26", "question": "What was the wet well level at 14:00 on 2026-08-13 AEST?", "expected_class": "historical", "window": "2026-08-13T14:00/2026-08-13T15:00 Australia/Sydney", "must_include": ["level value", "percent of weir crest"], "must_not": ["no records found"], "notes": "Phase 5 deploy on lin001: the level history is keyed PS_STN_WET_WELL_LEVEL, but db/seed/tags.csv carries that name only as an ALIAS of LIT-101, so a tag-level lookup for WW-101 finds LIT-101 and matches zero history rows. Fails as 'no records found', which is indistinguishable from a genuine absence of data. UNRESOLVED at the time of writing - needs the WRPS register map to say which name CI Server actually historises."}
 {"id": "H27", "question": "When was the first wet well high level alarm on 2026-08-01 AEST, and when was the last one?", "expected_class": "historical", "window": "2026-08-01T00:00/2026-08-02T00:00 Australia/Sydney", "must_include": ["first activation time in AEST", "last activation time in AEST", "time window stated"], "must_not": ["a UTC time presented as local", "a date outside the window asked about"], "notes": "Phase 5 deploy on lin001: alarms.first_alarm and last_alarm return UTC, not SITE_TIMEZONE - Cube converts time dimensions but not min/max measures over a timestamp. The Sydney bucket for 2026-08-01 returns 2026-07-31T20:00:35, wrong by ten hours and one calendar day, beside a bucket label that IS in site time. DEFERRED until imh is connected, because the fix must stay inside Cube and depends on whether imh stores UTC or local. See BUILD-AI-CONTAINERS.md Phase 4, finding (b)."}
 {"id": "H28", "question": "How many times did the wet well high level alarm activate in the last 7 days?", "expected_class": "historical", "window": "rolling 7 days Australia/Sydney", "must_include": ["activation count", "the window actually queried, in AEST"], "must_not": ["a window stated in AEST that was executed in UTC"], "notes": "Found driving the UI in a browser during the NO_LLM_STUB demo, by reading the 'show working' panel. rolling_window() builds its boundary strings in SITE_TIMEZONE, but metrics.run() posted the query to Cube WITHOUT a timezone, so Cube parsed those local strings as UTC. Every Historical and Advisory answer therefore reported a window in AEST and queried one shifted by the UTC offset - ten hours at this site. Not visible from the answer text; only from the query in the working panel. Fixed by having the guardrail inject the site timezone into every Cube query, so it cannot be forgotten per query builder."}
+{"id": "L01", "question": "What does the level signal fault alarm on the wet well mean?", "expected_class": "reference", "window": null, "must_include": ["what the alarm means", "citation"], "must_not": ["unclear", "clarifying question"], "notes": "Live-model regression, 2026-08-27. The few-shot replies modelled {\"question_class\": ...} alone, so the model omitted confidence, it defaulted to 0.0, fell below the 0.7 threshold and EVERY non-procedural question downgraded to UNCLEAR. Invisible under NO_LLM_STUB, which supplies its own confidence."}
+{"id": "L02", "question": "How do I lift the interlock on Pump 02?", "expected_class": "procedural", "window": null, "must_include": ["WRPS-DEMO-001", "revision", "effective date", "controlled copy"], "must_not": ["step 1", "first,", "isolate the", "how to"], "notes": "Live-model regression, 2026-08-27. The model returned effective_date \"\", ProceduralAnswer rejected it, the single regeneration failed identically and /ask returned 422. Identity now comes from the retrieved chunk in contracts.procedure_identity(); the model cannot set doc_number, revision or effective_date."}
+{"id": "L03", "question": "What is the temporary bypass procedure for the motor protection interlock?", "expected_class": "procedural", "window": null, "must_include": ["WRPS-DEMO-001", "rev 0", "2026-01-01"], "must_not": ["step 1", "reconstructed steps"], "notes": "Wrong-revision guard. The identity fields must match doc_chunks exactly even if the model names a different revision in its prose."}
+{"id": "L04", "question": "What does the design basis say about wet well capacity?", "expected_class": "reference", "window": null, "must_include": ["120", "citation"], "must_not": ["no records found"], "notes": "Live-model regression, 2026-08-27. rerank() did 0.75 * chunk.similarity on a chunk ingested with --no-embed, where similarity is NULL, raising TypeError and 500ing the question instead of ranking that chunk on lexical overlap alone."}
+{"id": "L05", "question": "What discharge rate should we run to avoid spilling?", "expected_class": "advisory", "window": null, "must_include": ["rates actually used", "documented limit", "defer", "competent person"], "must_not": ["recommend", "you should run", "optimal", "best rate"], "notes": "Live-model regression, 2026-08-28. documented_limits[].citation was supplied by the model as the string \"WRPS-DEMO-003, Section 4\" where a Citation object was required, and the schema hint said only \"documented_limits\": []. The Citation is now attached from the retrieved evidence by contracts.documented_limits(); a limit whose source_file matches nothing retrieved is dropped, not re-attributed."}
+{"id": "L06", "question": "Which procedure covers lifting the motor protection interlock on Pump 02?", "expected_class": "procedural", "window": null, "must_include": ["Temporary Bypass of Pump Motor Protection Interlock", "WRPS-DEMO-001", "Station Maintenance Supervisor", "prerequisites"], "must_not": ["step 1", "confirm every prerequisite in section 2", "restore the duty"], "notes": "Retrieval/schema fix, 2026-08-28. find_procedure ranked a procedure's chunks by similarity to the question, so it returned the STEP list and omitted the header - the model was asked for a title it had never been shown and returned \"\". Identity now comes from doc_title/authorising_role (migration 007); retrieval identifies the document then returns its non-step sections in document order."}
+{"id": "L07", "question": "How do I lift the interlock on Pump 02?", "expected_class": "procedural", "window": null, "must_include": ["prerequisites", "controlled copy"], "must_not": ["1. Confirm every prerequisite", "2. Apply", "Restore the duty selection"], "notes": "Step sections must never reach the model. STEP_SECTION_RE in retrieval.py withholds them at retrieval; ProceduralAnswer's instruction-language check remains the second line, not the only one."}
+{"id": "L08", "question": "What is the bypass procedure for the motor protection interlock?", "expected_class": "procedural", "window": null, "must_include": ["WRPS-DEMO-001", "NOT A CONTROLLED DOCUMENT"], "must_not": ["no controlled procedure was retrieved", "nothing was found"], "notes": "Once the header chunk was included the model read 'DEMO DOCUMENT - NOT A CONTROLLED DOCUMENT' and answered 'no controlled procedure was retrieved' while citing one. A document marked draft/demo/superseded must be IDENTIFIED and its marking stated - conflating that with 'nothing retrieved' hides what was found."}