Implements §2 and §3 of the parent design — the whole of Thread A. The two rungs ship together because they share one story and neither blocks the other: the layering rule gives each surface a different legal lever, and this part supplies both.

§2, the interactive surface. The daemon's result note currently coaches "answer 'nothing found' only after distinct vocabularies have failed" on empty results, and every one of the 45 recorded failures happened on non-empty results. Across 1,795 served calls, 79.9% never set mode, and the two negatives missed in all six replicates never escalated once. The memory_search tool description and the non-empty result annotation gain the answerhood check the probe validated: verify a note answers rather than relates, treat a related note as no answer, and treat "nothing answers this" as a correct and expected outcome. Exact wording is frozen before the mini-gate runs. This deliberately exercises the previous design's own re-audit trigger, which named a description change as the cheapest lever its investigation had not tried.

§3, the deliberate path. One Haiku call per search, taking the natural question and the candidate set. The question matters at the level of architecture: on the same instrument, a gate reading the reduced tool query fixed 8.9% of failures where one reading the natural question fixed 86.7%, so a gate placed where only the query is visible lands in the first row. Excerpting follows the probe's corrected instrument — IDF-weighted head plus best-middle plus tail, notes under ~3.5KB shown whole — because the thin version of that instrument mislabelled 43.2% of its apparent over-rejections. Output is a verdict attached as a label; every note stays present and readable. The prompt must handle derived answers, where strict answerhood preserved only 58.7% of episodic-temporal questions, and it is iterated offline on that slice before being frozen.

None. Both rungs are independent of Thread B and of each other.

§2 elicitation — a mini-gate of the 20 negatives plus a fixed 15-question answerable canary sample, stratified and chosen before any run, n=6 replicates, ~$21:

• Mean rejection ≥80%, against a 62.5% baseline whose replicate spread was 55–75% — roughly 2.5σ up.
• Canary answerable mean within one question of its own baseline slice.
• Both clauses hold, with the arm difference checked by exact permutation test. rc02 is this rung's watchlist case.
• Rebuild, reinstall, and launchd kickstart are part of the rung: launchd runs the installed binary, so a cutover that skips the reinstall measures the old path.

§3 labeller — deterministic offline replay on the frozen corpus, corrected instrument, all six replicates' served candidates:

• ≥80% of negative trials have every candidate labelled drop. Current evidence is 82.5% on the two fully re-checked replicates, before the derived-answer refinement, which loosens the gate and may push this down.
• ≥90% of answerable trials keep the expected note. Current: 86.6%.
• Episodic-temporal slice ≥80% preserved. Current: 58.7%. This clause is where the refinement either works or the rung refutes.
• A consumer whose labeller call fails renders unlabelled output with a visible degrade marker, mirroring the embedder's degrade contract.

This part implements one slice of AgentM Rejection and Vocabulary (Status: final). See the parent for Context, Alternatives Considered — in particular why the deleting keep/drop gate was priced and declined — the Quality Attributes overview, and the Operations strategy. Mid-execution changes to this part's scope must be appended to the parent's Document History.