Calibrated Abstention + Faithful Span Verification

File: src/handlers/recall.rs (abstention_decision) · src/handlers/verify.rs (verify_claim)

The problem

An agent memory that answers with a confident-looking wrong answer is worse than one that says “I don’t know.” Retrieval systems must know when to refuse. And a claim-verification step must be faithful: it should point at the exact span of text that supports a statement, not gesture vaguely at a document.

The reference

• Calibrated abstention — driven by a multi-signal estimator, not a magic score < 0.3 cutoff. The signal is the existing HeuristicEstimator’s Recommendation::ClarifyQuery (overlap + gap + lexical-density agreement across retrievers). This is the roadmap-required form: “abstain when the evidence is genuinely ambiguous.”
• Deterministic span verification — the honest, low-cost way to check a claim: case-insensitive substring match against a chunk’s text with byte-offset match ranges.

The implementation

1. Abstention (v1.5.0): when the estimator emits ClarifyQuery, /recall returns {decision: "low_confidence", hits: []} instead of top-1 garbage. Zero new compute — confidence + recommendation were already computed by the retrieval pass; abstention_decision() is a pure helper. v1.12 (Discern) added the graph-rescue before abstaining (see 05-hub-dampening.md).
2. POST /verify (v1.5.0): {chunk_id, claim} → {supported, decision, match_ranges}. Case-insensitive substring match over one chunk, O(content), no embeddings, no LLM. Bounded: MAX_QUERY (2000) on claim, MAX_MATCH_RANGES (100) on output. It reuses the /get/{id} SQL shape — one query, no new schema.

Measured ceiling

• Abstention is heuristic, not learned — ClarifyQuery is calibrated on rank-agreement signals, not a judged corpus. A judged corpus (brain eval --floor) is the operator step that turns it into a measured claim.
• /verify is lexical only — no semantic/paraphrase match. “Faithful” means the span literally appears in the text, which is exactly the right guarantee for a verifiable memory store, and exactly the wrong tool for paraphrase.
• /verify records no audit row (pure read) — reads are audit-able via the opt-in read-event audit (v1.15).

This is the “say ‘I don’t know’ in a way a reviewer can verify” story from the blog.