Skip to content

ADR-0191: Separately admitted relevance reranking

  • Status: Experimental implementation under ADR-0178; activation remains gated
  • Date: 2026-10-12
  • Related: ADR-0178, ADR-0185, ADR-0188
  • Detailed design: docs/plan/general-memory-recall.md

Context

The full-runtime development pilot found a stored coaching routine among permitted candidates but outside the twenty returned items. People context used its existing sparse path, where confidence, review status and reinforcement outweighed stronger relevance. Another source anchor ranked seventh even though its semantic similarity exceeded every distractor, so the three-source recovery bound did not reach it. These are ranking failures, independent of possession or a particular memory type.

Decision

Retain reciprocal-rank fusion as the baseline and add an independent, default-off reranking ablation. After visibility filtering and validated equivalence collapse, compute the strongest of distinct query-term coverage, structured-match strength and nonnegative semantic cosine similarity. Give that signal 0.8 of the final score and the existing pre-fusion metadata score 0.2. The existing duplicate penalty, item/token caps, lifecycle exclusions, confidence labels and uncertainty rendering remain mandatory. A relevance score is not factual certainty.

The dense-only comparison supplies no sparse signals. Sparse and People reads may use their permitted lexical/structured signals; this does not activate semantic People retrieval or widen an identity scope. Verified equivalent originals may transfer their strongest permitted relevance signal, but never create extra votes. Historical questions retain their existing revision selection and visibility rules. Frozen controls using recent candidate selection keep their original ranking.

The profile names reranking: baseline or relevance. A separate admission callback is mandatory, even for a relevance profile. Production composition supplies the same version-bound release admission used by the semantic services; missing or withdrawn evidence restores the baseline. Explicit isolated evaluation may supply its own admission for fabricated inputs. Traces identify relevance-rerank@1 only when it actually ran. Reranking adds no provider call, query, source expansion or model cost.

Evidence and limits

A regression with one provisional, inferred, review-flagged fact and eighty newer high-confidence neighbors failed for semantic, linked and lexical signals. The experimental reranker places the stronger fact first without changing its authority or lifecycle labels. Withdrawal restores the previous order. Full comparative quality, cross-family regressions and operational evidence remain required before activation. Raw cosine and lexical coverage are not calibrated probabilities; this transparent formula is an experiment, not a claim of optimal ranking.