search: R2 salience ranking boost needs a config gate and an outcome gate (or removal)

[itlackey]

Problem

The R2 boost (asset_salience.rank_score → search/curate ranking, ×[1.0–1.2]) shipped default-on in 0.9.0-beta.53 with no config opt-out — only the internal salienceRankScores: null API param can disable it. On live data it is both weak and biased:

• asset_salience has zero pack refs (%//% = 0 rows), while externally-installed packs dominate actual consumption (top shows AND curates). The boost can only ever favor self-generated personal assets — worst case ~+7% for an unread knowledge note over an equally-relevant pack asset.
• For 5,560/5,601 non-content-scored rows the encoding term is a per-type constant, so the boost mostly reduces to "is a personal skill/agent".
• Its dominant input (retrieval, w_r=0.60) counts search impressions as retrievals with no source filter — a second impression→rank loop on top of the utility EMA.
• The utility EMA contributor (weight 0.5) already delivers usage reinforcement from the same usage_events, with full pack coverage — R2 double-counts it for personal assets only.
• Live effect is ~noise: max rank_score 0.356 → max multiplier ×1.071, mean ×1.016.

Fix options (in preference order)

1. Remove the salience contributor from user-facing search ranking (keep rank_score for improve-internal selection) until a measured curate-golden-bench delta justifies it.
2. If kept: add a config gate (default off), and gate the boost on retrieval-backed outcome signal (asset_outcome.outcome_score > 0 — column exists, ~541 assets qualify) so it stops double-counting impressions and favoring the unread corpus.

Either way, a config opt-out should exist for a term that touches every query on the hottest read path.

🤖 Generated with Claude Code

https://claude.ai/code/session_018YjU1mGiShdDP4qFW7p4mv

[itlackey]

Triage verification @ 0.9.0-beta.57 (main): still open, not addressed.

Re-checked the current ranking path against this issue's asks (config gate, outcome gate, or removal). None are implemented:

• Boost still default-on for both search and curate. salienceRankingContributor is registered in defaultUtilityRankingContributors (src/indexer/search/ranking-contributors.ts:446-449) and multiplies score by min(1 + rank*0.2, 1.2) for any rank > 0 (ranking-contributors.ts:433-444; SALIENCE_WEIGHT = 0.2, SALIENCE_MAX_BOOST = 1.2 at :34-35). The live call site applyRankingRules(...) in src/indexer/search/db-search.ts:459-468 doesn't pass salienceRankScores, so it falls through to the default asset_salience load — active on every user-facing search/curate.
• No config opt-out. No key in src/core/config/config-schema.ts or schemas/akm-config.json disables the boost. (salienceThreshold at config-schema.ts:686 gates the unrelated improve-admission lane, feat(improve): automatic salience scoring #608.) The only disable path is still the internal RankEntriesOptions.salienceRankScores: null API param (src/indexer/search/ranking.ts:65,213-216) — exactly the mechanism this issue flagged as insufficient.
• Still gated on raw rank_score > 0, not outcome. loadSalienceRankScores reads SELECT asset_ref, rank_score FROM asset_salience (src/indexer/search/ranking.ts:103) with no join to asset_outcome.outcome_score.

No CHANGELOG entry for beta.53→beta.57 references a salience config/outcome gate. Keeping open — accurate as filed; note the code now lives in src/indexer/search/ranking-contributors.ts / ranking.ts for whoever picks it up.

---

Generated by Claude Code