SearchScore Labs / EXP-0001
Rejected ◆ Recommendation

Coverage-discount guard for homograph false positives

a null/negative result — published in full.

Hypothesis

Discounting a recommendation’s confidence by how much of the query the matched pattern explains ("coverage") will eliminate confident homograph false positives (e.g. "budget my salary" → Crawl Budget).

Method

Add a coverage factor to matchConfidence, gated to low-overlap queries. Sweep COVERAGE_EXP ∈ {0,0.5,1,1.5,2} × overlap-gate {0.5–0.8} over a 500-query adversarial set. Compare homograph FP, answer recall, ECE, and check flagship queries for regression.

Dataset

500-query adversarial sweep (18 homographs, GEO/SEO/commercial in-scope, foreign out-of-scope).

Metrics

homograph FP rate≤ 1%
answer recall≥ 80%
flagship-query regressionnone

Result

{ "off": { "recall": 0.829, "homographFP": "14/18", "ece": 0.133 }, "exp1.5_gate0.6": { "recall": 0.732, "homographFP": "5/18", "ece": 0.044 }, "flagship": "\"why isn’t ChatGPT recommending my company\" REFUSED at every strength that dents homographs (its coverage = 0, lower than the homographs’)" }

Decision

REJECTED. No lexical operating point separates homographs from legitimate low-overlap GEO queries. Gate disabled by default; matchCoverage retained as telemetry.

Lessons learned

  • Homograph disambiguation is a semantic problem; lexical coverage cannot solve it without refusing flagship queries.
  • A rejected experiment is a result: it redirected effort to embeddings (EXP-0003) with evidence instead of guesswork.

Confidence

HIGH

Evidence

500-query sweep — bench/history (calib-eval)confidence model — docs/kos-confidence-model.md
← all experiments