82 · V13.3s wired into the live picker (2026-05-19)

TL;DR

Until this change the live recommendation engine (src/lib/recommendation-engine.ts) ranked candidates by Declaration.signalScore only (V12, 10-factor composite). The V13.3s formula (V13.1g_stacked + V13.2 earnings overlay + V13.3 sector momentum multiplier) had shipped inside computeV13Score() (src/lib/signals.ts) and backtested to Sharpe 1.15 / DSR 0.53 / CAGR 43.2 % / MaxDD -14.1 % on the same OOS harness as V13.1g, but it was not reaching users.

This audit captures the wiring change: persist a new Declaration.signalScoreV13 column, bulk-rescore the whole DB with computeV13Score(), and switch the engine sort key to COALESCE(signalScoreV13, signalScore) DESC so the V13 alpha lands in production. Legacy signalScore is preserved as a rollback safety net.

Change set

Layer	File	Change
Schema	prisma/schema.prisma	New Declaration.signalScoreV13 Float? + @@index
Migration	prisma/migrations/20260519600000_declaration_signal_score_v13/migration.sql	Committed, not auto-applied. prisma migrate deploy runs it on CI/prod gate
Backfill script	scripts/_rescore-all-v13.ts	Batches of 2000 rows, idempotent, --apply gate, bulk UPDATE ... FROM (VALUES ...) for speed
Engine BUY ranking	src/lib/recommendation-engine.ts (getBuyRecommendations)	ROW_NUMBER() OVER (... ORDER BY COALESCE(signalScoreV13, signalScore) DESC, pubDate DESC)
Engine SELL ranking	src/lib/recommendation-engine.ts (getSellRecommendations)	Converted to raw SQL, same coalesce
Engine personal	src/lib/recommendation-engine.ts (getPersonalRecommendations)	orderBy: [{ signalScoreV13: { sort: 'desc', nulls: 'last' } }, { signalScore: 'desc' }]
Type	RecoItem.signalScoreV13?: number | null	Persisted on payload so the UI can label "V13" badge later
Strategy proof	src/lib/winning-strategy.ts	oosResults + monthlyPortfolio updated to V13.3s values

signalScore (V12) is intentionally left untouched. If V13 needs to be rolled back, drop the column or NULL it out and the engine falls through to V12 with no other code change.

Why persist, not compute on the fly

The V13.3 multiplier needs sectorReturnPre180d (already persisted), the V13.2 overlay needs earningsProximityScore (already persisted), and V13.1g_stacked needs clusterParticipantCount (±30d, same company, distinct insiders) and insiderRecentAlpha90d (mean priors r90d for the same insider, strictly before pubDate). Computing the last two per request would require a window scan and a priors join inside the hot path of /recommendations. Persisting the composite into one indexed column keeps the engine's candidate-fetch query index-bound (Declaration_signalScoreV13_idx + the partition-key composite).

OOS stats moved into STRATEGY_PROOF.oosResults

From V13.1g (audit doc 78) to V13.3s (audit doc 80, wired here):

Metric	V13.1g (prev live)	V13.3s (new live)	Δ
Sharpe (annualised)	0.70	1.15	+0.45
Deflated Sharpe (N=11)	0.33	0.53	+0.20
CAGR	25.4 %	43.2 %	+17.8 pp
Max DD	-29.7 %	-14.1 %	+15.6 pp
Hit rate (months)	55.7 %	57.1 %	+1.4 pp
Bootstrap CI95 Sharpe	[-1.18, 3.64]	[-0.62, 3.88]	width narrows

CI95 still straddles zero — the lift is meaningful but the sample (14 monthly buckets, n=140 picks) is too thin to be statistically robust against selection bias. Surfaced via the existing sharpeDisclosure() helper.

Cache invalidation

Searched src/: no revalidateTag() callers for reco-general / reco-sells / picks-cache / home-data. The caching layer uses unstable_cache with TTL windows (e.g. sectors, top-movers, clusters/recent). The reco page itself runs server-rendered against Prisma with no TTL cache wrapper, so the V13 cutover takes effect on the next request after the rescore completes — no manual purge needed.

Rollout

1. Land schema + script + engine changes (this PR).
2. CI green + Vercel deploy.
3. On prod: npx prisma migrate deploy (idempotent — applies the V13 column).
4. On prod box: tsx scripts/_rescore-all-v13.ts (dry, sanity-check sample).
5. On prod box: tsx scripts/_rescore-all-v13.ts --apply (writes signalScoreV13).
6. Watch /recommendations: top-10 BUY cards re-rank as V13 column populates.

Rollback path: drop the column or UPDATE "Declaration" SET "signalScoreV13" = NULL. Engine COALESCE falls back to V12 silently.

Test sample · top BUY picks before vs after

To be captured post---apply on prod. The audit will be amended with the diff table (top 10 V12-ranked vs top 10 V13-ranked) once the backfill completes. Expected pattern from the bake-off: 2 of 14 OOS months saw cohort reranking under V13.3s, and both flipped months were net wins. Day-1 live picks will likely look very similar to V12, with sector-tagged BUYs (17.5 % coverage) either lifted (mean-reversion band) or fade-dampened (euphoric sector band).