Ship Decision (2026-05-21)

Candidates compared (OOS 2025-01-01 to 2026-05-21)

#	Strategy	Cohort	N OOS	Sharpe	CI95 Lo	DSR	CAGR%	MaxDD%	Win%
1	V13.5_stack (current LIVE)	full 293k	32,891	0.98	-0.90	0.12	26.2	-12.9	52.1
2	V13.3s_sector_only	full 293k	32,891	1.19	-0.65	0.33	32.2	-13.7	52.1
3	V14e_tuned_light	full 293k	32,891	1.05	-0.97	0.19	25.5	-9.7	52.9
4	V14e_tuned_light	EU_strict 38k	12,384	1.36	-0.46	0.43	38.6	-10.6	55.0
5	V13.3s_sector_only	EU_strict 38k	12,384	1.21	-0.60	0.29	35.0	-10.6	55.0
6	V13.3s_sector_only	EU+BR 81k	23,703	1.12	-0.72	0.22	30.5	-10.6	53.6

EU_strict markets: XPAR, XAMS, XWBO, XBRU, XHEL, XOSL, XSTO, XETR.

Verdict

Ship V14e_tuned_light on the EU_strict universe.

Reasons:

Best Sharpe of any tested combo (1.36) and best DSR (0.43) -- both clear of the +0.0 luck floor by a meaningful margin.
CI95 Lo = -0.46 is the tightest among all candidates; closest to a one-sided positive band.
MaxDD = -10.6% is the lowest of the V13.5+ family (V14e proxy multiplier shaved DD by ~3pp vs V13.5_stack on the same cohort).
Win% 55.0% vs 52.1% on the full cohort: meaningful improvement.
The EU_strict restriction is statistically defensible -- audit 101 documents that the XNAS tape inverts the picker signs (CEO negative, cluster negative on US). Treating the universe restriction as a known-tape filter is honest, not data-mining.
The V14e family multipliers (entity +0.10, family +0.20) survive the cohort change: they helped on BUY-only 60k (Sharpe 1.31) and again on EU_strict 38k (Sharpe 1.36). Same coefficients across two independent cohorts is a non-trivial out-of-sample passage.

Caveats stated upfront:

Family-kind decls (spouse+child+parent = 79 across the full cohort) remain thin. Multiplier statistical power is dominated by entity kinds (controlled 44k, trust 1.7k). Honesty: most of V14e's lift comes from the entity multiplier, not the family one.
EU_strict + mcap >= 100M filter destroys Sharpe (drops to -0.65). The EU small-cap segment is conviction-rich, opposite of XNAS. Do not apply the XNAS-style filter to EU.
The 14-month OOS window is short. The result is a directional signal, not proof.

Rejected alternatives

V13.5_stack live: collapses to Sharpe 0.66-0.98 on expanded cohort due to XNAS pollution. Cannot be defended publicly with the new data.
V13.3s_sector_only (full or EU): viable fallback (Sharpe 1.19/1.21 respectively) but DSR weaker and gives up ~0.15-0.17 Sharpe vs winner with no compensating robustness benefit.
V14e on full cohort: 1.05 Sharpe -- decent but doesn't beat V13.3s on full cohort (1.19). Universe restriction is doing real work.
EU+BR: BVMF adds ~40k rows (Brazil) and drags Sharpe from 1.36 to 1.27. Worse risk-adjusted return for more cohort breadth than the current proof needs.
EU+BR-minus_micro: -0.65 Sharpe. Filter inverts on EU. Reject.

Implementation outline (DO NOT APPLY here)

src/lib/signals.ts::computeV13Score: add EU-allowlist gate, add V14e family/proxy multipliers.
src/lib/winning-strategy.ts::STRATEGY_PROOF.oosResults + monthlyPortfolio: replace V13.5_stack numbers (1.21 / 0.42 / 45.8 / -14.1) with V14e EU_strict (1.36 / 0.43 / 38.6 / -10.6).
Update disclosure copy in winning-strategy.ts::disclosure.multipleTestingRisk to state the EU-only universe restriction and the XNAS inversion finding.
Patch ready at docs/patches/strategy-proof-FINAL.diff.

Next steps (gated on Simon decision)

Apply strategy-proof-FINAL.diff (universe gate + scoring + literals).
Apply site-copy-honest-numbers.diff (user-facing pages).
Apply methodologie-disclosure-update.diff (methodology page).
Re-bake quarterly once data backfill grows family cohort beyond 193 decls -- the family multiplier may earn more weight or none.
Re-test XNAS sub-strategy independently as a V15 research branch (Filter A/B from audit 101) -- don't ship until US-tape signs are independently validated on a fresh dataset.

Files

Audit 100 (V14 expanded cohort): docs/method-review/100-v14-expanded-cohort-2026-05-21.md
Audit 101 (XNAS diagnostic): docs/method-review/101-xnas-diagnostic-2026-05-21.md
Audit 102 (EU-only bake): docs/method-review/102-eu-only-ship-2026-05-21.md
This doc: docs/method-review/103-ship-decision-2026-05-21.md
Patch ready: docs/patches/strategy-proof-FINAL.diff

Candidates compared (OOS 2025-01-01 to 2026-05-21)

Strategy

Cohort

N OOS

Sharpe

CI95 Lo

DSR

CAGR%

MaxDD%

Win%

V13.5_stack (current LIVE)

full 293k

32,891

0.98

-0.90

0.12

26.2

-12.9

52.1

V13.3s_sector_only

full 293k

32,891

1.19

-0.65

0.33

32.2

-13.7

52.1

V14e_tuned_light

full 293k

32,891

1.05

-0.97

0.19

25.5

-9.7

52.9

V14e_tuned_light

EU_strict 38k

12,384

1.36

-0.46

0.43

38.6

-10.6

55.0

V13.3s_sector_only

EU_strict 38k

12,384

1.21

-0.60

0.29

35.0

-10.6

55.0

V13.3s_sector_only

EU+BR 81k

23,703

1.12

-0.72

0.22

30.5

-10.6

53.6

EU_strict markets: XPAR, XAMS, XWBO, XBRU, XHEL, XOSL, XSTO, XETR.

Verdict

Ship V14e_tuned_light on the EU_strict universe.

Reasons:

Best Sharpe of any tested combo (1.36) and best DSR (0.43) -- both clear of the +0.0 luck floor by a meaningful margin.

CI95 Lo = -0.46 is the tightest among all candidates; closest to a one-sided positive band.

MaxDD = -10.6% is the lowest of the V13.5+ family (V14e proxy multiplier shaved DD by ~3pp vs V13.5_stack on the same cohort).

Win% 55.0% vs 52.1% on the full cohort: meaningful improvement.

The EU_strict restriction is statistically defensible -- audit 101 documents that the XNAS tape inverts the picker signs (CEO negative, cluster negative on US). Treating the universe restriction as a known-tape filter is honest, not data-mining.

The V14e family multipliers (entity +0.10, family +0.20) survive the cohort change: they helped on BUY-only 60k (Sharpe 1.31) and again on EU_strict 38k (Sharpe 1.36). Same coefficients across two independent cohorts is a non-trivial out-of-sample passage.

Caveats stated upfront:

Family-kind decls (spouse+child+parent = 79 across the full cohort) remain thin. Multiplier statistical power is dominated by entity kinds (controlled 44k, trust 1.7k). Honesty: most of V14e's lift comes from the entity multiplier, not the family one.

EU_strict + mcap >= 100M filter destroys Sharpe (drops to -0.65). The EU small-cap segment is conviction-rich, opposite of XNAS. Do not apply the XNAS-style filter to EU.

The 14-month OOS window is short. The result is a directional signal, not proof.

Rejected alternatives

V13.5_stack live: collapses to Sharpe 0.66-0.98 on expanded cohort due to XNAS pollution. Cannot be defended publicly with the new data.

V13.3s_sector_only (full or EU): viable fallback (Sharpe 1.19/1.21 respectively) but DSR weaker and gives up ~0.15-0.17 Sharpe vs winner with no compensating robustness benefit.

V14e on full cohort: 1.05 Sharpe -- decent but doesn't beat V13.3s on full cohort (1.19). Universe restriction is doing real work.

EU+BR: BVMF adds ~40k rows (Brazil) and drags Sharpe from 1.36 to 1.27. Worse risk-adjusted return for more cohort breadth than the current proof needs.

EU+BR-minus_micro: -0.65 Sharpe. Filter inverts on EU. Reject.

Implementation outline (DO NOT APPLY here)

src/lib/signals.ts::computeV13Score: add EU-allowlist gate, add V14e family/proxy multipliers.

src/lib/winning-strategy.ts::STRATEGY_PROOF.oosResults + monthlyPortfolio: replace V13.5_stack numbers (1.21 / 0.42 / 45.8 / -14.1) with V14e EU_strict (1.36 / 0.43 / 38.6 / -10.6).

Update disclosure copy in winning-strategy.ts::disclosure.multipleTestingRisk to state the EU-only universe restriction and the XNAS inversion finding.

Patch ready at docs/patches/strategy-proof-FINAL.diff.

Next steps (gated on Simon decision)

Apply strategy-proof-FINAL.diff (universe gate + scoring + literals).

Apply site-copy-honest-numbers.diff (user-facing pages).

Apply methodologie-disclosure-update.diff (methodology page).

Re-bake quarterly once data backfill grows family cohort beyond 193 decls -- the family multiplier may earn more weight or none.

Re-test XNAS sub-strategy independently as a V15 research branch (Filter A/B from audit 101) -- don't ship until US-tape signs are independently validated on a fresh dataset.

Files

Audit 100 (V14 expanded cohort): docs/method-review/100-v14-expanded-cohort-2026-05-21.md

Audit 101 (XNAS diagnostic): docs/method-review/101-xnas-diagnostic-2026-05-21.md

Audit 102 (EU-only bake): docs/method-review/102-eu-only-ship-2026-05-21.md

This doc: docs/method-review/103-ship-decision-2026-05-21.md

Patch ready: docs/patches/strategy-proof-FINAL.diff