Follows reports/jev_verdict.md. Local only, reuses the 341 saved
proxy-history verdicts (validation/jev_deepdive.py).
Headline: the proxy +1.02% is a momentum-echo artifact, not a Jev edge
The decisive test is §2b. Jev's absolute veto (avg 3.86 of 8 names dropped) gives D = +1.02%; simply dropping the bottom-4 momentum names gives +0.88%. The per-day rank veto curve and the momentum-drop curve track each other almost exactly:
| dropped | Jev rank-veto D | drop-worst-momentum D |
|---|---|---|
| 2 | +0.27% | +0.20% |
| 3 | +0.55% | +0.48% |
| 4 | +0.95% | +0.88% |
§3 explains why: within the proxy top-8, Jev p is ~a monotone inverse of
the momentum rank we handed it (spearman −0.917; +0.51 vs raw 126d
momentum). The proxy artifact's conviction field is the momentum rank, so
Jev is largely reading it back. Its filter therefore replicates the known
within-top8 momentum-concentration effect — information the strategy already
had — rather than adding judgment.
Consequence: the 341-week screen is near-worthless as evidence about Jev. It tested momentum, not Jev. Do not cite it as support.
Supporting findings
- Fragile even taken at face value. Per-year D: 2020 +3.13% (hit 83%), 2023 +1.39%, 2026 +3.36%, but 2021/22/24/25 ≈ 0 or negative. The H21 mean is a 2020 (COVID-recovery) artifact.
- Risk, not edge. JEV Sharpe 2.41 vs proxy 2.21 — the extra ~1pp of return comes with 9.7% vs 7.2% weekly vol; nearly all from concentration.
- Not a risk filter. Vetoed names' H21 mean was +1.25% (positive) — Jev drops relatively weaker names, it does not avoid absolute losers, so it would not protect in a drawdown.
- Regime has no signal. Jev=risk_off on 101 weeks; SPY was up in 63% of them. Neutral weeks 72% up. Jev never emitted risk_on. Reject.
Revised verdict
The proxy screen does not establish a Jev edge; it establishes that Jev echoes the inputs we gave it. Combined with the negative 4-day live replay, Jev currently has no credible positive evidence. Status stays CONDITIONAL CONTINUE but on thinner grounds: keep the shadow purely as data collection. The live Gemma-backed series (real theses/news, conviction not pure momentum rank) is the only remaining test, and its falsifiable kill criterion (D≤0 at 30 days) is unchanged and now more likely to bind.
Lesson for future Jev work
When testing Jev, the state must not encode the answer in a rank-ordered
field (here conviction). Either drop conviction from the state, or judge a
candidate set whose ordering is not already informative, otherwise the test
measures the echo.