Bat speed as an early signal — results (MISS)

H-BT confirmatory results — the early-signal finding dies here

One run against 2026 data, executed 2026-07-13 under the protocol frozen the same day in bt-preregistration-2026-07-13.md. We publish it because we said we would, and because a pre-registration you only cite when convenient is not a pre-registration.

PRIMARY — MISS

aval_May1-now ~ z(April aval) + z(April ev95)
                + z(April bat_speed)        n = 203

aval_apr    beta +0.107   p = 0.1347
ev95_apr    beta +0.372   p = 0.0019
bat_speed   beta +0.099   p = 0.3812        R2 0.274

Criterion (beta > 0 AND p < 0.05): MISS

The train years said April bat speed adds real signal beyond April results and April EV95 (2024: +0.413, p<0.0001; 2025: +0.231, p=0.015). On 2026 it added nothing distinguishable from noise. The twice-replicated exploratory finding did not survive its one pre-registered test.

Worth saying plainly: the incumbent won. April EV95 — the engine's existing signal — carried the model (+0.372, p=0.002) on the challenger's home turf.

SECONDARY — PASS

Bat speed year-over-year, 2025 -> 2026:
r = +0.944 (n = 151; criterion r >= 0.75)

Bat speed is a remarkably stable physical trait — the most stable metric we have measured, above EV95's 0.884 in our build. Stable is not the same as incrementally predictive.

TERTIARY — null, as pre-declared

2026 aval ~ z(2025 aval) + z(2025 ev95) + z(2025 bat_speed)
bat_speed beta +0.125, p = 0.337 (n = 188)

At annual scale bat speed remains EV95 in a different jersey (same-season r ≈ 0.89).

Post-hoc discussion (labeled as such — not evidence)

Why might the train effect have vanished? Candidates: the 2024 coefficient (+0.413, bat tracking's first public season) may have been the outlier that made two marginal results look like a pattern — a winner's-curse shape; and the 2026 outcome window (May 1 to July 12, ~2.5 months) is shorter and noisier than the train years' full May-onward windows, so power was lower. Either could be true. Neither changes the verdict: the criterion was frozen, and the answer was no.

Execution disclosures

• The confirmatory script crashed after the PRIMARY had executed and printed (a missing full-season 2026 file needed by the SECONDARY). The file was ingested and the run completed; the primary inputs were untouched by the fix and the completed run reproduced the primary numbers exactly as first printed.
• The original freeze commit accidentally included unrelated work-in-progress from a parallel session sharing the git index; it was soft-reset and recut with only the research files (6089203). The protocol content was identical in both cuts.

Verdict and the road not taken

Bat speed does NOT enter the engine. No board adjustment, no product integration — the pre-registered gate said PASS required the primary, and the primary missed.

One follow-up is banked, not promised: re-testing the early-signal hypothesis when 2026 completes (full May-September outcome window, restoring the train design's power) would require a NEW pre-registration filed before that data is examined. If it happens it happens in public, like this one.

Filed next to the clutch null, where honest dead ends live.