Verified 2026-09-12 against the uploaded private corpus. All 23 traces pass normal replay verification before the corrected ablation sweep prints tables. The corpus contains 57,783 samples, 42 recorded HEAD_UP cues, and 88.1 km of speed-integrated distance (19 traces with motion).
Aggregate evidence records the
compiler, replay binary and source hashes, a filename-free corpus digest, and
regenerated tables. The tested source tree is identical to remote revision
dd8a7db25522734de1dbec614f66113c69d80dcb. Raw exports, GPS coordinates,
ride names/identifiers and per-ride verification logs remain outside Git.
| Variant | Cues | Lead s min/med/max | In 5–20 s window | Top suppressions |
|---|---|---|---|---|
| baseline | 42 | 12 / 15 / 20 | 42/42 | TOO_SLOW 21,196, TOO_EARLY 6,733, INSIDE_EVENT 1,909 |
| no_memory | 42 | 12 / 15 / 20 | 42/42 | TOO_SLOW 21,196, TOO_EARLY 6,733, INSIDE_EVENT 1,909 |
| no_thresholds | 42 | 12 / 15 / 20 | 42/42 | TOO_SLOW 21,196, TOO_EARLY 6,733, INSIDE_EVENT 1,909 |
| every_zone | 83 | 34 / 226 / 1748 | 0/83 | TOO_SLOW 21,196, ALREADY_CUED 7,589, INSIDE_EVENT 1,899 |
| distance_gate | 41 | 6 / 19 / 73 | 24/41 | TOO_SLOW 21,196, TOO_EARLY 6,488, INSIDE_EVENT 1,906 |
| Variant | Changed observation decisions | Changed cue timestamps |
|---|---|---|
| baseline | 0 | 0 |
| no_memory | 16 | 0 |
| no_thresholds | 0 | 0 |
| every_zone | 7114 | 125 |
| distance_gate | 820 | 83 |
Memory removal changes 16 reason codes and no cue timestamps. Threshold and cooldown removal changes no emitted decisions. Removing the notice window as well increases cues from 42 to 83, with none in the calculated 5–20 s window. The distance gate yields 41 cues, 17 outside that window. Its 83 changed cue timestamps count additions/removals/changes at timestamps, not 83 unique cues. These are fixed-trajectory counterfactual replays, not measured rider responses.
Eight Pico sidecars contain 11,301 logged steps: 11,300 compared and one uncompared tail record. Recomputing all four decision fields finds zero mismatches. Stored comparison counters and divergence flags agree; reported orphan count is zero. All eight Pico and all 23 latency sidecars have matching trace filenames. This confirms recorded comparison coverage, not independent firmware identity, physical output, or the cause of the missing tail report.
Watch exports contain 49 dispatches: 27 marked delivered live and 22 queued without live delivery. Among the 27, 18 have a play verdict, four a duplicate verdict, and five predate verdict recording. Exported latency has median 542 ms (range 25–2,890 ms, n=27). These transport/verdict counts do not prove perception. The 49 dispatches and 42 kernel cues are different denominators.
Pico exports mark 15 actuations; 13 include microsecond delay values (138–173 µs, median 147 µs). No GPIO or sound was independently measured here. Rider grades total 33: 17 useful, seven too late, seven unrecognized, two too early, and zero false alarms. These descriptive grades do not establish a controlled rider-benefit effect or a general false-alarm rate.
This closes the missing-corpus replay gap and reproduces the historical aggregate numbers. Together with the host fault campaign, it supports an engineering experience report on execution consistency, comparison coverage and failure boundaries. Shared-kernel agreement cannot prove the shared policy is correct; the fault campaign shows both undetected coverage loss and divergence reported after an actuation request.
The raw corpus remains private. Holders can run the command in fault-injection.md on a new output directory. A public reader can inspect the aggregate evidence and reproduce host faults and synthetic ablations, but cannot independently reproduce these field results without access to the private corpus.