# Fresh-event replication: final results

Review of the completed September 12 campaign. All 18 stages, four 2,000-update
training runs and 90 evaluation cells completed. Each arm received 12,288,000
transitions; total campaign wall time across serial stages was 11.09 hours.
This note summarizes existing measurements, not a new simulator experiment.

## Frozen primary verdict

The primary condition is timed 1.5 m/s planar velocity increment plus 60 ms
shared delay burst. Counts sum evaluation seeds 9190–9192, 128 aliases each.

| Student training seed | Uniform completion | PLR completion | PLR minus uniform |
|---|---:|---:|---:|
| 9181 | 300/384 (78.13%) | 307/384 (79.95%) | +1.82 percentage points |
| 9182 | 297/384 (77.34%) | 297/384 (77.34%) | 0.00 points |
| Descriptive pooled total | 597/768 (77.73%) | 604/768 (78.65%) | +0.91 points |

The frozen directional screen requires a strictly positive effect in both new
training seeds. **It does not pass**, because seed 9182 ties. The pooled estimate
does not override that rule. A tie does not prove equivalence; these two new
student seeds cannot establish a precise algorithm-level effect.

## Secondary stronger-stress result

Timed 2 m/s + 80 ms exceeds this continuation campaign's severity support.

| Student training seed | Uniform completion | PLR completion | PLR minus uniform |
|---|---:|---:|---:|
| 9181 | 192/384 (50.00%) | 227/384 (59.11%) | +9.11 percentage points |
| 9182 | 194/384 (50.52%) | 211/384 (54.95%) | +4.43 points |
| Descriptive pooled total | 386/768 (50.26%) | 438/768 (57.03%) | +6.77 points |

PLR gains are +18/+12/+5 episodes for seed 9181 and +4/+6/+7 for seed 9182
across the three evaluation seeds. This consistent secondary direction is worth
an independently frozen confirmation. It must not be promoted retrospectively
to the primary endpoint. Historical seed 9111 was slightly worse than fresh
uniform on its own stronger-stress panel (212 versus 213/384); the total program
therefore does not show an all-seed stronger-stress win.

## Retention and interpretation

Both arms pass all six retention criteria on development and every test panel
for both new training seeds: **48/48 checks per method**. These are correlated
criteria, not independent trials. They permit up to a two-percentage-point
completion drop and 10% global/local MPJPE increases, separately under nominal
and full physics with zero actuator delay. All nominal episodes complete.
Some full-physics cells have one failure, within the frozen allowance.
Both methods pass the retention component; PLR still fails the overall combined
directional-plus-retention candidate screen because of the primary tie.

PLR allocates 55.14% and 59.00% of practice episode assignments to the strongest
push, versus approximately 25.1% for uniform. The allocation changes substantially
without replicating the primary gain. High critic error and frequent selection
therefore remain unvalidated estimates of useful practice. Assignment fractions
are not delivered transition or event-dose fractions.

Completion gains do not establish better stress tracking or recovery. For the
stronger panel, mean global MPJPE is 460.88 versus 527.79 mm (uniform/PLR, seed
9181), and 481.95 versus 507.02 mm (seed 9182). PLR has higher global error in
all six stronger-stress cells. These pre-termination aggregates have different
observation lengths when survival differs; they cannot separate longer survival
from worse tracking on matched intervals. Local errors are close on this panel.
Independent event-aligned recovery and failure-aware quality remain necessary.

## Next research decision

Keep fresh uniform and fresh PLR as controls. Do not claim general PLR superiority
or corrected latent-feedback benefit. The measured case for an emergency
retention fix is weaker in these two seeds, although the historical development
retention failure remains real.

The most useful next instrument is event-aligned, pre-reset task measurement:
pre-event competence, event delivery/reach, physical failure versus timeout,
and return to declared tracking bands with dwell and censoring. Report fixed
post-event horizons with survival alongside quality rather than hiding failed
episodes through success-only averages. This can distinguish improved disturbance
tolerance from surviving with persistent displacement.

Then freeze a study asking whether task-quality feedback improves that tradeoff
beyond critic-error prioritization, with fresh event generation and protection
held fixed. Compare dense task feedback before adding causal EMA and corrected
latent features. A cap-only control and another-seed schedule replay isolate
concentration and student-specific adaptation. Promote stronger-stress completion
to a primary endpoint only in a new, independently frozen campaign with new test
realizations; preserve this campaign's primary failure. No estimator/allocator
gate or multi-motion expansion is authorized by these results.

## Evidence checks

Campaign root:
`/home/linjiw/lucid-sonic/experiments/fresh_replication_20260912_a`.
Frozen campaign SHA256 remains
`d70a589dcf07cefb40187ba40cbbb1628bcc2f2cd89b92ea872d36bb0879f714`.
This review verified all 90 metric-file hashes against their receipts and all
eight analysis input evaluation-receipt hashes. Tables were recomputed from
`analysis_s{9181,9182}_test_es{9190,9191,9192}/analysis.json`; development retention
comes from the two `analysis_s*_development_es9130/analysis.json` files.
This is not an independent frame-level replay or a new uncertainty estimate.

Training URLs: [9181 uniform](https://wandb.ai/16726/lucid-sonic/runs/timed-aeeceb79064e951e),
[9181 PLR](https://wandb.ai/16726/lucid-sonic/runs/timed-b10a4a8da01cb23a),
[9182 uniform](https://wandb.ai/16726/lucid-sonic/runs/timed-a084dd3fcc064680),
[9182 PLR](https://wandb.ai/16726/lucid-sonic/runs/timed-afe5615a25b2ff49).
All four training runs were reverified as `finished` through the online W&B API
during this results review.