# Final report — V-REx, T1, T2 (2026-11-07)

## Outcome

No method met the joint target (honest LODO H≥0.85 / Acc≥78, AASD Acc>63,
Carta Acc>60).

## V-REx (closed)

- LODO mean H 0.8652 / Acc 82.10; fuse H 0.8794 / Acc 83.03.
- AASD H 0.6196 / Acc 51.10.
- Carta H 0.5412 / Acc 47.99.
- LODO is SUSPECT under the column-0 stem integrity finding. External results
  regress, so V-REx is closed and forbidden from rerun.

## T1 Step A — latency diagnostic

- Four frozen SW2 seeds; all AASD (18 subjects, 900 trials) and Carta
  (24 subjects, 480 trials); lag grid −16…+16 samples at 32 Hz.
- AASD peak: 0 samples, acc1 0.53977. No nonzero improvement.
- Carta peak: −1 sample (−31.25 ms), acc1 0.60441 versus 0.59908 at zero;
  Δ=0.00532, subject-bootstrap SE=0.00447, below the required 2 SE.
- Conclusion: no fixed conversion-offset evidence. T1 Step B causal lag-bank
  was correctly skipped.

## T2 ESR-stem — full run

Recipe: exactly odd `0.5*(g(x)-g(-x))`, no position embedding, replicate causal
padding, random-start schedule resynthesis, 1500 steps, batch 512, K=8.

- Gates: max `|f(x)+f(-x)|` = 0.0; position agreement = 1.0.
- LODO mean: H 0.7489 / Acc 71.91.
- AVGC: H 0.6998 / Acc 68.55.
- DTU: H 0.7865 / Acc 73.15.
- KUL: H 0.7416 / Acc 72.46.
- UHD: H 0.7406 / Acc 72.22.
- USTC: H 0.7760 / Acc 73.15.
- AASD: H 0.6073 / Acc 50.68.
- Carta: H 0.5429 / Acc 50.35.

Conclusion: ESR is structurally honest but materially below the honest baseline
and external leaders. Close as non-improving.

## T3 CF-Null — diagnostic passed, training failed

- Diagnostic: rho AASD **0.9765**, Carta **0.8652**; both exceed 0.25.
- Honest model: published SW2 classifier, not a stem.
- Seed-0 screen: lambda 0.1, 20 s counterfactual EEG shift, 4x250 steps,
  all five LODO folds and all5 external.
- LODO mean: H **0.7880** / Acc **71.74**.
- AASD: H **0.6367** / Acc **52.35**.
- Carta: H **0.6428** / Acc **58.57**.
- Decision: external Acc regressed and remained far below target; stopped
  before spending seeds 1-3.

## Target status

- Honest LODO H≥0.85 / Acc≥78: not reached.
- AASD Acc>63: not reached.
- Carta Acc>60: existing baseline remains best.

## Round 2 D0 — match–mismatch coupling

- Frozen SW2, both streams pooled against a same-stream 8–30 s past copy;
  attention labels were not read by the gate.
- Five-dataset in-domain subject mean M/M accuracy: **0.5574**.
- AASD: **0.5932**, deficit versus in-domain **−0.0358** (the adaptation
  route required a deficit of at least +0.05).
- Carta: **0.5774**.
- Subject Spearman(M/M, steady attention), diagnostic only: AASD **0.4530**,
  Carta **0.5148**.
- Decision: recording/subject coupling is not weaker on AASD. Skip R2-1
  MM-TTA and R2-2 SAA; proceed only with R2-3 PAD. Together with the 8.9 s
  median dwell and ~56% oracle delay-cut evidence, AASD Acc>63 is likely
  infeasible under the current causal/data constraints unless PAD reveals a
  presentation-specific failure.

## Round 2 R2-3 PAD — gate failed

- KUL dry/dichotic: normal Acc1 **73.59%**, `dom:=0` **69.80%**
  (−3.79 points).
- KUL HRTF: normal Acc1 **75.18%**, `dom:=0` **70.90%**
  (−4.27 points).
- HRTF−dry normal gap was **+1.58 points**, below the +2-point gate;
  dominance ablation hurt both presentations rather than selectively helping
  dichotic trials.
- Report-only external check: AASD **54.71→53.25%** (−1.46 points);
  Carta **59.91→58.65%** (−1.26 points).
- Decision: `proceed=false`; no PAD retraining. Round 2 is closed. The
  combined D0 and PAD evidence supports a task-level AASD ceiling rather than
  a correctable recording-coupling or presentation-dominance shift.

## Artifacts

- `results/research/stem_ms96_vrex.json`
- `experiments/multi_agent_opt/runs/vrex_stem_20261106/metrics.json`
- `results/research/lagbank_diag.json`
- `results/research/lagbank_diag.csv`
- `results/research/lagbank_diag.png`
- `experiments/multi_agent_opt/runs/t1_latency_20261106/metrics.json`
- `results/research/stem_esr.json`
- `experiments/multi_agent_opt/runs/t2_esr_20261106/metrics.json`
- `experiments/multi_agent_opt/runs/t2_esr_20261106/full_run.log`
- `results/research/cfnull_diag.json`
- `experiments/multi_agent_opt/runs/t3_cfnull_diag_20261107/metrics.json`
- `results/research/cfnull_l0p1.json`
- `experiments/multi_agent_opt/runs/t3_cfnull_train_20261107/metrics.json`
- `results/research/mm_diag.json`
- `experiments/multi_agent_opt/runs/r2_d0_mmdiag_20261107/metrics.json`
- `experiments/multi_agent_opt/runs/r2_d0_mmdiag_20261107/run.log`
- `results/research/pad_diag.json`
- `experiments/multi_agent_opt/runs/r2_pad_diag_20261107/metrics.json`
- `experiments/multi_agent_opt/runs/r2_pad_diag_20261107/run.log`
