- [Full-corpus EDA](full-corpus-eda.md) — done 2026-09-24: PDF report + tables in reports/eda_full; onboarding guide PDF in reports/onboarding; never read the Data Story HTML whole
- [Organiser rules](zscore-allowed.md) — lookahead exemptions: filtfilt (same order), removeBadChannel, repairBadSegment, full-trial z-score, each once; external data allowed; FFT prep non-compliant
- [No blocking waits](no-blocking-waits.md) — don't block turns waiting on long tmux runs; launch, say where to watch, return
- [Reply in Vietnamese](reply-vietnamese.md) — always answer in Vietnamese even if the user writes English
- [Decoder LODO findings](decoder-lodo-findings.md) — benchmark state: best = 4 vlaai_ws decoders (env+onset+dom +Fuglsang + SparrKULee pretrain, LDA weights) fused with direct classifiers (5-s and 10-s windows): H 0.778–0.781 (plateau); lookahead 0.5 / DEC_OFFSET=3 default and leak-tested; goal = maximise the competition score; where results live
- [RAM is the limit](ram-is-the-limit.md) — 24 GB RAM, no swap, machine shared with other users: heavy jobs only via the queue (RAM guard 6 GB); how to restart the scheduler and requeue an OOM-killed job
- [Blind test sets and Terms](blind-test-sets.md) — test rules read from Codabench 2026-10-05: only the trial under test, no IDs, within-trial causal adaptation allowed, 2 submissions (latest counts), test phase 9 Nov–8 Dec 2026 → use prep mat_rea (no EA carry)
- [Organiser reply: no data structure](organiser-reply-no-structure.md) — 2026-10-28: no answers to questions; don't tune on data structure; reverse-engineering may disqualify; stem_ms96 (eeg-thao.zip) exploits column order, 0.86 synthetic vs 0.55-0.62 real; new strategy = real-switch focus (REAL_SWITCH_STRATEGY.md)
- [Real-switch findings](real-switch-findings.md) — 2026-10-28..11-07: A/B/C/D idea outcomes and user decisions, foundation models, stage-2 Transformer, BMMS3, stimulus-only control; where results live
- [Design priorities](design-priorities.md) — from 2026-11-07 goal = raw per-subject Acc (kappa metric); improve real switches AND keep synthetic LODO; real-world/hearing-aid orientation; drop exempt preprocessing; params/FLOPs irrelevant
- [Metric changed to kappa](metric-kappa.md) — 2026-11-07: H = 2κL/(κ+L), κ = max(0, 2Acc−1), window ends at next switch; old H not comparable; Acc on real switches is the lever
- [Job queue namespace pitfall](jobqueue-namespace-pitfall.md) — start schedulers with JOBQUEUE_NAMESPACE inline; unset = default queue of other users (incident 2026-11-11)
- [Test submission 1](test-submission-1.md) — 2026-11-11 ref_t_bm_stem20 → H 0.625 / Acc 74.4 / SDT 8.3 s on the blind test; 1 submission left (latest counts); memorisation + loudness follow-ups
