Grand Diomande Research · Full HTML Reader

N'Ko Task List — 2026-04-25

- [ ] Monitor `nko_trajectory_ttt_290596` to completion - [ ] Run one-command refresh immediately after TTT completion: - [ ] `python3 Desktop/nko-brain-scanner/scripts/paper4_post_ttt_refresh.py --wait` - [ ] Verify TTT artifacts exist locally: - [ ] `results.json` - [ ] `test_predictions.jsonl` - [ ] `test_references.jsonl` - [ ] Verify refreshed five-run outputs exist: - [ ] `Desktop/Comp-Core/experiments/agp_mlx/asr_bridge/reports/paper4_same_snapshot_batch_replay_final/batch_summary.json` - [ ] `Desktop/Comp-C

Language as Infrastructure technical note experiment writeup candidate score 24 .md

Full Public Reader

N'Ko Task List — 2026-04-25

Active

  • [ ] Monitor `nko_trajectory_ttt_290596` to completion
  • [ ] Run one-command refresh immediately after TTT completion:
  • [ ] `python3 Desktop/nko-brain-scanner/scripts/paper4_post_ttt_refresh.py --wait`
  • [ ] Verify TTT artifacts exist locally:
  • [ ] `results.json`
  • [ ] `test_predictions.jsonl`
  • [ ] `test_references.jsonl`
  • [ ] Verify refreshed five-run outputs exist:
  • [ ] `Desktop/Comp-Core/experiments/agp_mlx/asr_bridge/reports/paper4_same_snapshot_batch_replay_final/batch_summary.json`
  • [ ] `Desktop/Comp-Core/experiments/agp_mlx/asr_bridge/reports/paper4_same_snapshot_batch_replay_final/batch_summary.md`
  • [ ] `Desktop/Comp-Core/experiments/agp_mlx/asr_bridge/reports/paper4_same_snapshot_final_correction/combined_bridge.jsonl`
  • [ ] `Desktop/Comp-Core/experiments/agp_mlx/asr_bridge/reports/paper4_same_snapshot_final_correction/gemma_correction_sft/manifest.json`

Recovery

  • [ ] Mark `gemma4_e2b_nko_correction_partial_real_stage1_stable` step 80 as invalid for quality use
  • [ ] Preserve the step-80 reports as failure evidence, not as deployment artifacts
  • [x] Inspect the clean recovery anchor:
  • [x] `[home-path]`
  • [x] Create a new recovery adapter path that does not overwrite the poisoned stable path
  • [x] `[home-path]`
  • [x] Add recovery tooling:
  • [x] `prepare_partial_real_recovery.sh`
  • [x] `partial_real_recovery_stage1_stable` launcher/chunk modes
  • [ ] Decide recovery mode:
  • [ ] restart from clean probe step 10
  • [ ] or restart fresh from the same partial-real dataset
  • [ ] Add an early 10-row validation gate before any long continuation
  • [ ] Abort immediately on any resumed `loss=nan`

Evaluation

  • [ ] Keep corrected evaluator behavior:
  • [ ] `eval_agp_nko_adapter.py` must parse `ASR candidate:`
  • [ ] Preserve the decisive fixed reports:
  • [ ] `eval-valid-10-base-fixed.json`
  • [ ] `eval-valid-10-fixed.json`
  • [ ] After recovery, rerun bounded base-vs-adapter validation before any broader replay claim

OCR Wave

  • [ ] Let the Mac5 `babamamadidiane` mirror continue in background
  • [ ] After matrix closeout, verify full channel mirror completeness on `HD1`
  • [ ] Start frame extraction
  • [ ] Start scene segmentation
  • [ ] Run OCR on extracted frames
  • [ ] Align OCR text with audio segments
  • [ ] Generate new training/evaluation pairs from OCR+audio alignment

Serving

  • [ ] Keep LearnNKo live inscriptions path stable:
  • [ ] `https://www.learnnko.com/nko?tab=inscriptions`
  • [ ] Keep Mac4 proxy healthy under launchd
  • [ ] Keep Mac5 speech endpoint reachable:
  • [ ] `http://[ip]:8899/transcribe`
  • [ ] Run a better speech sample audit after matrix closeout

Do Not Do

  • [ ] Do not continue training from the poisoned step-80 adapter
  • [ ] Do not use the step-80 adapter for proposal evaluation
  • [ ] Do not start the OCR training wave before matrix closeout

Promotion Decision

Attach run IDs, datasets, metrics, and reproduction commands.

Source Anchor

nko-brain-scanner/docs/handoffs/nko-task-list-2026-04-25.md

Detected Structure

Evaluation · References · Code Anchors