Frozen 2026-10-07 — development moved to https://github.com/DenisSud/arc-agi-3
- Python 88.6%
- Jupyter Notebook 10.3%
- TypeScript 0.8%
- Jinja 0.3%
Manual entry — the auto-journal was skipped because the run was killed mid-train at user request (step 374,000 of 451,170; best eval pred 0.102 @ 349,600; history.npz reconstructed from the block-buffered log, 96% coverage). Records the full decoder A/B on checkpoint_374000: the 98% changed-diff directive investigation through three readout arms — unconditioned 0.646 (reference, decoder_8k/), frame-conditioned 0.675 (CondDecoder), conditioned + changed-pixel CE weight 0.879 (K=25) and 0.904 (K=50, promoted to decoder/). Changed-diff 98% NOT reached by readout-side knobs; the 8k cosine horizon anneals each arm into a plateau, and the residual gap looks latent-limited (pixel-recon-aux fine-tune is the remaining lever). Also records the honest retrieval verification on the fixed retrieval_acc: acc@1 0.0 is REAL (best impostor ties/beats the diagonal in 100% of valid positions, median rank 818), mean diagonal cosine 0.894. |
||
|---|---|---|
| .pi/extensions/repl | ||
| data | ||
| docs | ||
| environment_files | ||
| examples | ||
| just-explore | ||
| su-jepa | ||
| tools_site | ||
| utils | ||
| writing | ||
| .gitignore | ||
| AGENTS.md | ||
| devenv.lock | ||
| devenv.nix | ||
| devenv.yaml | ||
| IDEA | ||
| pyproject.toml | ||
| TODO.md | ||
| uv.lock | ||