Roadmap

Phase 0 — Knowledge base (this folder) ✅

• agent.md goals
• docs: arch, inference, tracks, AOC, glyph
• maths distilled + findings

Phase 1 — Track A productionize

• [ ] Real tokenizer metrics (tiktoken / HF) in glyph_codec.py
• [ ] Chat export → palette+frame adapter
• [ ] Agent loop demo: one scripted multi-turn task A/B

Phase 2 — Track B MVP (one model, one/few heads)

• [x] Core object in-tree: new_ai_language/aoc/ (NumPy hybrid AOC, harvested Tier 0.5/1)
• [x] r_Σ + kmeans anchors + q-probe locate + local window payload
• [x] Report mem vs A-error vs dense (python -m new_ai_language.aoc.run_mvp)
• [x] r_Σ gate + dense fallback (iso synth → dense)
• [ ] Hook extract K/V or attn_out on small model (needs working torch/HF)
• [ ] Load real raw_kv/*.pt once torch unblocked or export .npz

Phase 3 — E2E quality

• [ ] Needle / short long-context tasks
• [ ] Compare AOC vs dense vs simple eviction
• [ ] Document failure domains (deep layers, high diversity)

Phase 4 — Optional advanced maths

• [ ] Paid novelty / chirotope-local retention only if Phase 2–3 need better online growth
• [ ] Residual recompute baseline (KV-Direct class) comparison

Explicit non-priorities

• Full high-d chirotope enum for d=64, m=32k
• Renaming known methods as novelty
• Vision-as-default for abstract reasoning