AOC package (Track B only)

Attention Output Cover / Manifold KV — engine representation, not text.

What this is

Hybrid cache object per head:

Piece	Role
`anchors`	r\* vectors ≈ cover of A(q)
`key_protos`	locate q → j\*
`local_K/V`	exact recent window
`r_Σ` gate	dense fallback if intrinsic dim high

Run

From monorepo root (v41_run/). NumPy backend (no torch required for synthetic).


python -m new_ai_language.aoc.run_mvp --synthetic --m 256 --d 64

python -m new_ai_language.aoc.run_mvp --raw_kv path/to/kv.npz   # keys K, V

# .pt raw_kv needs a working torch install

python -m new_ai_language.aoc.run_mvp --raw_kv "Research'/maths/exp_data_real_qwen05_256/run_1780840893/raw_kv/l0_h0.pt"

Kill tests

AOC payload < dense KV floats
Compression ≥ 2× when gate open
Finite mean ||A_hat − A|| on held queries
Not yet: e2e needle / PPL (Phase 3)

Sources

docs/06_aoc_engine_representation.md
maths/05_practical_mvp_from_harvest.md
Research'/maths/outputs/HARVESTED_COHERENT_SPEC.md
Research'/maths/outputs/harvested_practical_approx.py
Research'/maths/math_kv/attention_output_approx.py

Non-claims

Still needs model attention hooks to replace real KV cache.
Glyph / STATE-IR is Track A (glyph_codec.py) — separate.
No “constant memory for all 7B long-ctx” without e2e + gate rates.