New AI Language docs · static / Cloudflare Pages

Finding: AOC MVP wired in-tree (Track B)

Date: 2026-07-15

Status: keep (with open e2e gap)

Track: B — engine only (not text)


What shipped

ItemPath
Hybrid AOC objectnew_ai_language/aoc/hybrid_aoc.py
CLI harnessnew_ai_language/aoc/run_mvp.py
Runsnew_ai_language/aoc/runs/*.json

Implements harvested Tier 0.5/1 without the oracle bloat in harvested_practical_approx.py:

  1. Sample A(q) cloud (key-perturb + isotropic μ proxy)
  2. r_Σ = PCA 95%
  3. Gate → dense if r_Σ ≥ thresh
  4. r* ≈ 4 · r_Σ; kmeans anchors on A
  5. q-probes for locate (inducing qs; harvest surrogate)
  6. key_protos + local_window payload retained
  7. Optional hybrid Z-blend path; primary fidelity = pure anchors

Sources: maths/05_practical_mvp_from_harvest.md, HARVESTED_COHERENT_SPEC.md, math_kv/attention_output_approx.py.


Kill tests (synthetic, 2026-07-15)

Runmr_Σr*compresspure mean L2gate
clustered seed422567283.46×0.163AOC open
clustered seed75127284.83×0.047AOC open
isotropic seed125653dense1.0×0.0 (exact)dense fallback

JSON: aoc/runs/latest.json, synth_m512.json, synth_iso.json.

Pass: mem win when r_Σ low; finite A-error; iso high-r_Σ → dense (no false “always constant”).

Fail / open: e2e logits, needle, real Qwen/GPT2 raw_kv in this environment.


Mechanism notes

  1. q-probe locate is mandatory for low A-error. Key-proto-only locate gave pure mean L2 ~7 on the same clustered synth; probes dropped it to ~0.05–0.16. Matches harvest ApproxOutputO.probes path.
  2. A0 still strong baseline (mean L2 ~1.5 on clustered synth) but pure anchors beat A0 here once locate is correct.
  3. Hybrid Z-blend is secondary; pure full-cover anchors are the Thm3 fidelity metric. Hybrid needs engine-side mass estimates for production.
  4. Backend: NumPy. Local torch blocked (Win Application Control on c10_cuda.dll); real .pt load deferred.

Non-claims

  • Not plugged into HF attention / FlashAttn.
  • Not “constant memory for 7B long-ctx.”
  • Not Track A. Glyph remains separate.

Next kill tests

TestHow
Real head KVExport raw_kv/l0_h0.pt.npz or fix torch; run --raw_kv
Layer sweepSame build on deep-layer heads; expect gate → dense often
E2EHook one small model; needle + PPL vs dense

open_questions

  • Soft-k over anchors vs hard argmax for locate
  • Better Z_f estimator for hybrid without full far K
  • Whether auto-A0 switch threshold generalizes beyond synth