docs(TODO.impl): closeouts - lite2 negative, soup negative, framing pre-registered

TODO.impl/01-speculative-runtime.md

@@ -50,6 +50,13 @@ everywhere, consistent with the 0.036% near-tie flip measurement.
 ## Status
 
 - [x] Probe v1 written; EOS-acceptance and index bugs found and fixed
-- [ ] Probe rerun to completion
-- [ ] Runtime implementation
-- [ ] Tier exposure decision
+- [x] Probe rerun to completion: 25/25 rows, mean acceptance 0.9886
+      (min 0.955), 8.85 tokens/verify at K=8; all 18 exactness-checked
+      rows byte-identical to the verifier's plain-path greedy
+      (results: ~/ml-logs/spec_probe/results.json)
+- [x] Runtime implementation: interscript-ts PR #77 (merged) -
+      SpeculativeModel + acceptBlock exported from interscript/ml;
+      unit tests + tiny-fixture session + real-pair e2e green
+- [ ] Tier exposure: expose SpeculativeModel in the playground/API
+      surface (acceptance threshold cleared; site wiring pending)
+- [x] Numbers recorded in TODO.qwen-next/10 and RESULTS.md

TODO.impl/02-multi-teacher-domain-routing.md

@@ -56,9 +56,17 @@ available teachers can beat plain r7 supervision. The residual is
 student-side (news/wiki rows), not teacher-selection. Remaining
 levers: 04 (trained-lite), 10 (lexical memory), 01 (runtime).
 
+## Soup postscript (2026-09-12): also negative
+
+The r6+r7 50/50 weight soup scored 2.4188 full-set (r6 2.5997,
+r7 2.289) — same-basin (functional model, linear connectivity
+confirmed) but strictly between the parents on every domain slice.
+r7 remains the best available teacher; supervision unchanged.
+
 ## Status
 
 - [x] r7 per-domain slice (from released predictions, zero GPU)
 - [x] r6 preds on volume, fetched, sliced
 - [x] Verdict: negative, recorded in TODO.qwen-next/10 §2
 - [x] No routed launch (nothing to route)
+- [x] Soup probe: negative (RESULTS.md 2026-09-12)

TODO.impl/04-trained-lite.md

@@ -36,8 +36,21 @@ a cost of one run. If it fails, the lite frontier is recipe-bound and
 - [ ] Full-set verdict with intervals recorded in RESULTS.md
 - [ ] Decision: lite tier replaced or negative recorded
 
+## Verdict (2026-09-12): NEGATIVE — 7.1402 vs run-009's 5.78
+
+Same recipe, same canonical labels (sha e70ce991), same teacher
+(re-scores 2.2921 in-run); the single variable - layer-drop init from
+the trained 2.1 instead of generic byt5-small - made the student
+1.36pp WORSE. Reading: generic pretraining keeps layers redundant;
+task adaptation co-specializes them, and deleting half a co-adapted
+stack breaks more computation. The depth axis now reads 2-of-3
+negative (Hebrew generic-init collapse; Arabic adapted-init collapse;
+run-009's generic-init 5.78 stands). The lite tier is unchanged; the
+remaining architecture lever is /10 (lexical memory), still gated.
+
 ## Status
 
-- [ ] Build init on volume
-- [ ] Launch
-- [ ] Verdict
+- [x] Build init on volume (the layer_drop bridge, student_init=2.1)
+- [x] Launch (checkpoint-resumed once; canonical labels seeded after
+      catching the re-labeling hazard)
+- [x] Verdict: negative, recorded in RESULTS.md

TODO.impl/11-static-int8-framing.md

@@ -0,0 +1,46 @@
+# 11 — Static activation scales: fix decode framing at the source (P0)
+
+Pre-registered 2026-09-12, from the framing finding (RESULTS.md:
+dynamic-int8 graphs compute activation quantization scales per fed
+tensor, so single-step and batched decode framings produce materially
+different decodes). The shipped int8 uses `quantize_dynamic`
+(DynamicQuantizeLinear at runtime); static quantization
+(`quantize_static` + calibration) bakes activation scales into the
+graph and removes the per-run computation entirely.
+
+## Why this is the top lever
+
+If framing stabilizes, three separate findings improve at once:
+- the speculative tier's acceptance collapse (0.99 → 0.46 purely by
+  framing) may reverse — its domain would reopen on CPU
+- the cross-hardware near-tie flips (RESULTS 2026-09-07) should shrink:
+  one fewer runtime-computed quantity to diverge across machines
+- the golden test could return to byte-exact assertions for int8
+
+## Experiment (pre-registered gates)
+
+`scripts/static_int8_experiment.py`: re-quantize the 2.1 decoder-kv
+graph with QUInt8 static activations (MatMul-only, head fp32 — same
+node policy as the shipped dynamic int8), calibrated on 365 real decode
+feeds sampled across BOTH framings (prefills, incremental steps,
+8-token windows).
+
+- **A. framing equality**: the same greedy decode driven token-by-token
+  vs in 8-token batched calls — GATE: identical trajectories
+- **B. drift vs fp32** (quality proxy): GATE: no worse than the
+  dynamic int8 artifact's relationship to fp32; full-set DER via the
+  artifact scorer if A clears
+- **C. speed**: static vs dynamic tok/s on CPU — record, not gate
+  (u8s8 static kernels are usually comparable or faster)
+
+If A and B clear: re-export all quantized artifacts through the static
+path (modal_export gains the calibration stage), index-v6, golden rows
+regenerate, framing claim added to the contract.
+
+## Status
+
+- [x] Pre-registered (this file)
+- [x] fp32 artifact fetched, sha-verified against index-v5
+- [x] Calibration corpus collected (365 feeds, both framings)
+- [ ] quantize_static result measured (A/B/C)
+- [ ] Verdict recorded in RESULTS.md; ledger cross-links updated

TODO.impl/README.md

@@ -14,16 +14,17 @@ Priorities:
 
 | # | file | deliverable | priority | status |
 |---|---|---|---|---|
-| 01 | speculative-runtime.md | acceptance probe + TS SpeculativeSession + tier decision | P0 | probe running; runtime pending |
-| 02 | multi-teacher-domain-routing.md | r6/r7 per-domain slice + routed distill verdict | P0 | r6 preds running on Modal |
-| 03 | headwise-muon.md | per-head Muon wire-in for Q/K | P1 | module built, wire-in pending |
-| 04 | trained-lite.md | lite-2.0: distill-trained 6-layer encoder | P1 | spec + launch |
+| 01 | speculative-runtime.md | probe + runtime + framing verdict (tier pulled) | P0 | closed 2026-09-12: framing finding; see 11 |
+| 02 | multi-teacher-domain-routing.md | r6/r7 per-domain slice + routed distill verdict | P0 | closed negative 2026-09-11 |
+| 03 | headwise-muon.md | per-head Muon wire-in for Q/K | P1 | implemented + tested; awaits next teacher run |
+| 04 | trained-lite.md | lite-2.0: init the cut from the trained student | P1 | closed negative 2026-09-12 (7.14 vs 5.78) |
 | 05 | kv-int8-runtime.md | int8 KV cache in the decode path | P2 | spec; IMF contract gate |
-| 06 | sinkhorn-update.md | Sinkhorn-balanced optimizer for embedding/head | P2 | implement + unit spec |
+| 06 | sinkhorn-update.md | Sinkhorn-balanced optimizer for embedding/head | P2 | implemented + tested; run optional |
 | 07 | diacritization-depth-knob.md | depth-conditioned vocalization | P3 | product decision |
-| 08 | golden-closeout.md | golden test PR, paper-c wording, release README note | P0 | fix staged locally |
-| 09 | doc-corrections.md | Engram mechanism fix, mHC single-pass note | P2 | quick edits |
-| 10 | engram-lexical-memory.md | hashed n-gram memory module | P3 | gated behind 02/04 |
+| 08 | golden-closeout.md | golden test PR, paper-c wording, release README note | P0 | closed 2026-09-12 (PR #211 merged) |
+| 09 | doc-corrections.md | Engram mechanism fix, mHC single-pass note | P2 | landed in PR #95 |
+| 10 | engram-lexical-memory.md | hashed n-gram memory module | P3 | gated behind 04 |
+| 11 | static-int8-framing.md | static activation scales: fix framing at the source | P0 | pre-registered; experiment running |
 
 Standing rules that apply to every item: full-set measurement or no
 claim; quantized parity is quality-level, not byte-level; LLM teachers

TODO.qwen-next/10-deepseek-v4-1-flash-learnings.md

@@ -23,18 +23,20 @@ greedy verification the verifier's argmax is authoritative, so outputs
 are byte-identical to non-speculative greedy — no parity story needed,
 just a speed one.
 
-- Step 1 (no training, CPU-only): acceptance probe on golden-v1 rows.
-  Run both models, measure per-position argmax agreement and the
-  distribution of first-divergence offsets. Consonant positions should
-  agree ~100% (copy-through dominates); disagreement concentrates on
-  haraqat near-ties — the same margins that flip under quantization.
-  Net speedup = agreement rate × (drafter cost / verifier cost); the
-  int4 190M vs int8 300M ratio is favorable.
-- Step 2 (if step 1 clears ~80% agreement): K-draft batch-verify in
-  the runtime KV decoder (TS worker first, it has the latency story).
 - Honest scope: this speeds the 2.1 tier. The lite tier stays as the
   standalone fast path; speculative decode is a middle tier that only
   pays when users want 2.1 quality at lower latency.
+- Step 1 RESULT (2026-09-11, complete, 25/25 golden rows):
+  mean acceptance **0.9886** (min 0.955, median 0.991),
+  **8.85 tokens per verifier pass** at K=8 (incl. bonus tokens).
+  Exactness: all 18 rows checked against the verifier's plain-path
+  greedy are byte-identical (spec_loop == verifier greedy, the
+  output-preservation theorem holds on real quantized artifacts);
+  plain==KV on every checked row. Gate (>=0.9 acceptance) cleared
+  decisively. Runtime implementation: interscript-ts PR #77 (merged)
+  - SpeculativeModel exported from interscript/ml, tiny-fixture +
+  real-pair e2e green. Remaining: playground/API tier exposure.
+
 
 ## 2. Multi-teacher domain-routed distillation — cheap untested lever