[kingdonb]

Problem and result

Gemini-3.8-Flash's two documents (the Phase 2 roadmap here, and the Unsloth-comparison article in urmanac/nocknock) carried claims above what the recorded measurements support — including kernel-speedup numbers that misquote docs/QWEN36_METAL.md §10, a Phase 2 ticket (T8) for work that shipped in T0, and a T11 premise the 32k experiment refuted. This change delivers the full review and the corrected roadmap, and answers the maintainer's three research questions (why Unsloth prefills faster, what we have that they don't, whether a combined implementation is possible — yes, with the memory math).

Changes

• docs/QWEN36_METAL_REVIEW.md (new) — the referee report: verdict taxonomy; an 18-row claim-by-claim table for the NockNock article (each verdict sourced: §10's recorded numbers, code line references, the model card, the b11724-mix-c208e5f release, fork PRs tests: make iq_parity reproducible from a fresh checkout Niko1221/Strata#264/serve: clear the error string at the start of every request (fixes #183) Niko1221/Strata#194, the maintainer's screenshots); the 32k-evidence section (confound separated from recovery); the mechanism analysis with the deep-decode bandwidth arithmetic (assumptions labeled); the T6–T11 framing audit; the ready-to-run M2 measurement blocks; the evidence requests for Gemini; 11 numbered instructions to the NockNock author (no rewritten article text).
• docs/QWEN36_METAL_RESPONSE.md (new) — our public-voice reply: the honest ledger (where they lead measured / where our hypotheses point / when Unsloth Desktop is simply the right answer), and the tokens-per-joule-at-depth experiment as a testable hypothesis.
• docs/QWEN36_METAL_ROADMAP.md (edit) — Phase 2 rewritten per the review: T12 (re-baseline) and T13 (dispatch instrumentation) added first; T8 demoted to measure/divergence-prefix (the feature shipped in T0); T9 gets the parity-contract resolution and the measured ≥5x bar replacing the "150–200+ tok/s" prose target; T10 keeps the donor-reuse scope with the tests: make iq_parity reproducible from a fresh checkout Niko1221/Strata#264 evidence; T11 reframed from "break the ceiling" to "budget predictably at the ceiling". Explicit execution order: T12 → T13 → T6 A/B → T8 → T9 → T10 → T11.

Validation

Passed: grep audit of every cited path, symbol, line reference and knob name against the tree (push_c_gemv L884, push_c_lpr8 L1023, STRATA_Q36M_ATTN_SK L1915, prefix-matching L2419–2429, per-row comment L1274–1278, 38 kernel void k_ in q36.metal, dense_mm.metal, bench/results/2026-10-18-m2-q36-prefill). External claims re-fetched and verified the day of writing: the release body (20 pins, none touching Metal compute), the model card (geometry incl. 16 Q / 2 KV heads; MTP recipe and its -np > 1 restriction), Niko1221#264 (+16.4–17.8% pp on Qwen3.6, decode tie, CUDA-only), Niko1221#194 (~9–15% solo-decode cost, opt-in).

Not run (deliberate, per the agreed hand-off): the M2 measurement blocks (review §7) — this sandbox is Linux; the blocks run on the M2 via the maintainer/Gemini after this draft is forwarded, and the review marks every not-yet-run number "pending M2 run". Gemini's evidence captures (review §8: the model-load log; the clean 32k rerun) are requested, not collected.

Limits: docs-only change; no engine/backend code touched, so no CUDA/HIP/SYCL builds were required. The NockNock repo is referenced by name but not modified — the corrections go to its author as numbered instructions, per the agreed plan (no upstream PR).

Follow-up

Iteration with Gemini on his article between now and our own PR, per the plan: his evidence captures + our M2 runs fill the pending labels in the next round.

---

🔗 Open in Helix

📋 Spec:

• Requirements
• Design
• Tasks

🚀 Built with Helix