Roofline throughput-model consultation — GPT-5.6 Pro, 2026-07-10

Run 58m52s · gpt-5-6-pro (Pro-verified, completion_path: api, Ashitaorbis project) · conversation VERDICT: do not replace the anchor-fit model. A compute/HBM/fabric roofline with 1/batch weight amortization, MTP handling and ONE shared residual (η = 31.0% roof-utilization, fit on H800) predicts throughput-oriented operating points well — H20 −11/−16%, GB200 −2% (optimistic FP4 bound), Ascend@50ms −10% — but fails the whole-platform gate at Ascend's 15ms point (−39.8%, ≈6–8pp of margin), and the preregistered experiment is formally unrunnable: p=1 fitted parameter vs n_train=1 platform group; no genuinely untouched platform exists; H800 batch/MTP, GB200 precision bytes and Ascend KV dtype are post-hoc choices. Prefill transfers worse (worst −70.8%). Sensitivity: the favorable result is unstable to the undisclosed H800 batch (worst error 31.5–50.1% across batch 80–128) and to H800-MTP assumptions (η would drop to 22%, Ascend@15ms error −57%). DECISION ADOPTED: anchor fits stay primary; the roofline ships as an experimental diagnostic only (tests/roofline-diagnostic.mjs); a valid future test requires ≥2 training platform groups, a genuinely new held-out platform, frozen checkpoint/batch/MTP/precision metadata, and a max-error-per-held-out-operating-point gate (≈25% throughput ⇒ 5pp margin at 80%).

The full derivation (functional form, per-anchor operating-point assumptions, hand-computed fit and predictions, sensitivity tables, precondition-by-precondition verdict, and the future-test specification) is published verbatim — original bytes recovered from the receiving session's transcript, SHA-256-stamped — at consultation-2026-07-10-roofline-verbatim.html; the essential numbers are summarized above and in the updated LOAO methods note.