← Frontier Inference Margins · all research reports

Research artifact — each page states its own provenance class in its header: verbatim originals carry SHA-256 stamps; adopted-findings summaries and reconstructions say so explicitly. Conclusions are synthesized (and where needed corrected) in the main report; each page carries its own run date.

Final gate review — GPT-5.6 Pro on methodology v2.1, 2026-07-10

Run 32m01s · gpt-5-6-pro (Pro-verified, completion_path: api, Ashitaorbis project) · conversation VERDICT: SHIP-WITH-FIXES — "The model and its central conclusions have crossed the methodological threshold. The current deployment has not crossed the publication-integrity threshold." P0 closure score on the v2 commitments: 6 closed, 4 partial, 2 not closed. No new load-bearing arithmetic errors found in v2.1; all six §10 cards verified faithful to their archived dives. Twelve exact P0 edits specified (xAI lens sequence, remaining confidence-interval and DeepSeek-exclusivity language, provenance/artifact honesty incl. renaming model reviews away from "independent external," tariff-scenario prominence + hero relabeling, the hard incompatibility gate, China/Ant convention placement, §2d/ownership reconciliation, tone completion on loaded quotations, GB200 changelog repair, permalink privacy text, 2027 recast as sensitivity) + 8 runtime smoke checks. Coherence advice: the caveat layers repeat and drift — cut §1 quote theater, §2d, most of §6 narration, the §8 anecdote pile, and per-card parameter inventories (the dossier owns the ledger). All P0 items were applied in v2.1.1 together with the council final-review gate (see the changelog); the full verbatim review is preserved in the linked conversation.