Reception phase — corpus manifest (source-faithfulness audits)

Fixed BEFORE the audits ran. Selection rule: every post by (or featuring) the source that the page or its archived sweeps cite, plus the archived verbatim sweeps themselves (which contain surrounding-thread context the Grok agent captured). Known circularity limit, stated plainly: this corpus is the material the page's own research selected; the audits therefore test faithfulness to the cited record, not completeness against the sources' full timelines. A "no contradiction found in this corpus" verdict is bounded by that.

Verbatim context sources (archived in this annex, used by all audits): grok-sweep-margin-claims.md, grok-sweep-glm-gb300-plans.md.

TeorTaxes (@teortaxesTex) — bull position; [analyst] preset + §1 quotes

Scope: direct cost/margin ceiling claims ("at most $4/1Mt", "90 to 95", "80% margins" 2025) vs page-added utilization/fleet/stack parameters (the preset's rentMult 0.85 / util 70 / stack 1.25 are RECONSTRUCTION, labeled). Posts (15): 1895876331495477615, 1895876441117769922, 1895889315970044153, 1896141536884396528, 1901766584974155784, 1906521148726100430, 1926656992082337950, 2068833223228924229, 2068980091724419280, 2070786814097440805, 2070832301005537627, 2070833810145415518, 2071314837771014298, 2071315379004051513, 2071325290819641620 (all x.com/teortaxesTex/status/…)

Zephyr (@zephyr_z9) — semis-side; unit-vs-company coexistence

Scope: the 90–95% unit-margin allegation vs the ~70% company-GM / 15–20% FCF claims — does the page represent BOTH and their coexistence accurately? Posts (10): 2048029866633895944, 2069824935967748247, 2069832997218722140, 2070267916013187449, 2070726279540989971, 2070727375386718259, 2071528823422959737, 2074755502878470238, 2074917201589727588, 2075107995508408728

Dylan Patel / SemiAnalysis — middle position; owned-TCO preset

Scope: separate his DIRECT claim (Opus token margin "north of 80%") from the site-authored owned-TCO parameter reconstruction; audit InferenceX/InferenceMAX usage against the official benchmark material. Source-class caveat (from the plan review): the "north of 80%" citation is a PodcastAlphaX post (2072119494563262697) FEATURING him — quoted secondary, not his primary account. The audit must treat it as such. SemiAnalysis primary posts (7): 2023447723085885576, 2029693660686680529, 2037575752636301499, 2060043702450446552, 2070915303500849289, 2070915304629129460, 2074908975196221654 · plus https://inferencex.semianalysis.com/ and the two SemiAnalysis newsletter deep-dives cited in §9.

fleetingbits (@fleetingbits) — reported-margin camp

Scope: his reported-margin/accounting-perimeter claim ONLY (40–50%, cloud commissions booked under S&M). He does not stand in for every skeptic; the [analyst] Skeptic preset's parameter vector is RECONSTRUCTION, labeled. Posts (1): 2073528885149679622 — a single post; the audit must note the thinness of this base.

@_xjdr — deployment-fidelity audit (not a position audit)

Scope: verify every transcribed §2c number against the archived sweep verbatim: ~12000 api keys, ~300B tokens, 232/431 tok/s/GPU output avg/max, 2.1s TTFT "overage [sic]", 61s p95 TTFT (1M ctx), 81k avg input, 41% cache, 0 chat logs; config 60×B300 "15 trays", bf16 attention + fp8 experts/KV, no MTP/no Eagle; third-party cost estimate attribution (paxaral); separation from SGLang/NVIDIA DeepSeek-V4 and Baseten GLM-5.2 results. Posts (12): 2022517795733541146, 2047587676602425625, 2067741647941832818, 2067742918455279934, 2069038936362803544, 2069917609278218601, 2070947967574663249, 2071677153285210283, 2071835604095300079, 2071846027917943274, 2072058942025601533, 2072059152873210101