← Frontier Inference Margins · all research reports
nvlinkCapMinRatio)Owner voice note aca09d (2026-08-18): the H800's difference from the H100 must be modeled somewhere — he believes it matters mostly for training and less for inference, "but there is no way that it doesn't matter at all"; whichever assumption the model adopts must be explicit and adjustable, so a reader who disagrees in either direction can play with it and see the consequence. A GPT-5.6 Pro pre-decision review (2026-08-18) was folded before shipping: the first design's unconditional multiplier was rejected as double counting and replaced by the matched-counterfactual floor below; the copy moved from "uplift" to "fit-transfer assumption"; the bounds were tightened; H200 got its own lineage value.
The grounded delta (primary sources, captured 2026-08-18): the H800 SXM differs from the H100 SXM in NVLink (400 vs 900 GB/s aggregate bidirectional) and FP64 (1 vs 34 TF) only — FP8/BF16 tensor rate, 80 GB HBM3, 3.35 TB/s and 700 W are identical (NVIDIA H800 datasheet 2631447 Feb-2023 vs the H100 SXM datasheet; Lenovo Press LP1814: "the differences between the H100 and H800 are the NVLink interface speed and the FP64 performance"; Tencent Cloud HCCPNV5; Reuters 2023-03-21). DeepSeek-V3's report (arXiv 2412.19437 §3.2.2/§3.4) states NVLink 160 GB/s vs IB 50 GB/s in their H800 nodes and that the decode all-to-all runs "via direct point-to-point transfers over IB" — a fabric identical on both parts.
Correction to the 2026-08-16 row note. That note said the cap is "modelled nowhere" and that the
engine "has no fabric term on these rows at all". False of the compute path: HW_ROOFLINE carries
400e9 on h800 and 900e9 on h100/h200, and the live roofline consumes both (decode t_N = b·D/fabric,
prefill t_fabric). What is true: at every expert-parallel operating point the page ships the fabric
term is SLACK under the frozen max() form (≈3.0% of the binding memory term on the H800 at the page
default, ≈13% on kimi), so the two rows render identical throughput; and h100/h200 wear the η FITTED on
the capped part (F1 — DeepSeek's production disclosure ran on H800s) — a fit that may include
cap-related effects which cannot be separately identified. The H800 is therefore the measured part; the
borrowing rows are where the assumption lives; and the page's implicit assumption has been "the cap
costs nothing". On the dense tensor-parallel donor the fabric term already BINDS prefill and the H100
renders ≈2.25× the H800's prefill throughput with no lever at all.
What ships. A typed per-row lineage CALIBRATION.<row>.nvlinkCapLineage (closed set:
capped-anchor = h800; uncapped-borrows-capped-fit = h100 — identical constants;
uncapped-family-borrows-capped-fit = h200 — the provenance-distinct family transfer;
not-applicable = the rest, eligibility following provenance, never the nominal fabric); a SPECIFIED
lever nvlinkCapMinRatio ∈ [1.00, 1.25] step 0.01, default 1.00 = the historical model,
byte-identical, and the neutral key is never encoded into a link — the assumed MINIMUM ratio of a
borrowing row's throughput over the SAME row's capped counterfactual (its own constants with the
anchor's 400 GB/s fabric), per phase: T = max(T_raw, r × T_capped-counterfactual). Where the roofline
already renders the uncapped part faster (the fabric term binds — dense-TP prefill at 2.25×) that
advantage counts toward the ratio and the control adds nothing (already-modeled); where the two tie
(every shipped MoE-EP point) the factor is the ratio itself (floor-applied); it never touches the H800
and never rows with their own anchor. Per-leg lineage/disposition lines through one engine-owned
formatter, both phase factors and both roofline ratios printed. A COMPUTED readout
(nvlinkCapReadout) — the fabric-term share of the binding term on both parts, each borrowing row's
roofline ratio over its capped counterfactual, the engine's serial-exposure counterfactual (h100/h800
≈ 1.005× at the page default: t_iter,800 ÷ (t_iter,800 − t_N,800 + t_N,100), a bracket on the
model's overlap formulation, not an empirical bound), and the blend and stand-alone margins at neutral
versus the current setting plus a neutral slope per +0.05 — rendered beside the control and under the
per-accelerator chart. Nothing numeric is authored in a renderer.
No computed value moves at the default: the WIDE 270-state render parity re-mint is purely additive (stripping the one new typed leg field reproduces the prior baseline hash exactly), both tripwires re-derive exact, the default-state permalink bytes are unchanged, and the all-H800 DeepSeek replays are invariant under any value by construction. The sensitivity is real when moved: at ×1.10 the H100 stand-alone margin rises 70.3% → 73.0% and the default blend 68.98% → 69.53%. Evidence task E-2026-08-16-c (a matched capped-vs-uncapped serving observation) stays open — none is public; the setting is a declared belief about a fit-transfer effect, not a measurement.
Owner standing rule, 2026-08-16: "do not give any info about changes from previous versions on the
page, a change log buried where someone looking can find it is plenty, otherwise quite confusing to
users who have no idea what the old systems were." The 2026-08-16 sweep applied that rule across
the page, and then the cards-vintage rebuild later the same day put round-over-round comparison back
into the two estimate expanders — the faces moved to the round-3 readings (owner ruling
q-margins-cards-vintage, 2026-08-16T14:17Z) and the expanders narrated the move. The faces stay
where that ruling put them; the narration moves here, in full, so nothing published is lost:
GPT-5.6 Pro card. The round-2 pair the estimates rebuild was ruled on: 81.8 % effective billings / 84 % at list, bands 64–91 % / 68–92 %. The face now carries the round-3 self-authored revision — 83.1 % at the undiscounted list tariff, span 68–92 %, with a lead-only diagnostic of 79.65–85.89 % across 0–4 months. Its author stated round 3 at list only, which is why that face's basis line differs from the card beside it.
Fable 5 card. The round-2 reading: ≈77 % effective billings over 70–84 %, ≈80 % at list with band 74–86 %. The face now carries the round-3 self-authored reading — central unchanged at ≈77 %, span widened to 65–82 %, ≈80 % at list.
No computed value moves with this entry: it is prose relocation only, and the page's own render parity over all 270 states is byte-identical across it. The expanders keep every provenance line naming which authored statement each face quotes and which contextual review the analysis beneath it came from — that is citation, not version history, and the rule is about the latter.
The version badge moves 2.2 → 3.0 (plan D-10, ruled at M8): since 2.1 this line rebuilt the entire form (M1–M6), repaired the published numbers under adversarial review, added the two-reading FINAL-ANSWER surface, and now ships the last two development lines together. New in this release: the UX-C claim-bearing tails (typed tail commits, relocatable receipts, an honest close path); the M7 citation repairs; the provider-range calculator's derived-fleets statistic as the margin at the EXACT mean mix — the centroid of every distribution summing to 100 % inside the declared ranges, computed in exact integer arithmetic after float cancellation was twice shown to certify wrong answers — with the expected-margin claim rendered only when the serving regime and renormalization structure license it; and fixes for three user-reachable Share-link crashes (minting from modified states and after scope-crossing model switches). Honesty repairs: the FINAL-ANSWER landing-lead sentence now derives from the landing state instead of a hardcoded clause that had gone stale; mix labels speak the bounded providers' own total; and this changelog's earlier affineness claim carries its dated correction below. No computed reading moved: 255/255 production states verified byte-identical against v2.2.0 (research annex: headline-invariance evidence).
Owner ruling q-sliders-fleet-util-point, answering a question about whether fleet occupancy should
become a POINT. It reverses that and generalises it: "if there's a range for providers, then there
should just be an algorithm to sample the max and min based on those ranges … anything that doesn't
sum to 100 is going to be not included in the calculator because you need 100%." Single points, in
his words, "imply a level of confidence that nobody really has."
No declared range is deleted, and one is recovered. GPT Pro's round-2 blend declaration — 50-65 % NVIDIA, 35-50 % TPU, 0-15 % Trainium — was carried verbatim in this page's provenance as PROSE and was not expressible, because until now the engine had no compositional range that preserves a 100 % simplex. Its round-3 self-review said exactly that in its own words, declining to declare blend ranges because "the calculator has no compositional range that preserves a 100 % simplex" and logging the gap rather than faking it. The gap is closed, so the declaration ships where the calculator can read it. No authored number changed; the rendered point is still NVIDIA 60 / TPU 40 / Trainium 0.
A different geometry from the compounded band, deliberately kept apart. The existing band
corner-enumerates INDEPENDENT dials over a box. Shares are coupled by Σ = 100, so their feasible
set is a box sliced by a hyperplane — a polytope whose extremes sit at vertices.
Correction (2026-08-13, M8 exit gate). This entry originally claimed margin is AFFINE in the shares and that measured affineness is what licenses the exactness. The dual consult of 2026-08-10 sharpened that argument and the engine has carried the sharper form since: with the renderable set fixed, renormalization makes cost a ratio of two affine functions — LINEAR-FRACTIONAL, not affine — and a linear-fractional objective over a polytope still attains its extrema at vertices. What licenses exactness is a CONSTANT SURVIVOR-SET REGIME across the envelope, checked at every reported mix; affineness is a reported diagnostic, never the criterion. The affine measurement quoted here (272 pairs, worst deviation 6.8e-13) was real but described the unrenormalized case only.
The median follows his rule. Where the declared shares already sum to 100 they ARE the median — GPT Pro's 60/40/0 do, so the page shows that author's own mix rather than an average it invented. Where they do not, the median is the nearest distribution that does.
What it refuses. Ranges that do not contain the scenario's own mix (GPT Pro's reading of Anthropic's fleet is not used to bracket OpenAI's — 12 of 16 model states refuse, each naming the provider). A provider range and a leg range inside it, which would state one claim twice. And it never rescales a declared range to make it fit.
No computed value moved: 255 of 255 render states byte-identical, 255 of 255 saved links resolve to
the same margin (worstDelta: 0), no DEFAULTS move and no epoch bump.
What. The three pre-public citation-verification debts are resolved as attribution and scope-label repairs only. The Amodei "compute multipliers" term is pinned to the 2023-08-08 Dwarkesh Patel transcript and moves SOURCE ESTIMATE → CONFIRMED-primary for the term's existence and wording; its evidentiary axis is narrowed to a training-side security/secrecy posture, not unpublished inference efficiency, while heise remains the dated relay. The Information item remains CONFIRMED-secondary: its exact AI Agenda metadata is pinned, the paywall was re-tested, and the 13 captures in Wayback's digest-collapsed listing (2026-06-30 → 2026-07-23) were each fetched — none contains the article body — so the "more than half" and "a few hundred GPUs" figures remain relay-only. That finding is bounded to those 13 captures: it establishes nothing about captures outside the listing, or about other archives.
Epoch. The trends-page row is quote-dated to its 2026-08-09 re-fetch and "Updated Feb. 5,
2026" page stamp; the 3.0×/year software-progress statement is now explicitly pre-training,
while the confirmed +37–40%/year hardware line remains. The attributed 10–50×/year
inference-price band could not be re-confirmed at that URL on 2026-08-09. That is a failure to
re-confirm, not evidence Epoch removed it; the inference-price class remains carried by Epoch's
dedicated data insight in §2.3. No computed value moved. Resolution ledger:
research/b9-m7-citation-debts.md.
Two further corrections, folded from the leg's own documentation gate (round 1, HOLD). The OpenAI row said the "few hundred GPUs" figure covered free-tier ChatGPT. It does not: the confirmed segment is the logged-out guest/visitor tier, explicitly outside the free registered tier — a scope this repository's own x-sources record already called "the load-bearing finding", and which the row had been contradicting. Separately, two propagated status notes said "the quantitative figures above remain relay-only", sweeping in numbers that reach the repository by other routes; the relay-only status is now stated for The Information's own two figures and no others — an overclaim in the opposite direction from the usual one, and no more true.
A third round was needed, and the record says why. Gate round 2 confirmed both folds above and then caught two claims the fold itself had introduced: an archive conclusion drawn from two inspected captures while asserting coverage of all thirteen, and a gross-margin figure whose provenance had been silently upgraded to "audited-document reporting" by sitting next to a figure that genuinely carries that attribution. The first was answered by fetching all 13 listed captures rather than softening the sentence (none contains the article body; one lacks the gate string the other twelve carry, and the ledger says so instead of smoothing it); the second by separating the figures and their routes. The resolution ledger keeps all three passes visible, because a citation milestone that hid its own three-attempt convergence would be the wrong kind of document.
What ships. Every estimate preset now carries the RANGES its author declared, rendered as
per-dial bands under the hero at page open, with the compounded whole-box range behind an explicit
opt-in — never the headline. Both round-3 self-authored presets land (GPT-5.6 Pro and Fable 5, each
authored through the calculator itself), and the page-open default moves from the round-2 contextual
Pro preset to the round-3 self-authored Pro preset (owner-accepted 2026-08-08): the first number
a visitor sees is now ≈83% at the author's own +2-month lead, beside the author's own stated
≈83.1% (68–92%) and a derived gap line reading 0 points. The author's six identical NVIDIA leg
ranges are carried as one family-scoped range (values unchanged), so the landing default's
compounded band computes — 10 bounded assumptions, point 83.0561 unmoved to four decimals.
No saved link moves (measured: 180/180 minted-on-master tokens resolve byte-identically;
DEFAULTS untouched; the encoder epoch advances to v23r499 with this release, its first serving).
This release also puts the corrected +2.4-months stack-multiplier equivalence on the served page —
the v2.2.0 hotfix below was published to the mirror on 2026-08-07 but not deployed until now.
What. The hero tile's tail — every state-identity label, supplemental receipt and trailing
affordance under the headline margin — is now composed as a TYPED three-region structure
(.tile-mandatory / .tile-receipts / .tile-actions) committed atomically per render, in
place of one accumulated string. Receipts (and only receipts) can relocate into a "Result
receipts" popup on the shared dialog coordinator; on coarse pointers they are replaced by that
popup once JS readiness is proven, and restored for print. The incompatible-pair early return
now clears the analyst-share disclosure and the feasibility pair instead of leaving a previous
state's receipts under an "n/a" tile — the leg's one behavioral delta. The dialog close path
now REPORTS restoration failure instead of declaring success: a failed restore keeps the modal
(with the complete relocation plan registered for retry), every mutation that could destroy a
relocated source aborts while it stands, and pending work replays through one level-aware
dispatcher once the close succeeds. .id-epistemic (the anti-conclusion-shopping label) is
promoted in place. No computed value moved; the tail's fresh-state text is byte-identical
across a 21-state pre/post fixture; sinks 390→389 with zero class moves; index.html untouched.
Spec: research/b9-ux-memo.md §18 v6 (six adversarial design-gate rounds, P0 funnel
7→5→3→2→1→0). Manifest: research/b9-delta-manifests/b9-ux-c-delta-manifest.md. Implementation
047428e.
A figure this page stated about itself was wrong, and it was wrong on this page's own formula.
The serving-stack efficiency tooltip, its 1.25× tick label, the §"interlock" bullet in the report and
the v2.2.0 changelog entry below all said that a 1.25× stack multiplier is "≈ +2.9 months at 3×/yr"
of algorithmic lead. It is not. On the trend model this calculator actually implements —
months = 12 · ln(E) / ln(rate) — a 1.25× factor is +2.4374 months, which the page now states as
≈ +2.4 months. The +2.9 figure is the months-equivalent of 1.30×, not 1.25×: the two were
transposed when the copy was written, and nothing ever checked the sentence against the function
sitting a few hundred lines away from it.
Nothing computed changes. This was never an input to any number — no margin, cost, span or scenario reading moves, and no saved link is affected. What moves is a claim the page made about the size of an assumption it had retired, in the sentence explaining why it retired it.
What changed structurally, so this class of error is caught next time. The months-equivalent is now derived from the engine's own trend function inside the release test chain and asserted against the shipped copy, rather than being a number someone typed into a tooltip. A test did pin that tick label before today — and pinned the typo with it, asserting the literal string as though a hard-coded number could verify itself. A future edit that moves the trend model, the default rate, or the sentence will now fail the chain instead of shipping.
Found by the row-499 algorithmic-lead investigation while re-deriving every months-equivalent this project quotes.
The headline number changed, and this entry exists to say so plainly. Until today this page's
landing default — the number a first-time visitor sees before touching a control, opus at the
median perspective — read ≈77%. It now reads ≈69%, with ≈59% carried beside it as a
second, labeled public-evidence reference reading. Both are scenario readings of this calculator.
Neither is a measurement of anyone's margin, and neither has become "more correct" than the other:
they differ by exactly one declared assumption, the owner-ratified algorithmic-lead prior, and by
nothing else.
Why it moved — three mechanisms, not one, and they run in opposite directions. Each delta below was measured by executing that revision's own engine at the landing default, not inferred:
| # | what landed | when | landing default | move |
|---|---|---|---|---|
| 0 | the previously published engine (v2.1.x) | — | 76.7784% | — |
| 1 | serving-model re-engineering — the activated roofline display path, the parallel capacity-width solver, and serve-feasibility-filtered default fleet membership | 2026-07-19 → 07-23 | 35.1404% | −41.64 |
| 2 | the model-size revision — the current flagship adopting the 2–3 T community band at a 2.5 T scalar, in place of the earlier 5 T Musk-relative reading | 2026-07-24 | 37.2066% | +2.07 |
| 3 | the margin-evidence adjudication — the r4 run B §C1 repaired defaults | 2026-07-26 | 59.1806% | +21.97 |
| 4 | the owner-ratified algorithmic-lead prior (b9 M5) — the one assumption separating the two published readings | 2026-07-28 | 68.9840% | +9.80 |
The rebuild of how this page models serving is what moved the number most, and it moved it down; the two adjudications that followed moved it back up. The FA arc's own acceptance rows asked only for the size revision and the margin-evidence adjudication to be separated (G1A-1, G1A-3) — the measurement found a third and larger mechanism, so the on-page bridge names three. That bridge is the fifth block of the FINAL-ANSWER surface and names the ≈77% predecessor directly.
Old links keep telling the truth, in two different ways — both verified against the deployed page, not assumed. A share link in the pre-v5 format (everything the ≈77% page itself minted) no longer resolves under this engine at all: it raises the epoch-deprecation notice — "This shared link predates the v2.2 engine (v4 format) and can no longer be resolved. Showing the current central scenario instead. Your data was not changed." — rather than re-interpreting an old relative token against moved tables. A v5-or-later link, which carries the margin its sharer saw, restores and renders the drift notice on the whole-percent hero convention: "originally shared: ≈77% — current engine: ≈69%". Neither path silently re-renders an old link at the new number.
What else this release makes public for the first time (the whole b9 arc, previously development- only): the b9 M6 two-reading FINAL-ANSWER surface and its analyst-gap executive summary; the b9 M1 repaired defaults; the b9 M2 form-correction debt disclosure; the b9 M3 energy/procurement-basis dimension; the b9 M4 custom fleet builder and permalink codec v6; the b9 M5 family sliders and algorithmic-lead prior; the user-tunable speculative-decoding lever (default 1.00×, no credit); the IM3 roofline display path and the IM4 fleet/capacity solver behind mechanism 1 above; and the UX-A and UX-B legs. Several §10 provider cards move with them — most loudly Zhipu/GLM's dive replay to ≈−166% (policy-unclean, labeled as such on the card) and Google/Gemini's to ~84%.
Carried forward unchanged: v2.1.12's two owner-approved RAISE Summit podcast claims and their
analyst-characterization provenance tier. Version identity: ENGINE_REVISION, the footer
release manifest and package.json all read v2.2.0; DATA_AS_OF stays at 2026-07-26, the date of
the newest evidence incorporated.
Why. The 2026-07-27 owner ruling on q-im-podcast-candidates approved two claims for inclusion, both from Dylan Patel's (SemiAnalysis) on-stage remarks at the RAISE Summit 2026 (Paris; recorded 2026-07-09, published 2026-07-16) — surfaced by this project's podcast-mining pipeline on 2026-07-26. This is the first release to carry a podcast-transcript-sourced claim in the MARGIN_CLAIMS registry.
New provenance label: ANALYST-CHARACTERIZATION. Both claims are Patel's own account of pre-IPO/private financials he says he has reviewed, not a company disclosure or filing — the owner ruling required this be carried explicitly, so a new provenanceTier: "analyst-characterization" value is introduced (free-text field, no enum to extend) alongside a tierNote on each record spelling out the caveat.
Anthropic — first realized-profit claim. Patel said Anthropic turned its first gross profit in Q2 (June 2026) and will book slightly over $1B of operating profit in Q3, describing these as the financials Anthropic is preparing to disclose in its IPO. Added as a new patel-anthropic-first-gp-2626 MARGIN_CLAIMS record (binnable: false — a profit-milestone/dollar claim, not a percentage margin, so it cannot sit on the margin-% axis; carries its own reason per the registry's binnable:false convention) and as a dated paragraph in §7's reported-margin narrative, distinguishing it from every other figure in that section (all projections or modeled estimates, never a claimed realized profit).
OpenAI — a later, differently-cut margin trajectory. Patel said OpenAI's company-wide gross margin moved from ~30% (late 2025) to ~55% today, and — stripping away free users — from ~50% to ~65%. Added as a new patel-openai-margin-trajectory-2626 MARGIN_CLAIMS record (binnable: true, point at the current 55% company-wide figure, relation: "different-metric", matching the existing openai-33-company-gm record's pattern) and as a dated paragraph in the §10 OpenAI card, placed alongside the existing 33%/39% Information/Reuters figures it updates without replacing.
No preset, parameter, MFU, price, tariff or architecture number changed — both additions are evidence-board/narrative only, following the same pattern as the DeepSeek TI addition in v2.1.6. ENGINE_REVISION and DATA_AS_OF (now 2026-07-26, the date of the newest evidence incorporated) bumped accordingly; the footer release manifest and package.json version aligned.
What. The calculator gains a specDec scenario lever: a post-roofline, decode-only
throughput credit, default 1.00× (no credit), domain [1.00, 1.60]. It is available ONLY from the
"no MTP/disagg" stack setting (0.7) — a mutual-exclusion gate, because the stack slider's own
1.0 baseline is defined partly by the same MTP evidence the lever would credit, and crediting both
would count one improvement twice. It is inert under every published-operating-point replay, and it
is structurally exempt on any leg whose deployed efficiency already absorbs speculation
(h20, ascend) or whose speculative status this page cannot establish (gb300, tpu7) — so 37% of the
default fleet by weight never receives the credit, and every leg says on itself what it received.
The page's own reading does not move. The default is 1.00× and x * 1.0 === x exactly, so the
gate-shut path is bit-identical to the pre-leg engine: reference blend 0.5918058739356502,
ratified default 68.98395363429421, membership 7/0/100, and a 10-leg × 7-scope identity fixture
minted before the lever existed reproducing byte-for-byte afterwards. The exec-summary row stays
computed: false. Turning the lever up declares your scenario, not this page's finding.
Design first, then code, then a court for the bytes. The design was pre-registered as
research/b9-spec-decode-lever-memo.md and gated through thirteen rounds / 26 lens-runs before
any implementation existed, then frozen by adjudication once findings had been off-mechanism for
eight consecutive rounds — the stopping rule being that a review loop converges only when the
reviewed artifact is shrinking or the defect class is closed under review, and a whole memo under
adversarial review is neither. Every byte of reader-facing copy this leg ships or changes was
ratified by the court, not authored on the leg's own authority — 26 enumerated spans, each
pinned as the RUNTIME STRING a reader receives rather than as source bytes, because these constants
are built by concatenation and a source-level grep proving absence proves nothing.
Four live defects in already-shipped copy were found and corrected on the way. The most instructive: the page claimed acceptance figures for speculative decoding are unpublished "for any lab", which is false against its own registry — two non-flagship anchors carry them. The true claim is a frontier-fleet absence, and it is now what the page says, in both files that made it.
Full record: research/b9-delta-manifests/b9-specdec-delta-manifest.md.
What. The FINAL-ANSWER surface stops shipping one pinned reading and ships the five-part D-6
surface: two labeled calculator readings — the public-evidence repaired reference (algorithmic
lead 0 months, family multipliers 1.0×) and the calculator's own ratified-prior default — plus the
adopted analyst reading (above 80%, unchanged), the no-verified-central-comparator statement
(unchanged), and a bridge saying how they relate. Beside it, a collapsible analyst-gap executive
summary shows engine-computed single-control bridge rows from the default state, ordered by the
owner's declared plausibility order and never by margin. Every explanatory surface the milestone
reworks adopts the §20 owner rulings: legible explanation typography on the default skin
(.explain-body), a hyperlink-styled "Deeper explanation" trigger opening a near-fullscreen,
non-grayed, internally scrollable popup with a top-right X, and nothing expanded by default.
Design first, then code — with one disclosed exception. The design was pre-registered as
research/b9-m6-fa-memo.md and gated through ten review rounds (fresh GPT-5.6 Sol xhigh each, analysis-only) before any
implementation existed — findings 9 · 8 · 10 · 7 · 3 · 3 · 1 · 1 · 1 · 0-blocking. Three of the
first round's four blockers were false claims in the memo's own evidence, one later round caught
an assertion that could not fail, and one caught a governance question a design gate had no
authority to settle. The memo records what was wrong each time, not just the conclusions. The pre-registration claim is
made only where the commit history can back it: the memo's §§0–16 exist in commits that touch no
code at all. A later section (§17), closing five holes the post-gate owner amendment left open,
ships in the implementation commit itself — so its ordering rests on the session's account rather
than on the history, and the memo says so. One of those five is a copy amendment that the
implementation gate holds should have been adjudicated first; it ships flagged as a declared,
still-unadjudicated deviation rather than smoothed over.
No number moved. M6 changes which readings are LABELED and SHOWN, and how explanations are
rendered — not one computed value. The reference blend (0.5918058739356502), the ratified default
(0.6898395363429421), every fixture, every snapshot and the WIDE 180-state render hash are all
byte-identical across the milestone. The FA keeps its reference-pinned derivation and GAINS a
sibling prior reading; un-pinning it would have moved FA numeric fields.
Speculative decoding — a lever, deliberately out of the default. Following an owner ruling of 2026-07-30, the exec summary's spec-decode row is not a bare refusal: the lever exists, it is deliberately not in the default scenario, and the page says so and names what would be needed to price it. The mechanism is now vendor-officially on the record and dated (an OpenAI engineering post of 2026-07-29 credits an improved draft/speculator model with more than 15% additional token-generation efficiency; its 2026-07-30 pricing post says it is passing those gains on) — at ONE lab that is not this page's flagship, with acceptance rates and draft-model details published nowhere. That vendor claim and the independent open-stack measurements (about 14% at production-like batch to about 60% at modest concurrency) are reported as separate classes and never summed. The credit stays at zero here, stated as a conservative floor rather than as a finding that it is zero. Building the control itself is scoped as its own gated piece of work.
Superseded readings (history). Before the b9 repair this page's evidence-informed blend read ≈37% and its declared-topology blend ≈48%, on defaults the r4 adversarial review found defective (per-device batch/weight identity, b = 4 Trainium operating points, cross-platform throughput coefficients, no explicit MTP, several low or analyst-set rents). Those readings are superseded by the repair and are preserved in git history. They are not selectable, and this page does not present them as a coherent public-market-rent case or as a central estimate of any provider's economics. This milestone moved the last live pre-repair comparison off the FINAL-ANSWER surface and into this note and the methods box, which is where plan D-1 says it belongs.
Legacy retirement, honestly reported. The enumeration D-1 required was executed rather than asserted: all 12 perspectives, all 16 model presets, the three counterfactual fleets and the size bookmarks were inventoried at both bases. Nothing needed retiring — M1's data repair had already retired the broken readings STRUCTURALLY: no live perspective, preset or exploration route resolves through a retired operating-point identity. That, and not numeric absence, is the property the plan actually needs; numeric absence would be the wrong claim in both directions, since an ordinary slider state can still round to a retired headline (several do) while a retired route could be reachable showing some other number. So what M6 owed was placement, basis-labeling and one guard-scope gap, not a deletion pass. An enumeration that quietly found nothing and said nothing would look identical to one that was never run.
What. The mid-tier and big-picture levers (plan D-5/D-9.1; shared memo
research/b9-m45-ui-memo.md v2.1 §§8–13, design gate closed at 28d4b25). Two BROAD-UNSPECIFIED
lever groups now exist. Family efficiency multipliers — one ±% multiplier per hardware family
(NVIDIA / TPU / Trainium / Ascend, 0.50–1.50×) applying to every card in that family, in every
serving phase, in any fleet; nothing here is sourced, and a custom leg tagged unclassified is
exempt and says so. Algorithmic lead (months vs published open practice) — how far a lab's
private serving stack is ahead of published open practice in RESIDUAL, non-hardware,
accelerator-portable efficiency this page does not otherwise model, converting through
E = rate^(months/12) at a ratified 3×/yr (halving ≈7.57 months; Gundlach et al.
arXiv:2511.23455 with Epoch AI data) with modeled cost-out divided by E in every phase. Its
per-lab defaults are an owner-ratified scenario prior, not a measurement: Anthropic /
OpenAI / Google +3 months, DeepSeek +1, the other frontier Chinese labs 0 (their published open
work DEFINES the zero), any lab with no ratified prior 0 and labeled unassessed. Under a
published-operating-point replay the slider is inert and locked at 0 — the lab's actual
efficiency is already inside the replay, so months on top would double-count.
The interlock. Both groups describe the same kind of unmodeled efficiency, so composing them silently would count one improvement twice. Editing either group locks the other with the reason stated inline; because the ratified defaults start nonzero, the FIRST family edit replaces the prior — the lead is set to 0 and locked, loudly, in one attributed line — so every headline move stays attributable. Unlocking is always available, always deliberate, and leaves a persistent stacking banner. Explicitly-modeled levers (precision, serving regime, the cache controls) are NEVER locked by this machinery; where they overlap a live prior you get a non-blocking warning. In the same spirit, serving-stack efficiency was redefined: it now means measured stack COMPOSITION relative to published open practice, its 1.0 tick naming the same referent as the lead slider's zero, and the "frontier lab (assumed)" label is retired — a frontier assumption is exactly what the lead slider now represents defensibly (1.25× ≈ +2.4 months at 3×/yr — figure corrected 2026-08-07; this entry and the page's own tooltip both said +2.9, which is the months-equivalent of 1.30×, not 1.25×. The original wording is preserved apart from the number itself, and the correction has its own entry above). Slider VALUES are unchanged everywhere, including the §10 dive replays.
Also. A touch scroll-lock (edit / keep static this visit / lock + don't ask), deliberately distinct in icon, copy and placement from the interlock's lock. Permalinks carry the machine state explicitly (UNLOCKED is a user CHOICE, not derivable from values) with baseline-aware fail-closed validation; a pre-machine link gets an honest reconstruction, never an assumed FREE. Refused permanently and stated in the methods box: the published token-PRICE series (9×–900×/yr) measure tariffs, not serving efficiency, and capability lag (~4 months, Epoch-measured) is a different axis that is never this slider.
Numbers. This is the milestone whose acceptance is that the DEFAULT headline moves: the
flagship default goes 59.18% → 68.98%, inside the plan's ratified-prior sanity band ≈66–71,
and the move is exactly the reference with cost divided by E(+3 @ 3×/yr) = 1.31607. Everything
else is byte-unchanged and proven so: the trend-0 reference blend holds at 0.5918058739356502,
every numeric field of the FINAL-ANSWER surface is byte-equal to its pre-M5 value under an
interim reference pin (the FA states, on its own surface, that it reads at 0 months and that the
ratified-prior reading arrives with the final-answer rework), and feasibility, declared batch,
capacity and the status vectors are lever-invariant. Full old→new ledger:
research/b9-delta-manifests/b9-m5-delta-manifest.md.
Not shipped. No push, no deploy, no master commit — production stays v2.1.11.
What. The owner's max-granularity path (plan D-4/D-9.2; shared memo
research/b9-m45-ui-memo.md v2.1, design gate closed in three Sol rounds at 28d4b25): a
custom fleet BUILDER — the only composition-changing path — behind a "＋ Custom fleet…" entry
under the fleet switcher. Fleets are user compositions of 1–12 legs; each leg names a
CALIBRATION DONOR from the registered accelerators (the page's own LOAO test refuses MFU
transfer across platforms, so performance identity is always the donor's, disclosed per leg;
peak-rate fields render read-only in the drill-down). Editable per leg: rent $/hr, capex,
operating power, per-datacenter electricity $/kWh (live only under the owned/strategic-TCO
basis — inert-and-explained under rent, D-2), user-declared HBM capacity (BINDS feasibility
through the solver's one registry point), share %, label, family (donor's or unclassified —
exempt from the future family sliders, disclosed). Duplicates of one accelerator at different
electricity costs are first-class. Custom fleets are engine class user-custom: never a
default, never central-eligible, serve-feasibility fail-closed with the same welded disclosures
as named fleets; the disclosure carries a typed profile whose evidence quantities are
null-with-reason (user compositions claim no evidence). Permalinks move to codec v6:
custom fleets travel BY VALUE (self-contained links, fail-closed 4-level schema validation,
explicit size-guard refusal instead of truncation); v5 links keep loading under the established
drift-warning path — nothing is deprecated. Link-restored fleets are EPHEMERAL (loading a link
never writes device storage; "Save a copy" is the only write path). Persistence:
im_custom_fleets_v1, fail-closed and non-destructive, epoch-stamped.
No shipped number moved: reference blend margin byte-identical 0.5918058739356502; zero
fixture regeneration; the leg-path refactor is equivalence-proven for every existing state.
New suites: tests/custom-fleets-b9.test.mjs (75) + tests/custom-fleets-cdp.test.mjs
(15, real-browser builder round-trip); browser chain 198 → 225. Delta manifest:
research/b9-delta-manifests/b9-m4-delta-manifest.md. Production stays v2.1.11 — no deploy.
What. The calculator gains a physical energy surface and typed procurement bases (plan D-2/D-3;
memo research/b9-m3-energy-memo.md). Serving energy (Wh per M tokens) is computed for every leg
and every lens — operating power × PUE ÷ achieved throughput at the engine's own operating point,
with three declared conventions: TDP-as-operating-power proxy (typed boardPowerW override hook,
null everywhere — no per-accelerator operating-power measurement is public), NO utilization divisor
(the Wh figure is physical intensity; the dollar path's idle allocation is a different question,
stated), and the analyst-set 5% cache-read fraction doing disclosed proxy duty for cache-read
energy. The exact identity electricity $/Mtok = Wh/Mtok × $/kWh ÷ util welds the new surface to
the cost engine (suite-enforced to 1e-12). Every cost lens and every rent cell is typed to one of
three named procurement bases — public-capacity rent (the China public-cloud lens),
low/committed planning rent (the registered planning vector, this page's default), and
owned/strategic TCO — with a fail-closed mixing trap on every computed mix (load-bearing for
M4's per-leg overrides). Electricity DOLLARS stay explicit only under owned/strategic TCO; rent
lenses carry an embedded-in-rent chip plus the implied Wh/Mtok (info-only), because decomposing
rent without TCO assumptions would be fabrication. Registry rows additionally gain family
(consumed by M5's family sliders). No shipped number moved — the flagship reference margin is
byte-identical 0.5918058739356502, no fixture or snapshot regenerated, rent lenses byte-invariant
under kwh/pue sweeps (new suite tests/energy-model-b9.test.mjs, 35 assertions after the gate
fix round; browser suite 184→198 both copies, counting the full test:browser chain — run-app-tests.sh alone: 148→162; convention stated in the delta manifest). Full enumeration: research/b9-delta-manifests/b9-m3-delta-manifest.md.
What. The untyped per-device b/W_iter decode identity is replaced by run B §C4's
topology-aware field set, delivered as TYPES with no implementer defaults: a REQUIRED
decodeTrafficBasis (3-member closed set incl. expert-coverage), etaRepresentation bound to
the basis (η may not be read outside the representation it was calibrated in), a REQUIRED
batchQuantity discriminator on every operating-point cell (B_rep DERIVED as b × N_phys),
nPhysDeclared as a registry constant firewalled from the capacity solver, declared q/a
(all 1/1) + anchorAlreadyIncludesMTP derived per anchor with a no-universal-multiplier guard,
per-platform prefill hooks (every row still universal-transfer), and resolveDecodePlacement
with opus as a labeled declared-surrogate. No shipped number moved — the reference blend
stays 59.1806% and every leg's throughput is byte-identical; six of seven legs keep the
surrogate's numbers because the corrected ones are not identified by public evidence
(memo §2.2–2.3), and every remaining D1 instance is typed, declared, disclosed and SIZED.
The disclosure surface (family 9). Every calculator result carries an OPEN CALIBRATION DEBT
aside: the engine-computed flagship span (53.24–64.17%, 10.93 pp — how far the §C4 form swings on
the N_phys declaration alone, both ends outside the sanity tripwire), per-leg η-held
re-expression sizes, the Polaris-adjudicated trn2/trn3 "may be ~15× wrong" cross-quantity
exposure (run B §A3-vs-§C1 self-contradiction, stated and typed rather than resolved without
evidence), and the §C4 replication residual on topology-aware legs. Same surface on MCP
run_scenario.
Amendment 3 fold (2026-07-28). The external adversarial review corrected the coverage
exponent to per-token distinct selection — 1−(1−k/E)^{B_rep·q} — superseding the
with-replacement form the design memo carried; memo revised to v6, the identified span became
engine-computed (a pinned literal would have stayed stale exactly as the v5 figures did), and
probe8 re-derives every coverage-dependent figure both ways.
Also: tests/roofline-core.test.mjs repaired (11 FAILs + a crash → 478 PASS) and promoted to
GATING; new tests/form-equivalence-b9.test.mjs (155 assertions); family-9c browser assertions
in both suite copies. Full enumeration: research/b9-delta-manifests/b9-m2-delta-manifest.md.
Production continues to serve v2.1.11; nothing here is deployed.
Why. An owner-commissioned granular adversarial review of the calculator's input defaults
(research/reviews/im-adv-r4-REPORT.md; three independent legs — an Exa academic sweep and two
GPT Pro dives, blind-public and internal-registries) found that the ≈37% headline was manufactured
by a small set of identifiable defects that all pushed the same direction, concentrated in the
50% of fleet weight the engine's own eligibility harness already marked as failing both its
throughput and price evidence gates. The reductio: the two Trainium legs computed −43.6% and
−21.5% margin, costing output tokens at $60 and $52 per million against a $25 list price —
2.4× list-price loss on co-designed silicon that Anthropic demonstrably serves Claude on
commercially. Nobody does that. This release lands the review's exact repaired cell set.
The reference blend moves 37.2066% → 59.1806%, inside the review's repaired band 55.24–61.25 at its midpoint 59.20. The Trainium legs land at +34.8% and +65.1%. This is not a ship: production continues to serve v2.1.11, and nothing here is deployed.
The defects repaired (each cell carries the review's own basis; full enumeration in
research/b9-delta-manifests/b9-m1-delta-manifest.md):
measured: null and an assumed batch. Relabelled analyst-set at a declared assumed operating
point; its assumed batch 128 becomes a declared workload midpoint of 64.Two assumptions that were implicit are now stated. The review recovered both by inverting the published cost rows rather than reading them off the page: a cache-read serving cost of 5% of fresh prefill (analyst-set, unobserved, and distinct from the published 10% cache-read billing tariff — it puts ~70.1% of modeled direct cost on the prefill side at the current flagship default), and a KV/recurrent-state charge of ≈0.545 GB per live sequence. Both now appear in the methods box and the tip registry with their derivations. The review's scenario-only ledger — which cells this page can defend as a scenario and cannot defend as a measurement — renders alongside them.
A size-invariance property was lost, correctly. The headline used to be identical at 2.0, 2.5 and 3.0 T total parameters. The review diagnoses that as evidence the equation omitted total/resident geometry, not evidence of robustness. It now varies monotonically, and the three values reproduce the review's own independently computed table to the printed digit.
Two page-authored "80–90%" routes now land inside the band they were authored for (≈81.7 and ≈80.7, previously ≈79.6 and ≈78.5 and disclosed as falling short). The repair moved them; no route definition changed. The owned-TCO route still falls short of its ≥90 band at ≈89.1, and says so.
Verification. Four independent reproductions of the review's arithmetic, none curve-fitted: its
per-leg table (≤0.08pp), its repaired band (55.22–61.23 vs 55.24–61.25), its TPU quantification
(16.90 → 7.31/7.63/7.95), and its total-size sweep (59.87/59.18/58.49 vs 59.9/59.2/58.5). Full node
suite, contract harness, migration differential, MCP suite 39/39, worker build and the 496-point
fleet-policy harness all green; frozen protocol paths byte-untouched; site/app.js byte-identical.
The browser suites could not run in the authoring environment (headless Chrome unavailable) and are
owed at the gate.
Why. A fresh cold public-only GPT-5.6 Pro review of the v2.1.10 site (research/gptpro-reports/cold-review-v2110-2026-07-16.md) returned the same structural verdict as the v2.1.8 run: NOT SOUND for wide sharing as a quantitative provider-margin estimate / ranking, but a strong, unusually transparent scenario workbench. The reviewer credits the site's candor and holds that the remaining blockers are structural (the fitted-residual/extrapolation identification gap, prefill-driven default cost, unanchored default fleet) — remedied only by held-out validation or by removing provider point estimates, both of which are escalated to the owner, not attempted here. This release banks the review's cheap, non-structural, clearly-valid fixes.
Fixes applied (labeling/hygiene; no preset/parameter/engine number changed):
Deferred / escalated (owner decision — see Q-AUTO-2026-07-15-COLDFIX and the overnight handoff): the structural blockers (calibrated uncertainty propagation into the headline; phase-specific MoE serving model; default-fleet reweighting to fitted rows; per-dive component models; separate evidence boards per estimand; removing/replacing the normalized table; annex raw-report "80% CI" language). These are the honest floor of what a scenario calculator can be, and the two cold reviews agree the site is sound as a scenario workbench — the gap is positioning, not defects. Credited to the 2026-07-16 external cold review.
Why. The public-only cold GPT-5.6 Pro review of v2.1.8 (research/gptpro-reports/cold-review-v218-2026-07-15.md) returned NOT SOUND as a quantitative margin estimate / provider ranking while praising the site as a transparent scenario workbench. A full-context council (4 GPT-5.6 Sol perspectives + Opus synthesis; reports/codex-council/2026-07-15-margins-v218-review-*/) adjudicated GO-WITH-FIXES: the numbers are faithful to their source dives and the epistemic machinery (registry classes, weights, ranges) is correctly built — the defects are in the prose that sits on top of it. Finding-by-finding adjudication: research/adjudication-cold-review-2026-07-15.md. This release lands the safe labeling/disclosure fixes; the structural residue (calibrated uncertainty propagation, a phase-specific MoE serving model, default-fleet reweighting) is escalated to the owner and the queue, not attempted here. No preset, parameter, MFU, price, tariff or architecture number changed — labeling and disclosure only.
Blocking overclaims removed (P0). (1) §5 no longer calls the 2026-07-15 blinded GPT-5.6 Pro run an "independent blinded replication" that "corroborates the central band" — it is relabeled a blinded model-generated cross-check and explicitly framed as a robustness comparison, not an independent empirical measurement or matched-estimand validation (same model family, different basket/workload; the 47–89% band is wide enough that containing ~77% is weak corroboration). The registry who string is renamed to match; the record keeps its model-generated class, zero claimant weight and 47–89 range. Every council perspective plus the cold review flagged this — the highest-convergence fix on the site. (2) The Google/Gemini §10 card headline is reframed from "unit CM ~96%" to "not publicly identifiable": the ~96% is demoted to an explicitly-unanchored internal-cost scenario, aligned with the card's own "no public basis to compute Gemini's margin from TPU economics" admission. The engine value and the calculator's dive-replay are unchanged; only the card's prominent, screenshot-bound label changed.
Default-fleet honesty (P1 / cold B3). The hero's unanchored-share warning was factually wrong ("TPU/Trainium have no public serving anchors") and undercounted the problem. It now says TPU v7 & Trainium2 have public serving anchors that are not yet fitted (Trainium3 unanchored), and it names the GB300 leg — 15% of the default fleet priced at an analyst-estimated $6/GPU-hr with no public rack rate. §4 adds a matching clarity clause distinguishing the honest "function of rack-hour price" framing (of the newly-documented bridge anchor) from the deployed calculator's $6 scenario input.
Scenario-not-estimate framing (P1 / cold B1, #29–30). The persistent result-surface identity strip now carries a "selected scenario output — not an identified estimate or probability interval" chip (screenshot-safe). The normalized §10 table is retitled "Same-assumption scenario outputs … (not a ranking)" with the no-uncertainty-propagation caveat moved to the top of the block. The §10 intro reframes "reproducible" to headline-matched (the calculator matches each dive's rounded result, not its method), and the methods-box "Reproducibility" bullet is retitled "Implementation & regression consistency — not empirical validation" and now states that out-of-sample flagship validations, calibrated coverage tests and production-telemetry matches are all zero.
Source-caveat & provenance fixes (P1 / cold #5, #17, #25–28). AMD: the unverified timed-out figures (396–408 gen tok/s, 64–178 s TTFT, ≈$2/M TensorWave proxy) are removed from the page entirely and marked excluded pending source recovery; the confirmed 2,256 / 3,500+ tok/s/GPU figures are labeled aggregate. DeepSeek 70–80% relabeled press-reported (not "company-reported"), page and registry. Trainium 405B precision corrected to "FP8-rescaled target weights in a BF16 execution setting, plus speculative decoding." TPU v7 $6.42/M anchor labeled a vendor saturation run with no achieved TTFT/TPOT. NVIDIA: B300 $0.289/M labeled an accelerator-rental floor before non-GPU overhead; AWS Capacity Blocks labeled reserved/upfront. The FP4 1.85× multiplier is labeled a whole-model system benchmark / upper-bound scalar (selective-quantization models won't realize it fully). "Realized billings/price" → "modeled effective billings/price" on the hero and in the methods box. The Rubin bar in the generation chart no longer prints an absolute $/margin figure — it renders shape-only, marked "unpriced," with a caption caveat. "Full, unedited" annex language corrected to "full public reports," with a one-line disclosure of the public-release hygiene pass (conclusions unchanged). The serving-feasibility tile is labeled a weight-storage floor, not an SLO/topology check, and stops calling TPU/Trainium/Ascend "GPUs." Version identity aligned: footer engine string, ENGINE_REVISION and package.json all now read v2.1.10.
Regression coverage. tests/snapshots.test.mjs gains a content-contract block (§6e) asserting each P0/P1 label is present and each flagged overclaim string is absent, so a future edit can't silently reintroduce them. Full gate green (engine + traffic contract + browser suites). Credited to the 2026-07-15 external cold review and the internal council.
Why expedited. An external, public-only cold review of v2.1.8 (research/gptpro-reports/cold-review-v218-2026-07-15.md) identified one directly code-checkable MAJOR defect; the owner approved applying only that item immediately ("Expedite tonight") rather than holding it for the next weekly cycle, leaving the rest of the review's findings unadjudicated.
Chart-gen billing-mix bug, fixed. The "Cost per 1M output tokens across hardware generations" chart (#chart-gen) computed its billed price mix with a local formula that used the serving-side cacheHit value as the billable cached-input share and ignored the separately editable billCacheHit field entirely, along with cacheWriteShare/cacheWriteMult cache-write billing — all of which the hero's headline margin already respected. Defaults coincide (billCacheHit: null assumes equal to serving reuse, cacheWriteShare: 0), so the two surfaces agreed for every casual viewer; the divergence only appeared once a viewer moved the cache-billing controls away from their defaults, at which point the hero and the chart silently disagreed under supposedly identical "current settings." Reproduced directly against site/engine.js before fixing: at Opus/median with blend forced 100% onto the H200 generation and billCacheHit=20 / cacheWriteShare=30 / cacheWriteMult=150 (all non-default), the hero priced the mix at $5.245/Mtok (79.80% margin) while the chart's old formula priced it at $3.268/Mtok (67.58% margin) — a 12.2-point margin gap on the exact same hardware and settings. Fixed by extracting the billing math the hero already used into one shared engine function (computeMix(), consumed by both the existing blended workload() and a new single-accelerator workloadOnHw()); the chart and the existing per-accelerator margin chart now both call workloadOnHw() instead of any local reimplementation — there is no longer a second price-mix formula anywhere in the codebase for either surface to drift from. A grep of site/app.js for other local priceMix/costMix arithmetic found none: every other chart and table already computed margin via the shared workload() function; chart-gen was the sole offender.
Regression coverage, verified both ways. Two new checks were added and each was confirmed to fail against the pre-fix code before being confirmed to pass against the fix: (1) tests/snapshots.test.mjs gained a numeric parity assertion (workload vs workloadOnHw under non-default billCacheHit/cacheWriteShare — crashes with workloadOnHw is not a function pre-fix) plus a source-guard check that renderGenChart()'s body calls workloadOnHw( and no longer contains cacheReadMult/a local priceOut + formula; (2) tests/run-app-tests.sh gained an end-to-end browser check — a crafted permalink (blend 100% H200, billCacheHit=20, cacheWriteShare=30, cacheWriteMult=150) renders the hero at "unrounded: 79.80%" and the chart's H200 bar at "79.8% margin" (pre-fix the bar read "67.6% margin," which the test now asserts is absent).
No preset or parameter numbers changed — this is a bug fix to how one chart computes an already-correct set of inputs; the engine's fitted MFUs, prices, cache tariffs and architecture assumptions are untouched. Credited to the 2026-07-15 external cold review.
Why expedited. Second same-day expedited release: a round-2 GPT-5.6 Pro dive package targeted the NVIDIA forward side (Blackwell-Ultra/Rubin) and Huawei Ascend/CloudMatrix — two of the model's remaining anchor-quality gaps — and the owner ordered both applied immediately ("Expedite tonight") rather than held for the next weekly cycle.
GB300 NVL72 moves from throughput-unanchored-in-practice to audited. MLPerf Inference v6.0 (Apr 1, 2026) contains valid, reproducible single-rack GB300 generated-throughput results on DeepSeek-R1 FP4 — NVIDIA Interactive 250,634 / Server 400,437 / Offline 647,076 gen tok/s/rack; Nebius Server 575,580 / Offline 673,936 (the strongest one-rack results). GB300's rack rental price remains genuinely unanchored — no numeric rate is public on AWS, CoreWeave, Nebius, GCP, Azure, OCI or Crusoe as of Jul 15, 2026 — so the model documents GB300 $/M-output as a function of rack-hour price rather than a point estimate. The cleanest present Blackwell-Ultra anchor is a same-provider B300 pair instead: Nebius's 8-GPU MLPerf Server result (60,413 gen tok/s) at its own public $7.85/GPU-hr rate ⇒ $0.289/M generated output tokens. A GB200 rack bridge via AWS's Capacity Block rate ($761.904/rack-hr) gives $0.881/$0.630/$0.435 per M output at Interactive/Server/Offline. Added to §4's narrative and hardware table, and to the gb200/gb300 dossier notes in engine.js — no flagship engine numbers changed.
Rubin's absolute-economics negative is confirmed, precisely scoped. The same dive found no MLPerf submission, no InferenceX result (listed "Coming Soon"), and no public rental or purchase price for Rubin or Rubin Ultra. NVIDIA's only public claim is relative — up to 10× tok/s/MW and ~1/10 cost per M tokens vs GB200 NVL72 on Kimi-K2-Thinking (32K-in/8K-out) — not an absolute anchor, and NVL144/Rubin CPX/Rubin NVL8/R100/VR200 should not be collapsed into one generic "Rubin" figure. Added to the RUBIN dossier note and §4's hardware table/narrative.
CloudMatrix 384 / Ascend 910C gets its first full-system generated-throughput anchor. FlexNPU (arXiv:2606.04415, Jun 3, 2026) served DeepSeek-R1 W8A8 on a full 384-card system at ≈633,000 generated tok/s system-wide (≈1,646/card) under TTFT≤1s/TPOT≤50ms — an end-to-end, SLO-constrained measurement, distinct from the isolated-decode anchor (1,943 tok/s/NPU at batch 96) already fitted in this model. No public CM384 hourly rental price exists, so throughput is anchored but cost is not. A cross-source proxy for the older 910B chip (JD xLLM's 709 gen tok/s/card × China Telecom CTyun's public RMB 38.45/hr instance rate) gives ≈$2.09/M generated output tokens (range $1.61–2.80/M) — medium-low confidence, cross-source reconstruction, not audited COGS, and not the 910C row. Added to §4's narrative, the hardware table, and the ascend dossier note; the new full-system anchor differs from the currently-fitted isolated-decode anchor and is queued for a considered refit, not applied tonight (structural-change guardrail).
Total-vs-generated correction preserved. The widely-repeated "6,688 tok/s/NPU" CloudMatrix figure is confirmed as PREFILL/INPUT throughput and an idealized perfect-expert-balancing projection (the measured default is 5,655); the "1,943 tok/s/NPU" figure already on this page is DECODE at batch 96 (≈20 gen tok/s per active sequence) — these must never be blended as simultaneous end-to-end throughput. No site text conflated them, so this is a preserved-correct finding, not a fix.
Ascend 920 exclusion confirmed. No official product page, benchmark, deployment, cloud SKU or price exists; Huawei's own Sept-2025 roadmap goes 910C → 950PR → 950DT → 960 → 970 with no 920. Already excluded from this model (no HW row exists); the dive's negative is now documented explicitly in the ascend dossier note and §4 narrative.
Annex. Both new dive reports are published verbatim in the research annex (research/dive-nvidia-forward.html, research/dive-ascend.html).
Deferred to a considered pass (not done tonight). The CM384 full-system FlexNPU anchor (≈1,646 gen tok/s/card) is a materially different measurement basis from the isolated-decode anchor (1,943 tok/s/NPU at batch 96, ≈9.5% MFU) currently fitted to the ascend row's effDec/effPre. Reconciling the two — and deciding whether the roofline should refit to the full-system SLO-constrained number — is a structural engine change, out of scope for tonight's additive release; logged as a new queue item for the next considered pass.
No preset or flagship engine numbers changed — every item above is an evidence-quality upgrade or a precisely-scoped negative finding; the cost model's fitted MFUs, prices and architecture assumptions are untouched.
Addendum (same day, annex hygiene pass on four dive reports). The TPU, Trainium, blinded-replication and AMD dive pages (research/gptpro-reports/dive-{tpu,trainium,replication-blinded,amd}-2026-07-15.md, all first published in v2.1.7) had their provenance headers condensed for public share ahead of wider circulation: recovery-mechanics language describing the private CDP/browser recovery technique was replaced with plain "recovered from the ChatGPT conversation history after the dispatching session failed" phrasing; internal conversation-ID values were removed from the provenance tables; raw [cite: turn…viewN] / [filecite: …] interface citation markers were stripped throughout (sentence text otherwise untouched); "commissioned by owner" became "commissioned". The AMD page's honest 120-minute-hard-timeout provenance narrative is preserved intact. Verbatim pre-edit copies are archived at research/gptpro-reports/archive/pre-hygiene-2026-07-15/. Beyond the header cleanup and marker removal, the findings/analysis prose is unedited — no engine, data, or claim content changed.
Why expedited. Same-day second release: a "Targeted 4" GPT-5.6 Pro dive package landed four items the owner ordered applied immediately rather than held for the next weekly cycle — a correction to a figure already live in v2.1.6, plus three new evidence-quality upgrades. All four ship together as v2.1.7.
Q-007 correction, PARTIAL (AMD/DeepSeek non-NVIDIA throughput). v2.1.6 credited DigitalOcean/RadixArk's "3,500+ tokens/sec/GPU on AMD MI350X/MI355X" claim without qualification. A follow-up dive (reasoning-summary only — the source hit the ChatGPT-Pro MCP's 120-minute hard timeout) argued the chip is MI355X-only and the headline figure is total, not interactive, throughput. Verified against primary sources before shipping (full chain in logs/weekly/2026-07-15-expedited.md): confirmed — MI350X is not an InferenceX-benchmarked SKU at all, and the primary June 2026 InferenceX article independently describes extending the concurrency sweep to 1,024 as drawing out "the high-throughput, low-interactivity end of the frontier," closely matching the dive's own framing. Not confirmed — the dive's precise numeric split (3.56–3.67K total vs ~396–408 generated tok/s/GPU, 64–178s TTFT) and its derived ≈$2/M-output-token TensorWave anchor: this lives in InferenceX's interactive chart data, not extractable text, and could not be independently pinned. §10's DeepSeek card is corrected to reflect only what verified (chip identity, total-vs-interactive-throughput caveat); the unconfirmed figures are explicitly withheld and flagged for a future CDP recovery pass, not shipped as fact.
TPU now has real cost-per-output-token anchors. A dive found named-model, named-precision public serving benchmarks for TPU v5e, v6e and v7/Ironwood — the strongest is Qwen3-Coder-480B-A35B on four Ironwood chips, 518.86 output tok/s/chip, deriving $6.42/M output tokens on-demand ($2.89/M at 3-yr commitment) from current GCP list prices. This is real evidence-quality progress over "peak FLOPS only," added to the Google §10 card, the §3 anchor-fits methodology, the subtitle, the methods-box bullet and the hardware table — but the anchors are documented, not yet fitted into this roofline's MFU (no flagship engine numbers changed). The load-bearing negative from the same dive is preserved and stated explicitly: there remains no public basis to compute Gemini's margin from TPU economics — external rental list prices are not Google's internal fleet cost, and no Gemini-SKU→TPU mapping is public.
Trainium2 gets a narrow engineering anchor; Trainium3 and Project Rainier remain unanchored. Two AWS Neuron tutorials on a full trn2.48xlarge (16 chips), paired with the current Capacity Block rate, give infrastructure-only candidates ($68.90/M output tokens for Llama 3.3 70B, $98.65/M for Llama 3.1 405B) — engineering reference points (batch=1, concurrency=1), explicitly not production TCO. Added to §6 (Anthropic's own fleet section) alongside the more load-bearing negatives the same dive documented: no public Trainium3 throughput or instance price exists at all; Project Rainier is confirmed running Claude inference on ~500,000 Trainium2 chips (AWS, Nov 2025) but discloses no allocation, utilization, tokens or rate; no Trainium MLPerf submission exists through v6.0; AWS's 30–40% price-performance claim is unreproducible.
Independent blinded replication corroborates the flagship band. A dive rebuilt this page's own unit-serving-margin metric from scratch, bottom-up from public anchors, under an explicit instruction not to consult this site or its repository — and disclosed that neither ever appeared in its search results. Result for an equal-dollar GPT-5.5/Gemini 3.1 Pro/DeepSeek V4 Pro basket: central 73.3% (cross-model median 71.8%), range 47–89%. This page's deployed default (~77%) sits just above the central estimate and inside the range. Added as a new gptpro-blinded-replication-733 record in the MARGIN_CLAIMS registry (its own zero-claimant-weight group, same pattern as the existing GPT-5.6 Pro consult record) and as a corroboration paragraph closing §5.
Annex. All four dive reports (including the AMD reasoning-summary-only one, clearly labeled as partial) are published verbatim in the research annex for the first time this release (research/dive-tpu.html, dive-trainium.html, dive-replication-blinded.html, dive-amd.html).
No preset or flagship engine numbers changed — every item above is an evidence-quality upgrade, a documented negative finding, or a corrected/hedged prose caveat; the cost model's fitted MFUs, prices and architecture assumptions are untouched.
Evidence board (Q-001). The Information reported (Jul 14–15) DeepSeek's annualized revenue nearing ~$500M, a ≈¥50B (≈$7.4B) raise at a ≈¥500B (≈$74B) valuation with STAR Market IPO prep, and a follow-up 70–80% V4 API gross margin (up from an earlier ">50%" figure). Added to the evidence-board claims registry as a reported, different-metric figure (api-product-line-GM) spanning the 60–80% and 80–90% buckets — the first company-reported post-V4 margin datapoint identified in this research, versus prior community backsolves. §2a's disclosure narrative gains a dated paragraph noting the directional corroboration (with the metric-distinction caveat kept explicit, as for the 545%/84.5% figures). §10's DeepSeek dive card gains a dated paragraph with the TI figures and TeorTaxes's fleet backsolves (~5,662 GPUs at list-price/100% utilization vs ~9,000–18,000 at realistic bundle-cost utilization) as an independent check against the dive's own fleet-mix uncertainty, with the paid-API-only scope caveat TeorTaxes states himself.
Non-NVIDIA production throughput (Q-007). DigitalOcean and RadixArk publicly reported serving DeepSeek V4 on AMD MI350X/MI355X at 3,500+ tokens/sec/GPU (HIP graphs, a claimed ~10× gain over an undisclosed baseline), attributed to a June 2026 InferenceX result — added to §10's DeepSeek card as a dated third-party data point. It does not resolve the card's own "V4 production throughput" known-unknown (it is not DeepSeek's fleet), but it is evidence non-NVIDIA serving of V4 is technically viable at competitive per-GPU throughput.
No preset or parameter numbers changed — both items are reported/community figures and infrastructure disclosures, not first-party pricing or architecture disclosures; the engine's cost model, presets and dossiers are untouched. Both items independently corroborated by a recovered GPT-5.6 Pro consult (a private working-copy document, not part of the published annex), whose own confidence ratings (medium / medium-low) are reflected in the hedged prose above.
Addendum (same day, repo-governance pass). A GPT-5.6 Pro inspection of the freshly public repository (21 ranked findings; adopted where valid) drove: the annex renderer no longer converts single-tilde pairs into strikethrough — "~¥50B (~$7.4B)"-style money figures had been rendering as struck text in three annex pages (the fix lives in the generator; preserved source documents stay byte-identical); the public snapshot is now self-contained (MCP package build/test inputs, worker overrides, the feedback admin page and runbook ship with it) and self-verifying (GitHub Actions runs the engine, contract, browser, MCP and feedback-bundle checks on the public tree); the README was restructured around a reproduce-one-result proof block with the metric's exact list-price conditional and accurate review-history framing; CITATION.cff, SECURITY.md, CONTRIBUTING.md and third-party font notices added; the private release script now regenerates the grounding ledger and annex before the test gates rather than after.
Permalink round-trip defect (P0), fixed. The v4 encoder diffed the scenario against global DEFAULTS while the loader restores a clean-identity link as preset-plus-diff — so a field the user set TO a global-default value that happens to differ from the preset's own default was omitted as "unchanged" and silently reverted to the preset value on load. Concretely: GPT median with active set to 300B round-tripped to 105B, moving the margin 77.9% → 92.3% — a 14.4pp silent shift on an ordinary user action, in a shared link. The encoder now diffs against the loader's own baseline (the declared model/perspective/traffic identity, resolved exactly as the loader resolves it); MODIFIED-identity links keep the global-DEFAULTS baseline because that is the baseline the loader actually uses for them. 66 new release assertions sweep every model × every global-default-valued field case plus the concrete reproduction. Found by the 2026-07-15 GPT-5.6 Sol release inspection; links minted before this release that hit the dropped-field case are unrecoverable (the information was never encoded) — the engine-revision drift warning already flags them on load.
Source-currency corrections (P0). The incompressible-knowledge-probes citation now carries the paper's v2 figures (arXiv 2604.24827, revised Jul 5, 2026: median fold-error 1.48×, 86% within 3× — the page had v1's 1.59×/87.6%), pinned to the revision date. §4's cumulative 2024→2027 hardware-cost multiplier ("roughly 6–25×") did not reproduce from the page's own per-generation chain; it now states the product of that chain (~5–16×), with the commissioned fixed-model projection's band cited alongside (~4.5–10× by the 2027 Rubin ramp, ~7–17× by mature Rubin).
Smaller corrections. §10 xAI no longer says the prospectus "discloses a fleet of >440k accelerators" — it disclosed cluster counts; the total is a derivation (the dive always labeled it SPECULATION-derived-from-DISCLOSED). Freshness stamps unified at July 15, 2026 (the landing text said July 11 while the engine said July 13; all 12 tracked list prices were re-verified against provider pricing pages today — 12/12 confirmed, including Kimi K2.7 Code's $0.19 cache-hit after a K2.6-page false alarm). The annex generator's skin-boot extraction is now genuinely fail-closed (exactly-one-match asserted, overmatch into a foreign script refused) — previously a duplicated or unterminated marker could silently ship wrong boot code despite the fail-closed claim. The live feedback form posts via its custom domain instead of the name-bearing workers.dev host (deployed same-day, ahead of this release).
Adjudicated, not changed (Sol inspection findings held up as design). The Sonnet 5 tariff flip already has a stale-loud release gate (a traffic-contract assertion fails after 2026-08-31, and deploy.sh runs it before every deploy). Model switching in a MODIFIED state deliberately freezes model-owned fields (documented "Residual 2"; every surface labels the state modified and its permalinks travel the full numeric state, so nothing silently misattributes). Provider-dive uncertainty language lives in preserved consult documents, which this annex archives verbatim rather than edits. Queued for the next update (research/update-queue.md): grounding-ledger source rows gain URLs/dates; DeepSeek's announced mid-July peak surcharge needs live verification; Opus 4.7 fast-tier wording flips on its Jul 24 retirement; The Information's DeepSeek ~$500M ARR / 70–80% V4 API gross-margin report (Jul 14–15) enters the evidence board.
Evidence-scent label pass. Page-set scenario values lose every remaining trace of false empirical scent: "top-decile"/"near-top" stack wording (implying a measured distribution that does not exist) became "aggressive"/"above-baseline" with explicit page-set qualifiers; the utilization slider's "typical" tick — contradicted by its own tooltip's "no representative industry distribution is public" — became "central scenario"; "300B is our median read" became "this page's working estimate"; the "DeepSeek 5:1" tick was corrected to the file's own measured 4:1 profile; "FP8 (frontier default)" is now "(assumed frontier norm)" (closed labs don't disclose serving precision); "evidence median" — naming a fallback that is actually the central scenario, in three places including report prose — now says what the code does; and dossier rows where page-chosen levels wore COMMUNITY ESTIMATE labels were reconciled to SPECULATION. A pre-existing display incoherence was fixed structurally: Kimi/Grok model dossiers annotated traffic parameters with dive-mix sources while rendering Reference-convention values beside them — the dive-only annotations are removed and the dive mixes keep their provenance in the effective-traffic line, where those values actually render. The exploration-surface lint's regex gap (caught "assumption" but not "assumed") is closed. Four GPT-5.6 Sol find→fix→verify rounds, looped to clean on P0/P1.
GPT-5.6 Pro outside-perspective review, adjudicated and remediated. A cold outside-expert review of the live site (18 findings; archived with the adjudication in this annex) was adjudicated finding-by-finding against what the site already says: most findings were already mitigated by welded caveats the reviewer had no context for, several relitigated deliberate design, and six were valid — all cured here. §3's "from first principles" (a derivational status the very next paragraph disowns) became "from an explicit cost identity". §5's "for any frontier model" universal was falsified by the site's own engine (DeepSeek V4 Pro at its own list never enters the 90–95% zone on a GB300 fleet at any utilization) and is now bounded to Western-lab list tariffs with that counterexample stated. The §3/methods-note validity domain now states the size axis plainly: all published serving anchors are ≤~50B-active models while flagship scenarios assume 120–300B active — the model's largest unanchored extrapolation. §5's "margin machine" carries the billable-share caveat inline (86.2%→61.5%). The Zhipu card no longer converts the audited 18.9% segment margin into a "floor" (it is a segment-margin anchor) and its reseller price ceiling is now stated as conditional — the same conditionality the Kimi card always carried. Two currency errors fixed against primary sources: §5's "~500k Trainium" contradicted the site's own hardware table and Anthropic's Apr 20, 2026 statement (">1M Trainium2 in use"); TPU7x GA is Mar 31, 2026 per Google's release notes, not Apr 22 (a Cloud Next conflation).
Provenance upgraded, not just defended. The Anthropic inference-infrastructure margin record (38% → >70%, SemiAnalysis) had been carried as reported-unverified via X relays under the registry's mis-attribution rule ("the newsletter is paywalled"). The margin sentence in fact sits in the newsletter's free portion; it was fetched directly, archived verbatim in this annex (curly apostrophe and all), and the record now quotes the primary source — the same upgrade pattern the Patel/Sequoia citation still awaits. The grounding ledger's version stamp — which claimed "cannot drift" while stamping the previous engine version — regenerates with this release (the no-drift guarantee now includes its own stamp).
Still owed (unchanged). The five-human comprehension test remains the open ship gate; the Patel primary-citation swap and Zephyr re-capture remain flagged owner actions.
The redesign. Person-named analyst presets (TeorTaxes, Zephyr, Dylan Patel/"semi", "skeptic") are reworked into a mechanism-first calculator — the hero margin is still driven only by model/traffic/lens, never by a target — plus a first-class secondary margin-range evidence board fronted by a route-first entry ("what would have to be true for a ~90% margin?") rather than an evidence catalog a reader has to read past. A typed MARGIN_CLAIMS registry bins claims, not people: floors like Dylan Patel's "north of 80%" render as compatible-with every bucket at or above 80 (never as interval membership — a floor is not a ceiling); Zephyr's 90–95% is carried in its own unnamed-subject group, never presented as an Anthropic claimant; company-GM figures (Zephyr's own 70%), reported figures, DeepSeek's disclosure anchor, and the GPT-5.6 Pro model-generated consult each sit in visibly distinct, differently-labeled groups so no serving-context number can pose as the calculator's unit metric. The four analyst presets are absorbed into de-named, page-authored "[range exploration]" routes — numeric-identical to the retired presets, migrated so old permalinks and browser-saved presets resolve to the same numbers under the new ids. Routes exist only where popular discourse actually places a claim: the 60–80% bucket has none (no unit-serving claimant sits there — the central scenario is its own anchor), and the ≥90% bucket has exactly one, because the engine's own reconstruction of the cited "increase the batch size … drive margins from 90% to 95%" mechanism tops out around 87–89% at market rates — the routes that did clear 90% were owned-fleet procurement stories, not the batch mechanism, and were kept separate rather than mislabeled.
Gated the way v2.1.x releases are. A pre-implementation plan review (GPT-5.6 Pro plus a four-persona council) ran before any code, and a second, blind four-persona final gate ran against the built M1–M4 milestones. The final gate returned NO-SHIP, unanimous, with a P0 cluster — all remediated in this release: share-link identity integrity is now fail-closed — a v4 token whose declared model or perspective does not resolve is refused whole and the default state renders, rather than applying the numeric diff under a substituted or mislabeled identity (the exact corruption class v2.1.1 had already fixed once had regressed on two new paths: copying a link after mutating a loaded route, and a crafted token naming a nonexistent model while real selectors stayed on screen); saved presets now carry typed origin and reload as an explicit modified state rather than silently inheriting whatever lens happens to be selected at load time. Board framing lost its "who claims/who places frontier inference margins there" language — an umbrella phrasing that quietly turned a floor into an endorsement despite the correct row-level "compatible with" badge — for explicit relation-aware wording, and reported figures are now labeled "reported figure," never "verbatim," when no archived quotation exists. The hero's ">90%" caption is now interval-aware: it asserts "within the cited 90–95% range" only when the value actually lands there, and says the value sits above that interval otherwise (it had been claiming every ≥90% result, including ~100%, was "within 90–95%"). The model-sizing name block (which had been leaking real names into a loaded route's dossier under a prose disclaimer) moved into a structurally separate, separately-collapsed panel outside the route dossier — quarantine by structure, not by caveat. The evidence board also picked up an intentional 2×2 layout at desktop widths (auto-fit had been stranding the fourth bucket card 3-plus-1) and the v4 permalink schema now validates metadata-carried traffic against the same bounds the sliders enforce, closing an out-of-range/bad-mode injection path.
Also fixed a pre-existing v2.1.2 defect the deep review surfaced along the way: moving a traffic slider or tick updated the hero margin correctly but left the dossier's "Effective traffic mix" line — and the accompanying status note — showing the previous value, because renderAll/fullRefresh never re-rendered the dossier on a plain state change (only certain load paths did). The hero margin itself was never wrong; only its own explanation of itself was stale. The dossier now re-renders on every state change, so the displayed traffic identity always equals the resolved traffic and the number the margin was actually computed from.
Source provenance (honest). The four load-bearing X posts behind the routes and claim records were re-checked against primary sources rather than re-trusted from the archived sweep alone. Dylan Patel's "north of 80%" is now independently confirmed at a primary source — the Sequoia Capital podcast, 2026-07-02, transcribed by three separate outlets — where the site had only a secondary clip account (PodcastAlphaX); swapping the citation to the primary transcript is a recommended follow-up, not yet done. The two TeorTaxes posts and the Zephyr post that the ≥90%/80–90% routes and the Zephyr unnamed-subject claim rest on are unretrievable by automated means — X is login-walled for this research and no Wayback or mirror capture was found — so their durable provenance in this repo is the dated archived research sweep (research/grok-sweep-margin-claims.md), transcribed at the time, not a live-primary re-verification, and that is stated plainly rather than papered over. Zephyr's load-bearing "names no lab" scope point in particular rests on that sweep transcription and is flagged for manual re-capture; the site errs conservative in the meantime — it treats the 90–95% figure as about an unnamed provider and never attributes it to Anthropic.
Still owed (not softened). The five-human, 90-second, 4/5-comprehension-bar test remains open and unrun — it is the actual gate this whole redesign exists to pass, and the two review rounds (plan review, final gate) are a code/copy/provenance pre-check with no established error rate against real readers; they do not substitute for it. A smaller, label-only gap also remains: some status notes omit the traffic-mix figure even though the dossier's own traffic line always carries the correct number (no display shows a wrong value, only an incomplete one in a few paths) — logged in BACKLOG rather than left undocumented.
The plan for this release was itself reviewed pre-implementation by GPT-5.6 Pro and a four-persona council (both archived in this annex); both said "revise before implementing," and the revised plan adopted every P0 condition (the final gate later found one adoption — the traffic state contract's application layer — incomplete, and a second remediation round closed it; see below) — including renaming the new axis honestly (TRAFFIC MIX, not "workload"), spec-first fixtures, and redesigning the planned named-person reception simulations into corpus-bounded source-faithfulness audits with no role-play and no approval outputs.
Traffic mix — I/O + cache (the scoped ontology step; the full resolver remains future work). A third selector owning only ioRatio/cacheHit: provenance-named profiles (no generic "chat/coding" archetypes), model defaults that visibly re-resolve, replays as atomic composites with locked traffic, an anti-lens-shopping span computed at byte-identical traffic ("cost-lens span at [profile]"), v3 permalinks carrying traffic identity, numeric-identity migration for v2 links and browser-saved presets, and a 180-pair frozen baseline proving parity (documented intentional deltas only). The state contract was committed as failing fixtures before implementation.
Coherence cuts. §6 became an assumption-by-assumption divergence table (the full 54-minute consult is now published verbatim with a SHA-256 stamp — recovered from the receiving session's transcript after an earlier revision mislabeled a 39-line digest as "verbatim"); §8 states a longitudinal selection rule; §10 cards carry metric/lens/traffic lines, top-3 evidence lists with the dominant driver marked, and would-update-on triggers.
Durable exports. The full roofline derivation (verbatim, SHA-256), the adopted grounding ledger (machine-generated from the preset registry — the authoritative parameter record), and a dated GPT-5.6 Pro re-emission of the expired preset pack with a checked delta table (the ledger wins).
Final gate — NO-SHIP, then round 2. The four-persona council final gate NO-SHIPPED the first remediation with a precise list: a schema-downgrade permalink forgery (a v3. token declaring an inner v2 schema routed through the lock-overriding legacy-migration path and rendered ≈71.6% under the locked replay's name), type-only sanitizer checks, replay "exit" that was actually a warning on one input path, unimplemented v2/browser-preset identity migration, an incomplete metric rename, a lens span computed at the hidden model default under locked replays, and a release stamp that made the served tree unreproducible. All were fixed in a second round, each encoded as a permanent regression test; the reception resolution ledger's affected rows carry the reopened-and-re-closed history. GPT-5.6 Pro's independent final gate ran in parallel (archived in this annex).
Reception phase — and what it caught. Nine simulated-reader audits ran against the frozen preview (five third-person source-faithfulness audits for TeorTaxes, Zephyr, Dylan Patel, fleetingbits and the @_xjdr deployment numbers; four anonymous archetypes: ML-infra engineer, calibration reviewer, cost-accounting reader, information-scent lurker). They found 16 P0-class defects the two prior review rounds missed — among them: the metric name "list-price" contradicting the realized-billings denominator; a forged-permalink path that bypassed replay-traffic locking; the §3 equation prose claiming interactivity applies to prefill; the TeorTaxes preset pricing H100 below its own cited $2.40 anchor; a site-authored 8:1 ratio presented as ncode deployment telemetry; §1 and dossier scope-expansions of Zephyr's unnamed-lab allegation; unknown /research/ URLs serving the homepage as HTTP 200; and the normalized table falsely claiming the §10 ranges "still apply." All findings are dispositioned in the published resolution ledger (45 fixed after two rounds, 0 rejected — five first-round closings were reopened by the final gate and re-closed; the ledger keeps that history); the hero was de-named (bands describe cited ranges, never people), analyst presets carry always-visible PAGE-AUTHORED RECONSTRUCTION boundaries, precision dropped to whole points with the unrounded value as a diagnostic, and the release gates grew to 499 engine + 220 contract assertions plus a 17-check application-level browser suite that renders the page headless and replays the live attack payloads. Raw outputs, prompts and the corpus manifest are published verbatim. This exercise is not validation and does not satisfy the still-open 5-human-reader test.
Two attribution-fidelity defects found live by the v2.1.2 plan-review round (four-persona council, empiricist) and fixed: the §2c _xjdr blockquote had silently normalized "2.1 sec TTFT overage" to "average" and dropped the "61 sec p95 TTFT (1M ctx)" and "0 chat logs kept" lines — it now matches the archived sweep verbatim ("overage [sic]"), and the prose carries the 61-second p95 as the observed cold-prefill tail. Also repaired a grammatical break the Jukan removal introduced in §6. Both plan reviews are published in this annex (consultation-2026-07-11-plan-review-*).
Removed the §1 "attribution check" passage about Jukan and the related §6/method-note handle-dispute sentences (owner call): since no margin claim of his was ever found, the two research engines' disagreement over his handle changed nothing in the analysis and spent reader attention on research-process trivia. His two substantive citations remain as bare links — "inference is memory" (§4) and the $5B-spend→$15B-ARR lease-economics read (now in §7, labeled a spend-to-revenue multiple, not a margin).
A second four-persona council review plus a GPT-5.6 Pro final gate ran against the deployed v2.1 and returned NO-SHIP with a surgical fix list (archived in this annex). All gate items repaired; the numeric engine is untouched (the reviews verified 5/6 dive replays to a point and called the arithmetic sound):
Backlog implementation round, consulted and gated: a four-persona council design review (typed-ontology and ship-list adjudications), a GPT-5.6 Pro preset grounding pack (first-party tariff verification), and a GPT-5.6 Pro roofline consultation (negative result, adopted) — all archived in this annex.
Position dossiers. Every model and perspective preset now expands into a dossier: whose position it is (quote-anchored), an attribution-honesty line (quoted-position vs reconstruction vs calculator-synthesis — a quote anchors a claim, not a parameter vector), every parameter it pins with source + evidence label, what it assumes away, and what would falsify it. Values render live from the presets (set-canonical), so dossiers cannot drift; the test suite asserts bidirectional key coverage.
Typed scenario ontology. The perspective dropdown now types its entries — [lens] / [analyst] / [replay] / [SLO replay] — making visible that cost-accounting lenses, quoted analyst positions and measured-operating-point replays are different kinds of object. The hero tile shows the margin range across all lenses for the selected model, as a guard against lens-shopping.
New presets (per the council's ship list). Models: DeepSeek V4-Flash (disclosed 284B/13B), GLM-4.7 (disclosed 355B/32B — a distinct price floor from 5.2), and GPT-5.6 Terra/Luna + Gemini 3.5 Flash as clearly-marked TARIFF SCENARIOS (price is the only identified quantity; excluded from the normalized table). Perspectives: xAI cash-marginal and opportunity-cost lenses (xAI-scoped, pairing-warned), China public-cloud on-demand (≈6.15× IDC rates, dated), and an Ant Group H20 SLO replay (Pro tier, measured-operating-point convention). Rejected with reasons: Kimi K2.6 (economically redundant), Grok 4.3 (adds surface, not identification), Anthropic-strategic (merged into the existing strategic-partner lens), Epoch (deferred until a roofline exists), reported-margin-implied (an inverse manifold, not a preset — now a §7 diagnostic sentence). All tariffs verified at first-party sources on 2026-07-10.
Roofline: a documented negative result. The commissioned compute/HBM/fabric roofline (one shared residual, no per-platform efficiency) predicts throughput-oriented operating points within 2–16% but fails the whole-platform gate (Ascend@15ms −39.8%) and the preregistered validation is formally unrunnable on public data (one fitted parameter vs one training platform; no untouched holdout). Anchor fits remain primary; the roofline ships as tests/roofline-diagnostic.mjs and the methods note carries the full result.
Also: permalinkable scenarios (versioned, dated, with a share-time privacy note), pairing warnings for cross-scoped model/perspective combinations, cache-WRITE billing (revenue-side, conservation-safe, default-off; storage remains a documented omission), and the §7 bridge re-framed in dollars per Mtok with a named unreconciled-accounting-perimeter residual.
Revision driven by two independent external reviews, both archived unedited in this annex: a four-persona GPT-5.6 Sol council with Claude Opus synthesis, and a GPT-5.6 Pro adversarial review. Both reached the same verdict — sound thesis, publication-blocking presentation — and their findings substantially overlapped.
Calculator (engine):
tests/snapshots.test.mjs) runs before deploys.Report:
Initial publication: Anthropic-first investigation, interactive calculator, per-provider deep dives (§10), Chinese accelerator support, research annex.
f361c29)What. Owner ruling R-4 says the long explainers at the bottom "shouldn't be fully expanded to begin with." Measured, the full report was 63,907 characters visible at first paint across ten always-open sections — the page's only violation of that ruling. Those ten sections now collapse, each keeping its heading visible, and every long explanatory surface on the page — the six provider dossiers, the methods box, the perspective and model dossiers, the evidence catalog, the range detail — gains a legible measure and a "Deeper explanation" route into the near-fullscreen popup M6 built. On a touch screen the inline expansion is replaced by that popup, which is what R-2 asked for; on desktop the collapse stays and both routes work, which is what R-3 asked for.
The report's section headings deliberately stay outside their own collapsible bodies. That is
not a style decision. Two separate runtime extractors serve report-s1…report-s10 to MCP clients
by slicing from one <h3 id="sN"> to the next, so wrapping a heading inside its <details> would
hand consumers structurally unbalanced fragments. Keeping the heading outside also means a deep
link still lands somewhere visible when its section is closed.
What an MCP consumer receives did not change by one byte, and that is executed rather than
claimed. The expand/collapse label is interface chrome, not part of the document, so it is
stripped from the two transport extractors by a rule scoped to the report wrapper alone — the six
provider-dossier summaries inside §10 are ordinary <summary> elements and had to survive, which
they do by construction. All eleven ids (the ten sections plus the whole front page) are
byte-identical to a fixture captured before any of this markup existed, in both normalized and raw
text, and identical between the Node and Worker transports. The test carries a control that fails
when the strip is switched off, because a parity test that cannot fail proves nothing.
Nothing is copied into the popup. The dialog holds the same live nodes the page was showing, moved and moved back — so the evidence catalog's buttons still work inside it, nested disclosures keep their open state, and no id is ever duplicated. That made the restore path load-bearing for the first time, and it was not safe: a failed restore used to be swallowed while the dialog was removed anyway, taking the live content with it. Restoration now carries fallback coordinates, the dialog is only removed once every moved node is provably back in the document, and if both routes fail the dialog is kept and flagged rather than deleting anything. The suite injects both failures.
Two defects the acceptance suite found in the author's own work. Moving the popup trigger above
each collapsible body — so that expanding §10 does not bury the affordance 33,472 characters down —
made that trigger the heading's immediate sibling, which silently broke the rule that opens a
section when a citation links to it. And selectors like #prov-openai > .prov-body stop matching
while an ancestor is temporarily relocated. Both are fixed; the first keeps the broken version as a
test control.
No number moved. The 180-state render hash is unchanged, both tripwires re-derive exact, and the count of claim-bearing render sites is identical before and after — moving an already-rendered node into a dialog is not a second rendering of it.
032e32f)The page carried explanatory notes that no reader could reach. TIPS.specDec is 3,352 characters;
it rendered a 1,328px box in an 844px viewport positioned at top: -494px, and because the tooltip
is pointer-events: none there was no way to scroll to the part that was cut off — which was the
beginning. showTip's clamp was one-sided on both axes, so any box taller than the viewport
produced a negative coordinate that nothing corrected; at a 320px viewport the horizontal bound went
negative too and every tooltip sat 10px off screen.
The clamp is now two-sided on both axes and holds even if the CSS caps fail. Clamping alone would
only have truncated the note silently, so every ⓘ control now opens the full text in the
explanation dialog — on touch devices that is the only route, since no hover exists there and the
tap target was 16×12 CSS px. It is 44×44 there now. The hover preview stays deliberately
non-interactive: drawing a scrollbar on text you still cannot reach would have been the same defect
wearing a fix.
Adding a second dialog client is what forced the rest. The final-answer dialog closed through a private path and removed dialogs by its own marker, which with two clients could leave two dialogs open at once. Both now route through a single coordinator with one idempotent close bound to every exit — including a throwing payload builder and a page rebuild that destroys the trigger the dialog came from. The final-answer surface itself is unchanged: its payload construction is byte-identical and its own browser suite passes completely unmodified.
No computed value moved. The 180-state render hash is unchanged and both tripwires re-derive exact.