← Frontier Inference Margins · all research reports

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

Reception phase — synthesis & resolution ledger (2026-07-11)

What this is: a simulated-reader exercise — nine independent GPT-5.6 Sol (xhigh) audit runs against the frozen v2.1.2 preview (deploy 201b22a7, git 9f1fd0f), designed per the plan reviews: five corpus-bounded source-faithfulness audits (third person, no role-play, no approval outputs) and four anonymous audience archetypes. What this is not: validation. These are model-generated adversarial findings with no established error rate; the absence of a finding is not evidence of correctness, a null result is bounded by the manifest-selected corpus, and this exercise does not satisfy the still-open 5-human-reader comprehension test. No aggregate approval score exists or will be published. Raw artifacts (verbatim): prompts + corpus manifest + all nine raw outputs · run metadata in each file's header. Synthesis rule: findings were deduplicated and dispositioned; none were upgraded, and every row links back to a raw output.

Headline

The audits re-ran the deployed suites and byte-checked the deployed assets against the frozen commit (both green; the anti-lens-shopping span, the three-deployment separation in §2c, the xAI segment-margin handling, and the $5B→$15B labeling all drew explicit "no contradiction found in this corpus") — and then found 16 P0-class defects and ~26 P1s the two prior review rounds had missed. The strongest theme, hit independently by five audits: attribution boundaries existed in expanded dossiers but vanished at collapsed point-of-use surfaces (selector names, hero zone labels, preset notes, ledger source cells). The single most damaging class: the calculator's own voice converting page-authored choices into source-attributed facts.

Convergent findings (multiple independent audits)

Final-gate re-audit — round 2 (2026-07-11, later)

After the first remediation round, the release went to its final gate (a four-persona council + GPT-5.6 Pro, blind-first). The council returned NO-SHIP: four of the dispositions below were overstated as first closed — ACC-1 (metric rename incomplete in the tooltip/§5/§6), MLI-1 (the sanitizer blocked the exact audit payload but not a schema-downgrade forgery or out-of-range/enum/blend values), MLI-8 (a warning is not an exit; ticks/radios/selects/blend/URL mutations bypassed it), LUR-5 (a printed commit is not reproducibility — the stamp step itself dirtied the tree) — and the MLI-5 deferral was challenged (an 86.2/76.8/61.5% Opus envelope is material). A second remediation round closed all five: token-prefix schema authority with disagreement rejection; full range/enum/blend-shape validation; divergent overlays and every edit path (sliders, ticks, radios, selects, blend, URL overlays, saved-preset loads) now EXIT replay identity to an explicit "[modified scenario]" state; v2 links and browser presets recover named-profile identity on exact match; the metric rename completed; the cost-lens span computes at the resolved replay traffic; a "Billable cached-input share" slider makes the cache-conflation assumption breakable (default remains the labeled assumed-equal); and releases are now reproducible (clean-tree gate, release-stamp commit, served-asset hash manifest at /asset-manifest.json). An application-level browser suite (tests/run-app-tests.sh) encodes the live attack payloads as permanent regressions. The rows below carry their round-2 status.

Resolution ledger

Severity is the auditor's; disposition is this page's, with the deterministic acceptance check applied. "Modified remedy" = the objection was accepted but implemented differently than the audit's exact wording; the reason is stated.

# Audit Sev Finding Disposition
ACC-1 accounting P0 "List-price" metric name contradicts the realized-billings denominator at default batch/discount FIXED (round 2) — first round missed the engine tooltip and §5/§6 phrasing (final gate reopened it); now renamed sitewide. §10 cards keep "list-price" because every dive replay pins batch/discount to 0, test-asserted
ACC-2 accounting P0 Idle capacity simultaneously included (util divisor) and declared excluded (methods) FIXED — cost-perimeter statement rewritten; util tooltip states slack allocation explicitly
ACC-3 accounting P0 §7 dollar bridge used stale arithmetic ($3.10/$3.72 vs actual $3.26785/$3.71875) FIXED — exact current values in prose; suite now pins all three numbers so drift fails tests
ACC-4 accounting P1 "Non-additive bridge" is a sensitivity map inviting bridge use FIXED — renamed "Sensitivity and perimeter map — not an accounting reconciliation"; new "Already in the calculator?" column; cloud row split into cost-side procurement vs revenue-side channel share
ACC-5 accounting P1 Heterogeneous margin observations presented as a time series ("margin ladder") FIXED — replaced with a per-figure perimeter/method/status listing; "not a time series" stated
ACC-6 accounting P1 Hero and collapsed §10 cards unsafe as standalone business-margin claims FIXED — hero label "Unit serving margin — not company GM"; card headlines prefixed unit CM / output-token CM; meta description corrected
MLI-1 ml-infra P0 Forged permalink bypasses replay-traffic locking (71.6% forged vs 26.9% authentic while the UI claims locked 3:1/0%) FIXED (round 2) — the first-round sanitizer blocked only the exact payload; the final gate produced a schema-downgrade forgery (inner _meta.schema:"v2" routing through the lock-overriding legacy path) and out-of-range/enum/blend bypasses. Now: token-prefix schema authority (disagreement = rejection), full bounds/enum/blend validation, and replay-divergent overlays exit replay identity. All attack payloads are permanent app-suite regressions
MLI-2 ml-infra P0 §3 equation prose claims interactivity applies to prefill; the engine scales decode only FIXED — prose corrected; decode-0.5×/prefill-1.0× assertions added
MLI-3 ml-infra P0 Published test suites don't run from the published tree (../site/engine.js) FIXED — deploy script rewrites imports for the served copies; standalone execution verified (both suites green from a bare download layout)
MLI-4 ml-infra P1 Calibration tests skipped the GB300/H20-anchor/Ascend-anchor table claims FIXED — three named assertions added
MLI-5 ml-infra P1 One cacheHit drives both serving reuse and billable cached share FIXED (round 2) — the final gate rejected disclosure-only deferral ("expose it as an unknown input"). A "Billable cached-input share" slider now breaks the coupling explicitly (revenue-side only, engine-tested); the default remains the labeled assumed-equal scenario, and the methods box quantifies the envelope (Opus default: 86.2 / 76.8 / 61.5% at billable 0% / =serving / 95%)
MLI-6 ml-infra P1 Masthead overstated the anchored basis (35% of default traffic on unanchored MFUs) FIXED — masthead names the unanchored platforms; hero shows "Unanchored-MFU traffic share: N%" live
MLI-7 ml-infra P1 Interactivity multipliers look calibrated FIXED — control renamed with explicit multipliers, "uncalibrated", "decode only"
MLI-8 ml-infra P1 Edited replays keep their published-operating-point attribution FIXED (round 2) — the first-round warning covered only slider drags and was not an exit (final gate reopened it). Every mutation path — sliders, ticks, radios, selects, blend inputs, URL overlays, saved-preset loads — now exits to an explicit "[modified scenario]" identity with replay attribution removed
MLI-9 ml-infra P1 "Storage epsilon" asserts immateriality; cache-write=0 hidden from the price tile FIXED — "may be material" wording; live tile note when cache reads are modeled with write billing at 0%
TEO-1 teortaxes P0 The bull preset priced H100 at $2.04/hr, below its own cited $2.40 anchor FIXED — rentMult 0.85→1.0; margins moved (Opus 90.9→89.3%) and the parity fixture carries the documented remediation delta; dossier cell states the DeepSeek-calculation context of the $2.40
TEO-2 teortaxes P1 Reconstruction boundary hidden while the named result is visible FIXED — always-visible boundary prefix for analyst presets + selector rename
TEO-3 teortaxes P0 Ledger source cells converted non-claims into source-attributed assumptions FIXED — every analyst parameter cell now reads "Page-set …; no value in the cited corpus"
TEO-4 teortaxes P1 Unclassified $4/1Mt ceiling treated as an output-token validation target FIXED — §5 and tooltip state token-class is unresolved
TEO-5 teortaxes P1 §1 compressed Anthropic-specific and conditional western-provider claims into one 90–95% attribution FIXED — claims separated with scopes and conditions; "no unconditional 90–95% claim in the cited corpus" stated
TEO-6 teortaxes P1 300B described as a "median" partly sourced to a qualitative Fable observation FIXED — "page-authored scenario value" with the qualitative/relative nature of the cited posts spelled out
TEO-7 teortaxes P1 Hero assigned real-source names to arbitrary states FIXED — de-named bands (see convergent findings)
ZEP-1 zephyr P0 The archived sweep's own "Frontier/Anthropic-style" characterization exceeds the quoted post FIXED (modified remedy) — the artifact is archival, so a dated ERRATUM block corrects both characterizations rather than silently editing the record; the audit's exact-string removal test is intentionally not satisfied, the misattribution correction is
ZEP-2 zephyr P0 §1 "makes the same claim … one analyst's head" exceeded the corpus FIXED — replacement paragraph separates the unnamed-lab allegation from the explicit Anthropic company-GM post
ZEP-3 zephyr P0 Dossier claimed both metric scopes were explicit ("he says both explicitly") FIXED — dossier states which scope is explicit and which is this page's interpretation
ZEP-4 zephyr P1 Parameter cells like "semis-market insight"/"Implied by position" FIXED — page-set phrasing
ZEP-5 zephyr P1 "TeorTaxes/Zephyr 90–95% zone" hero string FIXED — de-named
DYL-1 dylan P1 Quoted-secondary source class undisclosed at both attribution points FIXED — §1 names the clip-account post and the source class; dossier anchor labeled
DYL-2 dylan P1 Preset name + §6 "reproduce ~93%" collapse the claim/reconstruction boundary FIXED — selector renamed; §6 sentence states site-authored translation
DYL-3 dylan P1 $0.99 lost its workload/model scope in the §7 row FIXED — Opus 4.7 agentic scope, 300:1, >90% cache, "not Anthropic-wide"
DYL-4 dylan P1 $1.60 TPU estimate extrapolated fleet-wide under a reported-evidence label FIXED — §6 cell scoped; ledger label now "SPECULATION derived from CREDIBLY REPORTED"
FLB-1 fleetingbits P1 One post promoted into a named "camp" FIXED — "One reported-margin objection"; selector renamed "Reported-margin diagnostic (single-post anchor)"
FLB-2 fleetingbits P1 Hero converts the reconstruction into a source-associated 21.6% FIXED — reconstruction prefix + de-named band
FLB-3 fleetingbits P1 Labeling weak at parameter point of use FIXED — boundary prefix + page-set cells
FLB-4 fleetingbits P1 Tooltip/dossier implied a unit-margin interpretation and inversion not in the post FIXED — tooltip source line and dossier identity/assumption/falsifier rewritten (falsifier now warns a failed inversion does not falsify the post)
XJD-1 xjdr P0 Unknown /research/ URLs served the homepage as HTTP 200 FIXED_redirects for the misnamed sweep path + a real 404 page (no more SPA fallback impersonating artifacts)
XJD-2 xjdr P1 §2c added interpretations beyond the corpus ("read as average", "~30×", "mostly prefill") FIXED — audit's replacement adopted; "overage" left uninterpreted; the p95 stated without a causal story
XJD-3 xjdr P0 The 8:1 in "ncode coding deployment 8:1/41%" is a site-authored assumption presented as deployment telemetry FIXED — profile renamed "ncode-informed scenario", provenance split (41% observed / 8:1 site-assumed), GLM dossier labels corrected to SPECULATION, §10 cardmeta reworded; numeric state unchanged
CAL-1 calibration P0 "The §10 ranges still apply" under the normalized table is false (DeepSeek −4% vs 45–83%) FIXED — replaced with the non-transport statement + no-uncertainty-propagation note
CAL-2 calibration P1 "Roughly 80% of plausible modeled outcomes" implies an undefined measure FIXED — "analyst-elicited scenario ranges … no stated coverage probability"; all six cards read "(scenario range …; uncalibrated)"
CAL-3 calibration P1 Undefined "median" (§5, preset name) FIXED — "[lens] Central scenario"; §5 states the deployed central scenario ~77% and an unassigned-coverage span
CAL-4 calibration P1 §6 verdict lacks model-generated provenance labeling at point of use FIXED — provenance banner inside the verdict box
CAL-5 calibration P1 Tenths-of-a-point hero precision FIXED — whole-point ≈ display; unrounded value in the tile's title diagnostic
CAL-6 calibration P1 Reference convention called "realistic" FIXED
CAL-7 calibration P1 Tooltips smuggled speculative population priors (chat 5-20:1; industry 30-60%) FIXED — named observations/conventions only; "no representative distribution is public"
CAL-8 calibration P1 A forecast inside a "not forecast" paragraph FIXED — hardware sensitivity scenario only; provider-conduct prediction removed; "No probability is assigned"
CAL-9 calibration P1 Direct annex landings still say "80% CI" FIXED — probability-label banner on all six dive wrappers (bodies stay verbatim)
LUR-1 lurker P0 Long selector names cause document-wide horizontal overflow on mobile FIXED — CSS width constraints + wrap
LUR-2 lurker P1 Standalone hero reads as a general margin claim FIXED — hero label/note/diagnostic changes above
LUR-3 lurker P1 Masthead overgeneralized anchor-fitting FIXED (= MLI-6)
LUR-4 lurker P1 "Evidence median" gave an empirical scent to a mixed-evidence scenario FIXED (= CAL-3)
LUR-5 lurker P0 Footer manifest didn't identify the frozen source FIXED (round 2) — a printed commit was not reproducibility (the stamp step dirtied the tree; final gate reopened it). Releases now require a clean worktree, create a release-stamp commit containing the stamped footer plus a served-asset SHA-256 manifest (/asset-manifest.json), and deploy that exact tree

Dispositions after round 2: 45 fixed · 0 deferred · 0 rejected (five rows were overstated at first closing and re-closed under the final gate — the honest history is in each row). Release gates: 499 engine + 220 contract assertions plus a 17-check application-level browser suite encoding the live attack payloads; the audits' own deterministic acceptance checks applied where automatable.

What the audits explicitly could not establish

Corpus completeness (all null results are bounded by the manifest-selected archive); the real people's actual opinions (by design — nothing here is an endorsement); human reading behavior (the lurker run is an information-architecture audit; the 5-human-reader gate remains open); byte-level screenshot geometry (headless captures failed in several runs; DOM/CSS reasoning was used instead).