Synthesis: Ship/No-Ship Review of inference-margins.pages.dev (methodology v2.1)

All four agents fetched the live page, engine.js, app.js, and the annex; three explicitly confirmed the deployed assets byte-match the local release artifacts [architect, risk-analyst, empiricist]. This is a well-corroborated review — the big findings appear independently in all four outputs with matching exact quotes and matching numeric endpoints, which raises my confidence substantially.

1. Consensus

• The core arithmetic engine is sound — this is a trust-layer problem, not an arithmetic rollback. 452 release assertions pass; five of six §10 blended headlines reproduce within one point (OpenAI ~94.7%, Google ~95.9%, xAI ~66.9%, DeepSeek ~68.1%, Zhipu ~60.0%). [unanimous]
• Cache-write accounting is conservation-safe and default-off — total write tariff, partitioned fresh input, prefill charged once, v2 baselines preserved. The prior cache-contamination bug is genuinely fixed. [unanimous] (Cache double-count is the one attack-list item all four score PASS.)
• The anti-shopping lens range is broken and does the opposite of what it advertises. lensRangeForCurrentModel() unions every non-replay perspective — including [analyst] positions and pairing-warned, scope-incompatible lenses — ignoring the compatibility warnings generated elsewhere. All four independently reproduced the same absurd endpoints: Opus −79.3% to 97.1%, Grok −337.4% to 92.8%, DeepSeek V4 −867.5% to 85.8%. [unanimous]
• The quoted-position attribution line contradicts its own tables. The rendered sentence "the parameters below come from the source's own disclosure" is shown for xaiopp, chinacloud, and anth20, whose tables label utilization, stack, batch share, discounts, etc. as SPECULATION. This is textbook attribution laundering in the mechanism built to prevent it. [unanimous]
• Both v2 §10 language commitments regressed. Ranking language survives ("The highest central margin on the page", "highest raw headline among the three Chinese-provider cards") and confidence-interval language survives ("margin confidence intervals", "the widest confidence interval on the page") — directly contradicting the page's own "not statistically calibrated confidence intervals." Card badges are correctly labeled "judgmental range," but "everywhere" is not met. [unanimous]
• The advertised "runnable diagnostic" is not deployed. /tests/roofline-diagnostic.mjs (and /tests/snapshots.test.mjs) return the HTML homepage with content-type: text/html. The local diagnostic's numbers are accurate (H20 −11.4/−16.2%, GB200 −1.9%, Ascend@50ms −9.8%, Ascend@15ms −39.8%), but external readers cannot run the artifact the methods note names. [unanimous]
• Permalinks are schema-tagged, not reproducible. They encode only diffs from mutable DEFAULTS, carry no engine/data hash, and discard _meta. The versionless-URL attack is not actually pre-empted. [unanimous] (framing disagreement below).
• The typed ontology is cosmetic. ioRatio/cacheHit/blend remain MODEL_OWNED_KEYS; [lens]/[analyst]/[replay] are labels on the old perspective axis, not a refactored model/workload/SLA/lens separation. Changing "model" still silently changes workload and procurement. [skeptic, risk-analyst, empiricist; architect treats it as the structural root of P1.1]
• The China 6.15× scalar is arithmetically exact but is a scenario scalar, not a coherent observed cross-SKU on-demand multiplier (H800 matches ~$10.6; H20 → $6.15 vs clean ~$4.48 observation; 910C has no reproducible public retail). P2 — caveats prevent it from being an undisclosed error. [unanimous]
• The explanatory surfaces compete, and the problem is not verbosity — it's that the layers disagree. The 29% vs 36.5% mismatch, stale CI language, and ranking regressions are symptoms of duplicated truth across methods box / preset note / dossier / §10 cards. [architect, risk-analyst, empiricist; skeptic frames it as "over-explained and under-specified where it matters"]

Attack-list scorecard — the four converge: attribution laundering FAIL, model/workload conflation FAIL, cache double-count PASS, versionless URLs PARTIAL/FAIL, dossier decoration FAIL, coverage theater FAIL/PARTIAL.

2. Disagreements

A. Is anything a P0? (Severity taxonomy)

The disagreement: Skeptic calls the xAI lens comparison and the anti-shopping range two P0s; Architect, Risk Analyst, and Empiricist find no P0s and classify both as P1.

• Skeptic's strongest argument: The anti-shopping range "converts invalid pairings into the most prominent uncertainty display… worse than omitting the range because it gives nonsense a methodological imprimatur." A flagship trust feature that actively misleads about the product's headline guarantee is not a P1 wording fix — it requires re-architecting what enters the statistic.
• The other three's strongest argument: P0 should mean arithmetic corruption. The engine math is correct in every case; these are presentation/provenance defects layered on a sound core. Empiricist: "no P0 arithmetic corruption."

Adjudication: Largely definitional and operationally moot — all four block ship on both findings regardless of label. But Skeptic's severity instinct is directionally right on one point the taxonomy obscures: the anti-shopping-range fix is not a lint, it's a re-architecture of range membership. I'd record both as ship-blockers and not litigate the number. Where the label genuinely matters is the xAI case below.

B. What is the correct xAI opportunity-cost number — ~27% or 36.5%? (The sharpest genuine disagreement)

The disagreement: Skeptic says 36.52% is a bug artifact and the true controlled value is ~27%; the other three say 36.52% is the legitimate calculator output and 29% is merely stale prose to be relabeled.

• Skeptic's strongest argument: The page advertises "Same operating point as the dive; only the valuation changes." But the cash/opportunity lens presets omit ioRatio and cacheHit, so the preset merge falls back to the global 15:1 / 60% cache defaults — a different operating point than the dive's 3:1 / 0% cache. Holding the dive workload genuinely fixed, Skeptic recomputes the opportunity lens at 26.95%, near the dive's ~29% — with no "model-shape divergence" needed. So neither 36.5% nor the "model shape" explanation is a controlled result; the ~10-point gap is a workload-merge bug.
• The other three's strongest argument (Architect's is cleanest): The deployed calculator genuinely emits 92.77% → 66.94% → 36.52%; the 29% belongs to the external dive's coarser replica model. The defect is a 7.5-point inconsistency — the methods box and Grok dossier still print 29% as if it were the calculator's third lens. Fix = harmonize the prose to "~93% → 67% → 36.5% calculator; ~29% annex replica."

Adjudication: Skeptic has the stronger case, and it modifies the Architect's fix. Skeptic did the controlled recomputation the other three did not — they verified the engine emits 36.52% (true) but not whether 36.52% is computed at the same operating point the page claims. It isn't. Architect's P1.3 fix (relabel 29% → 36.5%) does not survive Skeptic's critique: it would harmonize the prose onto a number that itself violates the page's "same operating point, only valuation changes" claim. The correct sequence is Skeptic's: fix the workload merge so all three lenses inherit the dive's 3:1/0% operating point, recompute (expect opportunity ≈ 27%), then harmonize the prose. Residual uncertainty: this assumes "same operating point" is the intended semantic. If the author intended the lenses to carry their own default workloads, 36.5% is legitimate and only the prose needs fixing — but then the "only the valuation changes" claim must be struck. Either way there is an inconsistency to fix; the magnitude of the required fix depends on author intent (Open Question C).

C. Permalinks: version-drift (future) vs identity-corruption (now)

The disagreement: Skeptic/Architect frame the permalink defect primarily as future drift (a later DEFAULTS change silently reinterprets old links); Empiricist and Risk Analyst find a same-deployment, right-now corruption.

• Version-drift framing (Skeptic, Architect): No engineRevision/snapshot hash; diffs merge into whatever defaults exist later. Architect explicitly offers a downgrade: "if shared links are explicitly guaranteed same-deployment-only… these findings could drop to P2."
• Identity-corruption framing (Empiricist, Risk Analyst): The URL serializes only numeric state — not modelId/perspectiveId/scope/provenance. On load, selectors stay Opus/median, the dossier doesn't render, and the lens range is computed for the selector's model. So a Grok+xAI-opportunity permalink displays Grok arithmetic while its selectors and anti-shopping range refer to Opus — misattribution within the same deployment.

Adjudication: Empiricist/Risk have the stronger, more specific finding, and it invalidates Architect's downgrade condition. "Same-deployment-only" does not rescue the permalink, because the identity corruption manifests on the current deployment. The fix must serialize full state (model/perspective IDs + scope + engine/data hash) and restore selectors/dossier/warning/range together — not merely add a version hash. This is a cross-check where the Risk/Empiricist failure mode is not addressed by the Architect's proposed fix — flagged.

D. Are tariff point-estimates and the ontology refactor ship-blockers?

The disagreement: Skeptic lists tariff-as-point-estimate as P1 ("not genuinely range-first"); Empiricist calls tariff labeling "mostly good" (P2-ish); Architect scopes the full workload/model ontology split to a v2.2 refactor, explicitly inappropriate as an emergency gate.

Adjudication: Architect's staging is right and I weight it. The labeling (names, speculative marks, table exclusion) passes for all four; the residual concern (point margins from arbitrary architecture midpoints) is real but is a symptom of the same MODEL_OWNED ontology. Correct call: a cheap mitigation now (suppress/qualify the tariff point-margin headline, e.g. "illustrative architecture scenario" in the result tile) + defer the true workload/SLA/model separation to v2.2. Do not gate v2.1.1 on the full refactor.

3. Open questions

• Are permalinks intended to be same-deployment-only? [architect raised as a downgrade condition] — but Empiricist's identity-corruption finding suggests even that wouldn't save them. Needs the author's design intent.
• Is the roofline diagnostic intentionally private? [architect] If yes, the fix collapses to deleting the word "runnable"; if no, deploy it. Cheap either way.
• Is "same operating point, only the valuation changes" the intended semantic for the xAI lenses? [synthesis-emergent, from Disagreement B] This single answer determines whether the correct opportunity value is ~27% (fix the merge) or 36.5% (fix only the prose and strike the "same operating point" claim). Only the author can settle it.
• Does any genuinely untouched roofline holdout exist? [risk-analyst, empiricist] Both conclude no — zero residual DoF, post-hoc operating-point choices, no untouched platform. This bounds how strong the roofline "validation" claim can ever be.
• Tencent H20/H800 and the 910C-proxy rates are unverifiable from the annex (no primary URLs) [empiricist]. Downgrades confidence in the China multiplier evidence specifically.

4. Final recommendation

NO-SHIP / hold v2.1 as deployed. Ship after a surgical v2.1.1 gate repair — no engine rollback. (Risk Analyst's "PAUSE, fix the P1s first" is operationally the same verdict; the four are effectively unanimous.)

Confidence:

• Very high (~95%) that it should not ship as-is. This is not a judgment call — four independent reviews on a byte-matched deployment corroborate the three load-bearing failures (anti-shopping range, attribution laundering, §10 regressions) with matching exact quotes and matching numeric endpoints.
• High that the surgical fix set below is complete and correct, and that the core numeric engine needs no changes.
• Medium on the exact xAI opportunity number (~27% vs 36.5%) — it hinges on an author-only design intent (Open Question C), and I weight Skeptic's controlled recomputation but cannot close the intent question from the outside.
• Lower on whether tariff point-estimates / the ontology refactor are gate items — a genuine judgment split; I side with Architect's v2.2 staging.

Ship gate (deduped across all four):

1. Anti-shopping range — compute only over compatible kind === "lens" entries passing pairingWarning() === ""; exclude analyst positions and replays; move exploratory/incompatible pairings to a separately labeled range. Assert contributor names. [unanimous]
2. xAI lenses — make all three inherit the dive operating point (3:1, 0% cache); recompute (expect opportunity ≈ 27%); then harmonize the methods box and Grok dossier. Do not merely relabel 29% → 36.5% (Skeptic's correction to Architect's fix). [skeptic-weighted]
3. Dossiers — reclassify xaiopp/chinacloud/anth20 from quoted-position to reconstruction (or split "quoted anchor" from "reconstructed vector"); annotate every dynamic dive/set override; make displayed values derive from final applied state. [unanimous]
4. §10 language — remove ranking phrases and remaining "confidence interval" language; add a static lint (highest|lowest|best|worst|confidence interval|\bCI\b). [unanimous]
5. Roofline diagnostic — deploy the script + frozen inputs + snapshot test as real artifacts; assert content-type/hash in release. [unanimous]
6. Permalinks — serialize full state incl. model/perspective IDs + scope + engine/data hash; validate on load; fix the same-deployment identity corruption first, then version-safety; add round-trip + migration fixtures. [unanimous, empiricist/risk-sharpened]
7. Moonshot/Kimi — expose an output-token-margin tile so the headline literally regenerates, or drop the "every headline reproducible" claim. [architect, risk-analyst, empiricist]

Include in the same release (P2s that prove the P1 fixes held): the dossier DOM test (assert rendered values, not key membership) and engine-generated bridge/test-count numbers (fixes the "29 assertions" → 452 and the "$3.10/Mtok" → $3.72-list staleness Risk Analyst uniquely caught). Defer: China per-SKU multiplier rename, roofline "preregistered holdout could not be executed as designed" phrasing, and the full workload/SLA/model ontology split (v2.2).

Conditions that would change the recommendation:

• If the diagnostic is intentionally private → item 5 collapses to deleting "runnable."
• If "same operating point" is not the intended xAI-lens semantic → item 2 shrinks to a prose fix (relabel 29% → 36.5%) plus striking the "only the valuation changes" claim.
• Note the one downgrade that does not hold: Architect's "if permalinks are same-deployment-only, drop to P2" is defeated by Empiricist's identity-corruption finding — do not rely on it.

Cheapest next action that moves the decision forward (Empiricist's, and I weight it heavily): write the ~1-hour semantic release test — four fixtures + §10 lint that catch every ship-blocker above:

1. Serialize+reload Grok+xAI-opportunity; assert model/perspective IDs, dossier, warning, margin, and range all survive.
2. Assert no pairingWarning()-triggering pairing contributes to any lens range.
3. Assert every quoted-position parameter is genuinely source-selected, else fail classification.
4. Fetch the public roofline URL; require JS content + expected output.
5. Static §10 lint for ranking/CI terms.

This test file is the regression gate that would have caught all of these pre-deploy — build it first; it converts the fix list from "trust the patch" into "trust the assertion."

5. Attribution map