The platform has rules.

The methodology has a hash. The hash does not change without a new model family.

The same problem Berg-Koelbel-Rigobon (2020) documented in ESG ratings — six major raters reaching dramatically different conclusions because their methodology drifted silently — is now happening at scale on AI platforms. Model weights update without notice, system prompts change between API calls, evaluation benchmarks rotate without versioning, and users have no way to verify whether the AI giving them an answer today is the AI they evaluated last quarter. Aequara publishes the SHA-256 of every active methodology and every routed model configuration. Operational amendments are versioned without rotating the hash. Substantive changes — model swaps, weight rebalancing, signal additions — require declaring a new family and a parallel publication cycle. The old methodology continues for the deprecation horizon, so users never lose continuity.

NPX-100 v2.1 · methodology hash cfe274c9... · operational amendment, not methodology change.

Aequara accepts no revenue from rated entities.

The conflict of interest that destroyed credit-rating credibility in 2008 — issuers paying the agencies to rate them — is structurally excluded from Aequara's revenue. Subscription, licensing, and Trustee Tier programs only. Rated firms cannot purchase, bypass, expedite, or influence their inclusion, weighting, or removal. The firewall is enforced through three mechanisms: (1) a sequestered NPX-BL index family for Aequara internal entities; (2) a public revenue-source disclosure published quarterly alongside index publication; (3) an external auditor attestation reviewed annually.

Every AI prediction is scored. Every error is published.

The Brier score (Glenn Brier, 1950) is the oldest and most honest calibration metric: mean squared error between forecast probability and realized outcome. Aequara registers every AI output as a Brier prediction with a stated horizon and confidence band — across every routed model, every steward, every product surface. At horizon resolution, the prediction is scored, the error is published, the running cohort calibration is updated, and any drift across model versions is flagged. We earn trust by making mistakes legible — the alternative is the modern AI-platform model where "confidence" is a vibe, evals are marketing, and nobody publishes what the system got wrong last month.

NPX-100 live cohort: Brier 0.143 · n=247 · calibration 0.91 · published 2026-05-10.

The DJIA has been wrong about the right things for 130 years.

The Dow Jones Industrial Average is methodologically inferior to nearly every modern index — price-weighted, 30 constituents, hand-curated. It has also been wrong about what counts as an "industrial" for a century. And it remains the most-cited index in financial journalism. Why: brand coherence over time beats methodology purity in any given year. Aequara optimizes for the methodology-hash stability of §1 over the methodology-perfection of any one moment. We will be wrong about the right things on purpose, slowly, in public.

Headline N=100. Operational band 97–105.

Per the indexation precedent set by S&P (S&P 500 ≠ exactly 500) and Russell (Russell 1000, 2000, 3000 are all loose-band): a tight headline integer with an operational band reduces forced-trade churn at the boundary. Constituents enter and exit on the band, not on the integer. NPX-100 publishes headline at 100, operates at 97–105, and discloses the active count alongside every cycle.

The first Tuesday of February, May, August, and November.

Russell, Fortune, S&P — every durable index has a publication ritual that doesn't move. The ritual creates a shared expectation; the shared expectation is half of an index's brand. Aequara publishes on the first Tuesday of February, May, August, and November, with the prior cycle's Brier scoring, the methodology hash, the constituent reconciliation, the revenue-source disclosure, and a sealed methodology-amendment ledger. The ritual does not move. Holidays adjust to the next business day; nothing else.

At least three model raters and one human reviewer per call.

Single-rater calls are the path to drift, blindspots, and the Diether-Malloy-Scherbina (2002) divergence trap. Every Aequara verdict is composed from at least three model raters (selected for independence by Jaccard input-overlap) plus one human steward who carries personal accountability. Multi-rater agreement is decomposed into strong confluence (independence ≥ 0.7, signals genuinely independent) versus weak confluence (correlated agreement = echo, treated as a single signal). Divergent panels — when raters genuinely disagree — are flagged as high-information setups requiring research, not consensus suppression.

9-state confluence matrix · Geman-Hwang 1983 Bayesian compounding · independence ≥ 0.7 floor.