Inputs are unadjusted
Prices are stored raw. Splits, dividends, and other corporate actions are applied by the calculation engine at computation time, per the index’s methodology. Adjusted-price inputs are refused because they encode CA choices that cannot be inverted without loss.
Computation is deterministic
Given the same inputs, the engine produces byte-identical outputs. Randomness (Math.random, unseeded shuffles, floating-point order-of-summation) is refused inside the engine. Every price-level, weight, and divisor is derivable from the recorded inputs.
Each level is computed twice, by two independent paths (divisor-scaling and position-scaling). Publication requires the two paths agree to 1e-9 at every point in the series. Disagreement fails the pipeline before anything is written.
Published levels are append-only
Every published level row is enforced immutable by a database trigger — UPDATE and DELETE are rejected outright. The only permitted change on a published row is setting the superseded_by pointer, once, when the row is retired by a restatement. The pointer is a link forward, not a rewrite; the original value stays intact.
Corrections restate; extensions do not
A restatement publishes a new version of the series with a mandatory reason, a per-date diff (old / new / delta for every changed date, plus added and removed dates), and a fresh hash chain. Prior versions remain fully retrievable, forever, at every date.
Version numbers only tick on corrections. A daily catch-up that adds new EOD sessions without changing any existing value extends the current version in place; there is no version bump on Tuesdays. This keeps the version count meaningful: v6 means five distinct corrections have been recorded.
The audit trail is hash-chained
Every daily bundle is fingerprinted (SHA-256 of the canonical levels CSV) and folded into a chain hash covering every prior entry, from a fixed genesis string. Any tampering — a single altered digit, a reordered row, a removed line — breaks all subsequent chain hashes. A reader holding only the current chain head can detect any modification to any historical bundle.
Restatements start a new chain, anchored at the new version. The prior chain is archived alongside the retired version so it can be re-verified in perpetuity.
The publication row is write-once
The index_publications row that records a version’s reason, inputs hash, and publication timestamp is also enforced immutable after INSERT. The only change permitted is setting the row’s superseded_by_version pointer, once, when the version is later superseded by a restatement. Reasons cannot be rewritten. Timestamps cannot drift. Inputs hashes cannot be swapped.
Provisional-date gate
The engine publishes a date only when every constituent has a bar at or after that date. Constituents on markets that have not yet closed for the session are legitimately incomplete; publishing them as if they were finalized invites a same-week rewrite. The gate produces the cutoff; anything beyond it is deferred, and shows up in the next run.
Per-index methodology documents
Each index carries its own methodology document, published by the index’s owner, covering eligibility, weighting, currency handling, rebalance and reconstitution calendars, corporate-action treatment, discretion policy, and complaint channels. BenchmarkRegistry does not sponsor or edit those documents; the registry publishes what the engine produces from the methodology, and links to the methodology as the definitive source.
Alignment
The publication discipline described on this page is aligned with IOSCO Principles for Financial Benchmarks in the areas of methodology documentation, integrity of the calculation process, transparency of restatements, and record-keeping. This is a statement of alignment, not a claim of compliance.