A number is only worth citing if you know who can change it, under what rules, and what happens when it's wrong. The methodology page shows how the number is made; this page shows how the making is governed. Everything below is already in force — each rule links to the place where you can watch it operate.
The method is versioned, and every change ships with a date and a reason in the public version history. There are no silent revisions: the current method text is hashed, and that hash is submitted to the independent OpenTimestamps calendars (servers we don't control), so even our own yardstick can't move quietly — check it at the anchored root. When a change would alter a published figure, the old figure stays on the record with its date rather than being overwritten.
The index's forward calls are graded by a single written rule that was sealed — hashed and dated — before any verdict matures. Nobody, including us, gets to pick a friendlier grading standard after seeing the results; the rule's hash is on the record, and the first verdicts publish from December 2026. Every resolved call publishes — hits, misses, and unresolved alike — in the Journal. A call that matures without an observable record resolves as unresolved and counts in no accuracy rate: absence of evidence is not a grade.
When we find an error — and we have — the correction is published with its date in the open log, the wrong figure is retracted rather than quietly replaced, and the sealed record keeps the original because the ledger is append-only. The largest correction in the index's history (a retracted headline ratio, July 2026) is documented there in full. If a live figure fails review, the surfaces show the method and a hold notice instead of a number: silence over a wrong figure, always.
Every published number is derived from its source data by code, from a single canonical definition — the same figure, the same wording, the same caveat on every surface, served from one place. There is no editorial step in which a person types or adjusts a published figure: figures change only when their source data or the versioned derivation changes. When a figure and its source file disagree, the automated integrity checks flag it and the figure goes on hold until the derivation is fixed.
The index is operated by an automated system: collection from primary sources, scoring, sealing, self-auditing, and the freshness and integrity watchdogs all run without a person in the loop, and the system's internal checks run against every code change before it deploys. Humans act at enumerated gates only — irreversible public steps (a regulatory filing, an outward communication) require a named human decision, by design rather than by bottleneck. The result is the property a reference needs: the number cannot have a bad day. It is computed the same way at 3 a.m. as at noon, and it does not depend on any one person being at a desk.
LOADSTAR sells derived analysis. It has never accepted payment, sponsorship, or data-in-kind from any entity it measures, and there is no mechanism by which a score can be purchased or negotiated — scores are computed from the public record by the pipeline above, which has no input labeled "customer". The index carries no advertising. We publish under the organization's name rather than personal bylines so the authority rests on the verifiable record, not on a personality.
The proofs are built to outlive the operator. Sealed predictions verify by recomputing their hashes from the published fields; the ledger root's external timestamp lives on infrastructure we don't control. If LOADSTAR stopped operating tomorrow, every sealed call, every date, and every correction already on the record would remain independently checkable — no cooperation from us required. Data and the sealed record are backed up off the production system on a fixed schedule.
Found an error, a broken derivation, or a rule above not holding? That report is worth more to us than a compliment — tell us. Corrections land in the public log, dated.