Snapshot passed all guards: global sanity ($3,500–$120,000), per-reference bounds (derived dynamically from trailing 6-week median ×0.65/×1.35), week-over-week change ≤±30%, and minimum 5 qualifying listings. Clean snapshots are used in RRGI and signal computation.

Snapshot rejected by one or more guards. Dirty rows are stored for audit but excluded from all index and signal calculations.

When Cloudflare blocks the scraper for a reference, no snapshot is written for that week. The reference is absent from that week's RRGI calculation and its weight is redistributed proportionally to constituents with clean data.

The original value, correction date, delta, and cause are written to market_index.notes for every corrected row. No correction is silent — each patch is reported to site administration before deployment.

The June 14, 2026 RRGI was corrected on July 5, 2026. Two separate causes applied: (1) a 28.83 pt upward adjustment from applying the Oyster Perpetual 41 bridge (OP41 was blocked by Cloudflare on that date and its clean anchor prices were not used in the original live computation); (2) a 30.42 pt residual gap that cannot be reproduced from current database records — the original inputs were not preserved. The revised value (1,021.25 pts) is the best recomputable figure under current methodology. The irrecoverable portion is documented in the notes field of that row.

A definition drift introduced July 26, 2026 caused that week's and the following week's (August 2) grey-market range to be recorded as the true low/high of the trimmed listing set rather than a band around the median — the mid computed from it was the midrange of that range, not the trimmed median this page describes. July 26 was reverted to its independently-verified clean value (1,014.29 pts) on August 5, 2026. August 2's true medians cannot be recovered (per-listing data behind that week's read was not yet being persisted), so it currently has no published RRGI reading rather than a restated or spliced one — a genuine gap, consistent with the bridging policy above. It will be filled once the following week's ingest runs on the corrected median basis, using geometric interpolation between the July 26 and next-available anchors, and flagged on the history table like any other bridged row.

The index base (1,000 pts, 7 June 2026) predates the data-quality guard system this page describes: the minimum-5-listings and per-reference-bounds guards could not run on week one — there was no trailing history to derive per-reference bounds from, and listing counts were not yet being captured at all. Under current methodology the base week would not qualify as a clean snapshot; it is grandfathered only because it was inserted before the guard system existed, not because it was evaluated and passed. A read-only model that excludes 7 June and rebases the chain to 14 June (first guard-eligible week) shows every published value from 14 June onward shifting by a near-constant −1.90% — e.g. 9 August 2026: 1,031.40 → 1,011.81 pts. The offset is flat, not compounding: only the first chain link (14 Jun/7 Jun) carries the error, so week-over-week changes, trends, and signals computed from this page's formula are unaffected by it. Decision: the base is not being rebased. The index level is arbitrary by declaration (1,000 pts, by construction) — rebasing would restate every already-published value, including two newsletters and social posts already distributed and not recallable, trading irreversible external inconsistency for a correction to a number that was arbitrary to begin with. This note is the audit trail for that decision; the full weekly model output is versioned in the site repository (model-rebase-14jun.mjs) so the analysis can be reproduced or revisited without redoing it.