GPUDECK · CORRECTIONS REGISTER · APPEND-ONLY

When the record was wrong, it says so here.

A frozen edition is never edited. When an error is found in one, a correction record is appended here, pointing backward at the affected issue with its severity, the affected file, and the digests before and after. Records are never removed or rewritten — a wrong correction gets its own correcting record. Machine-readable: corrections.json.

2 correction(s) on record · page generated 2026-09-23 05:21 UTC

TYPOwording or formatting; no figure or file changed
METADATAlabels, digests, or packaging; dataset rows unchanged
DATApublished rows or cells were wrong
METHODOLOGYa rule error - affected figures need re-reading under the corrected rule
RIGHTSGPUDeck lacked the right to publish it - the figures are withdrawn, not restated
C-2026-08-26-002 RIGHTS discovered 2026-08-26

The /prices page carried the banner 'PRICE CONTEXT - PERMISSIONED SOURCES, NEVER SCRAPED' directly above medians derived from the CC BY 4.0 GPU Price Tracker. The licence is real, but the dataset's own published documentation states it is 'collected twice daily since January 2026 by scraping provider pricing surfaces via the gpuhunt library', across 13 providers including RunPod, Vast.ai, TensorDock and CloudRift - four providers whose published terms GPUDeck has itself verified prohibit exactly that. A licence conveys only the rights the licensor actually held, so the word 'permissioned' was not GPUDeck's to use for that data and the phrase 'never scraped' was false as published. GPUDeck had treated the licence as settling provenance and had never read the dataset card to see how the licensor obtained the data. The series is WITHDRAWN rather than relabelled: a chain-of-title filter now removes every upstream lacking its own recorded basis before any aggregation runs, so no median, floor or count is computed over a row GPUDeck may not publish. Each provider returns only on archived evidence of an authorised collection route or a completed provider-specific analysis concluding no permission is required. The marketplace providers are not expected back. Found by reading the source dataset's own card while implementing METHODOLOGY v5, after Codex identified the chain-of-title gap in the round-2 settlement and GPUDeck recorded it as an open item rather than resolving it - recording it was not enough.

Affected editionissue 253
Affected fileprices (multi-provider tracker series) and its export gpu-price-context.csv
Affected rows/cellsthe entire licensed-tracker series: every marketplace and cloud median, floor, trend and sample count derived from it
Corrected in issue254
C-2026-08-26-001 METADATA discovered 2026-08-26

The first freeze of issue 252 wrote the edition page BEFORE the site-wide analytics-beacon transform ran, so the bytes later served differed from the frozen file, and no SHA-256 had yet been published in the register. The edition was re-frozen the same morning POST-transform, the register began publishing the full digest of the served bytes, and a staging gate now fails any build whose staged edition differs byte-for-byte from its frozen source. Found by the Codex round-4 audit before any digest was published or cited.

digest of the pre-repair frozen source file (git blob 781f442); the as-served pre-repair variant carried an injected beacon and its digest was never recorded - that unverifiability was the defect

Affected editionissue 252
Affected fileissues/issue-252-2026-08-25/index.html
Affected rows/cellsedition page bytes only; no dataset rows affected
Corrected in issue252
Old SHA-25625f67f3ec27661ef0b32c57840dc833b4c7782593a1db9efa3e955b47f965471
Replacement SHA-256d319d636e1f4b06510137799c1790188e5efc0afb3468422ea43298bf6c0e509
Methodology impactnone - no published figure or rule changed; the freeze procedure gained a byte-identity enforcement gate

Disputing a figure

If you operate a provider named on this site and believe a published figure misrepresents you, that has its own register with its own response clock: /disputes. This page records GPUDeck’s own errors; that one records challenges from outside, and states what happens to a figure while a credible dispute is open.

The protocol

Why not just fix the file? Because a citation to issue N must mean the same bytes forever. An edition that can be quietly repaired is an edition that cannot be cited. The fix lives in a later issue; the record of the error lives here.

What links where. Each record names the issue it corrects and the issue that carries the corrected data. The affected issue's own diff.json lists the correction id from the issue that corrects it onward — frozen diffs are never rewritten either.

Severity is a reading instruction. TYPO and METADATA leave every published figure standing. DATA means specific rows were wrong — the record names them. METHODOLOGY means a rule was wrong: re-read affected figures under the dated correction in METHODOLOGY.md.