Corrections log

On 8 and 9 October 2026 we withdrew 919 change records that had been first published between 18 April and 7 October 2026. They had been public for a median of 108 days (between 1 and 174).

Each recorded a difference between two captures of a page that lay in page furniture, not in the document: navigation menus, footers, page titles, language pickers, related-article lists, embedded page data, live counters and clocks. Our text extractor stopped reading these in two steps: navigation and footer stripping, deployed 10 September 2026 (890 of the 919 were first published before it), and extractor version 5, deployed 7 October 2026. Version 5 also drops page titles, video fallback text, menus built as select lists, link labels, loading placeholders and related-post lists, and folds invisible characters and typographic quotes. Re-extracted with version 5, the two captures behind each record contain the same text.

Each record has its own entry in this log, quoting what it described. Its page stays readable as a withdrawn record.

This change record described a change to Microsoft's Privacy Statement that Microsoft never made. On Mar 5, 2026 we stored two copies of the Privacy Statement that we had edited ourselves, to test our summarisation pipeline, as if they were captures of Microsoft's page. This record compared Microsoft's real page against that edited copy (version 1637), so what it described as Microsoft's change was our own edit appearing or disappearing. The edited copies are now quarantined and are not evidence of anything Microsoft published. Measured Oct 5, 2026: the record's later version rests on test snapshot 6, whose stored note reads "Test injection v2".

This change record described a change to Microsoft's Privacy Statement that Microsoft never made. On Mar 5, 2026 we stored two copies of the Privacy Statement that we had edited ourselves, to test our summarisation pipeline, as if they were captures of Microsoft's page. This record compared that edited copy against Microsoft's real page (version 1637), so what it described as Microsoft's change was our own edit appearing or disappearing. The edited copies are now quarantined and are not evidence of anything Microsoft published. Measured Oct 5, 2026: the record's earlier version rests on test snapshot 6, whose stored note reads "Test injection v2".

This change record described a change to Microsoft's Privacy Statement that Microsoft never made. On Mar 5, 2026 we stored two copies of the Privacy Statement that we had edited ourselves, to test our summarisation pipeline, as if they were captures of Microsoft's page. This record compared that edited copy against Microsoft's real page (version 1635), so what it described as Microsoft's change was our own edit appearing or disappearing. The edited copies are now quarantined and are not evidence of anything Microsoft published. Measured Oct 5, 2026: the record's earlier version rests on test snapshot 5, whose stored note reads "Test injection for summarization pipeline testing".

We withdrew 2 of the 31 clauses we published for this document. They were extracted from the capture of 2026-09-11 (version 6758), which our own guards had refused: it was stored as evidence but never published as a version of this document. A defect in the extraction step read 'the active version, or the newest capture', so when nothing was published it mined the refused capture instead and presented the result as this document's clauses. The fallback was removed on Sep 30, 2026; extraction now refuses to run when there is no accepted version. The remaining 29 clause(s), which were mined from captures we accepted, are unaffected and stay published. Measured Sep 30, 2026: a clause was counted here when its quoted text is absent from the version this page publishes and present in that refused capture.

We withdrew 9 of the 106 clauses we published for this document. They were extracted from the capture of 2026-09-11 (version 6707), which our own guards had refused: it was stored as evidence but never published as a version of this document. A defect in the extraction step read 'the active version, or the newest capture', so when nothing was published it mined the refused capture instead and presented the result as this document's clauses. The fallback was removed on Sep 30, 2026; extraction now refuses to run when there is no accepted version. The remaining 97 clause(s), which were mined from captures we accepted, are unaffected and stay published. Measured Sep 30, 2026: a clause was counted here when its quoted text is absent from the version this page publishes and present in that refused capture.

We withdrew 14 of the 49 clauses we published for this document. They were extracted from the capture of 2026-08-15 (version 5831), which our own guards had refused: it was stored as evidence but never published as a version of this document. A defect in the extraction step read 'the active version, or the newest capture', so when nothing was published it mined the refused capture instead and presented the result as this document's clauses. The fallback was removed on Sep 30, 2026; extraction now refuses to run when there is no accepted version. The remaining 35 clause(s), which were mined from captures we accepted, are unaffected and stay published. Measured Sep 30, 2026: a clause was counted here when its quoted text is absent from the version this page publishes and present in that refused capture.

We withdrew every clause we published for this document. They were extracted from the capture of 2026-09-14 (version 6941), which our own guards had refused: it was stored as evidence but never published as a version of this document. A defect in the extraction step read 'the active version, or the newest capture', so when nothing was published it mined the refused capture instead and presented the result as this document's clauses. The fallback was removed on Sep 30, 2026; extraction now refuses to run when there is no accepted version. No clauses remain on this document's page. We would rather show none than show clauses we cannot stand behind, so the clause surface is empty until a capture we accept can be re-extracted. Measured Sep 30, 2026: a clause was counted here when its quoted text is absent from the version this page publishes and present in that refused capture.

From Sep 9, 2026 to Sep 28, 2026 (19 days), this page published a 290-byte strip of site navigation as the Target Terms and Conditions. It was not an older version of the policy; it was not the policy at all, and no part of the Target Terms and Conditions was on this page for that period. The version restored in its place is the most recent capture we hold that IS the policy, taken 32 day(s) before the bad capture: that gap is currency lost by the restore, a separate cost from the wrong content. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The correct version was restored by hand on Sep 28, 2026.

From Jul 18, 2026 to Sep 28, 2026 (70 days), this page published about 900 bytes of raw page markup (a doctype and head tags) as the Postman Privacy Policy. It was not an older version of the policy; it was not the policy at all, and no part of the Postman Privacy Policy was on this page for that period. The version restored in its place was captured 1 day(s) after the bad one, so the restore cost no currency: it is the most recent capture we hold, and it is the policy. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The correct version was restored by hand on Sep 28, 2026.

From Jul 18, 2026 to Sep 28, 2026 (70 days), this page published about 900 bytes of raw page markup (a doctype and head tags) as the Patreon Privacy Policy. It was not an older version of the policy; it was not the policy at all, and no part of the Patreon Privacy Policy was on this page for that period. The version restored in its place was captured 1 day(s) after the bad one, so the restore cost no currency: it is the most recent capture we hold, and it is the policy. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The correct version was restored by hand on Sep 28, 2026.

For 3 nights, this page published the version captured Sep 12, 2026 as the current Uniswap Privacy Policy, while our own nightly capture was already fetching the version of Sep 17, 2026. Two numbers describe this and they measure different things: the page was wrong for 3 night(s), and the version it showed was 4 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 30, 2026.

For 6 nights, this page published the version captured Sep 1, 2026 as the current Uniswap Terms of Service, while our own nightly capture was already fetching the version of Sep 16, 2026. Two numbers describe this and they measure different things: the page was wrong for 6 night(s), and the version it showed was 15 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 30, 2026.

For 18 nights, this page published the version captured May 5, 2026 as the current TikTok Advertising Terms, while our own nightly capture was already fetching the version of Sep 10, 2026. Two numbers describe this and they measure different things: the page was wrong for 18 night(s), and the version it showed was 127 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 24 nights, this page published the version captured Jul 14, 2026 as the current Google Privacy Policy, while our own nightly capture was already fetching the version of Jul 30, 2026. Two numbers describe this and they measure different things: the page was wrong for 24 night(s), and the version it showed was 15 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 24 nights, this page published the version captured Jun 12, 2026 as the current Google Generative AI Prohibited Use Policy, while our own nightly capture was already fetching the version of Jul 30, 2026. Two numbers describe this and they measure different things: the page was wrong for 24 night(s), and the version it showed was 48 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 24 nights, this page published the version captured Jun 12, 2026 as the current Google Generative AI Terms, while our own nightly capture was already fetching the version of Jul 30, 2026. Two numbers describe this and they measure different things: the page was wrong for 24 night(s), and the version it showed was 47 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 35 nights, this page published the version captured Jun 6, 2026 as the current OpenAI Data Processing Addendum, while our own nightly capture was already fetching the version of Aug 20, 2026. Two numbers describe this and they measure different things: the page was wrong for 35 night(s), and the version it showed was 74 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 169 nights, this page published the version captured Apr 18, 2026 as the current Duolingo Terms of Service, while our own nightly capture was already fetching the version of May 5, 2026. Two numbers describe this and they measure different things: the page was wrong for 169 night(s), and the version it showed was 17 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 216 nights, this page published the version captured Apr 18, 2026 as the current Netflix Terms of Use, while our own nightly capture was already fetching the version of Apr 19, 2026. Two numbers describe this and they measure different things: the page was wrong for 216 night(s), and the version it showed was 0 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. The page corrected itself on the first capture after that fix, on Sep 29, 2026.

For 216 nights, this page published the version captured Apr 18, 2026 as the current Netflix Account and Content Policies, while our own nightly capture was already fetching the version of Apr 19, 2026. Two numbers describe this and they measure different things: the page was wrong for 216 night(s), and the version it showed was 0 day(s) older than the one the site was fetching. Neither number alone is the whole picture — a long displacement of a barely-older version is a smaller error than a short displacement of a much older one. The cause was in capture: it decided whether a page had changed by comparing against the most recently captured version rather than the one being published, so a page we were re-fetching correctly every night never replaced what was on the site. That comparison now uses the published version. This one is NOT yet corrected: the page still publishes the same version. Its capture now quarantines on an attribution conflict with the other document we hold at the same URL, which is a separate open issue. This entry will be updated when it is resolved.

We have withheld this change record, but we have not established that it is wrong. The raw capture it was based on was stored on a server that has since been decommissioned, so we can no longer check it against the source. It is withheld until it can be verified, and it may be accurate.

We have withheld this change record, but we have not established that it is wrong. The raw capture it was based on was stored on a server that has since been decommissioned, so we can no longer check it against the source. It is withheld until it can be verified, and it may be accurate.

We have withheld this change record, but we have not established that it is wrong. The raw capture it was based on was stored on a server that has since been decommissioned, so we can no longer check it against the source. It is withheld until it can be verified, and it may be accurate.

We withdrew this change record because it duplicated one we had already published: the same change, at the same address, on the same day, recorded a second time because we were monitoring the same address under two source records. The duplicate source record has since been retired. The original record remains published: https://conductatlas.com/change/2026-09-12-anthropic-anthropic-privacy-policy-5047/

We withdrew this change record because it duplicated one we had already published: the same change, at the same address, on the same day, recorded a second time because we were monitoring the same address under two source records. The duplicate source record has since been retired. The original record remains published: https://conductatlas.com/change/2026-07-09-anthropic-anthropic-privacy-policy-3556/