Changelog

CHANGELOG.md

@@ -1,21 +1,26 @@
-# Releases
-
-- v1 (2026-10-28): 31 subjects, 29754 claims, 0 verdict changes; method 0.1.0+4bff4432, 0.1.0+ecd0c1d5.
-- v1 README corrected (30 Oct 2026): the stability line printed "? of ?" and now gives the measured figures (152 of 164 on a repeated run, 102 of 109 researched again from scratch); the caveat that every report was made through Claude Code on a subscription plan, with the API agreement check (276 of 310), was missing and is added. The data files are unchanged.
-- Known defects in v1, found 3 Nov 2026 and fixed in the stored reports for v1.1 (v1's files are unchanged): a proposition naming where a subject lives, and a joke about a subject's own death scored as a prediction (both removed, with 50 more claims about private life); six verdicts given before the date their criteria judge them at (to be not yet due); claims from X posts dated by the end of the window they were read from, and Scott Alexander's 2016 predictions dated by the post that scored them (re-dated to the post they were made in); 88 scored claims that were extra copies of another claim (removed). Every change is listed in `data/corrections/2026-11-03-cohort-1-changes.csv`; the reasons are in `docs/DECISIONS.md`, "Cohort 1 record fixes of 3 Nov".
-- v1.1 (2026-11-04): 33 subjects, 28707 claims, 0 verdict changes; method 0.1.0+03d23c0e, 0.1.0+3e03e2df, 0.1.0+3f84d14d, 0.1.0+5c636cf5, 0.1.0+5e91aed6, 0.1.0+7831dd9d, 0.1.0+8bea457b, 0.1.0+c876030c.
-  What changed since v1, and why:
-  - **The X subjects re-run.** In v1 most claims from X posts carried the date and link of the three-month window they were read from, not of their own post, and no post had been checked for deletion. The 24 subjects with an X handle were run again on 4 November 2026 on the current pipeline, from the same list of web documents as v1, with their posts read again through the X API, then topped up to about 80 resolved claims as v1 was. Every claim from X now links to its own post and carries that post's date, and every post quoted was found still on X on 4 November (2,170 posts; none deleted). The re-run judges every claim of those subjects again, web claims included, so their verdicts and scores can differ from v1's; re-running the same claims has kept about 92% of verdicts in earlier measurements.
-  - **The record fixes of 3 November.** The changes listed above under "Known defects in v1" are in this release for the seven cohort subjects not re-run; for the 24 re-run subjects the same rules now apply in the pipeline itself (duplicates, restated statements, not-yet-due verdicts, private life in propositions).
-  - **Two more top-ups.** Warren Buffett (74 to 78 resolved claims) and Scott Alexander (78 to 79) were topped up on 4 November 2026, keeping their stored verdicts.
-  - **Scores by claim type.** The dataset page now shows each kind of claim on its own (factual claims, predictions, promises), with rank order only within a kind and only at medium or high confidence; the combined score is kept in `scores.csv` as the `overall` row, but nothing is ranked by it. The README says so.
-  - **Slate Star Codex is not X.** v1's page counted 275 Slate Star Codex claims as X posts; the count now goes by the link's host.
-  - **Two subjects added:** Ørsted and GiveWell (October 2026 reports), so the release holds every subject the claims viewer at https://epistacklabs.com/explore shows, claim for claim.
-  - **New file:** `claims.parquet`, the claims table as Parquet.
-  - Judge B, the second judge, was unavailable for part of the re-run (the provider account ran out of credit). It was asked again for the 468 claims affected, on the same evidence; judge A's verdict stands under the judging rule, so no verdict changed, and 197 of them are now flagged as judged differently by judge B.
-- v1.1.1 (2026-11-05): 33 subjects, 28705 claims, 0 verdict changes; method 0.1.0+03d23c0e, 0.1.0+3e03e2df, 0.1.0+3f84d14d, 0.1.0+5c636cf5, 0.1.0+5e91aed6, 0.1.0+7831dd9d, 0.1.0+8bea457b, 0.1.0+c876030c.
-  What changed since v1.1, and why (no verdict or score changed):
-  - **Two claims removed for being about private life:** Mark Zuckerberg's "I love surfing." and a remark of Rodney Brooks's about a ride-hailing service that refers to his house. Neither could be checked, so no score changes. Listed in `data/corrections/2026-11-05-privacy-changes.csv`.
-  - **Replies:** a subject's `respond_url` now explains how to reply; replies are made with the form on each claim's own page at https://epistacklabs.com/claims/<claim_id>.
-  - **Deletions:** every post quoted (2,170) was found still on X by lookups on 5 November 2026; none deleted. A post found deleted later is left out of the next release; files already published are replaced by that release, and the dataset's git history keeps earlier files.
-  - `datapackage.json` now lists `claims.parquet`.
+# Releases
+
+- v1 (2026-10-28): 31 subjects, 29754 claims, 0 verdict changes; method 0.1.0+4bff4432, 0.1.0+ecd0c1d5.
+- v1 README corrected (30 Oct 2026): the stability line printed "? of ?" and now gives the measured figures (152 of 164 on a repeated run, 102 of 109 researched again from scratch); the caveat that every report was made through Claude Code on a subscription plan, with the API agreement check (276 of 310), was missing and is added. The data files are unchanged.
+- Known defects in v1, found 3 Nov 2026 and fixed in the stored reports for v1.1 (v1's files are unchanged): a proposition naming where a subject lives, and a joke about a subject's own death scored as a prediction (both removed, with 50 more claims about private life); six verdicts given before the date their criteria judge them at (to be not yet due); claims from X posts dated by the end of the window they were read from, and Scott Alexander's 2016 predictions dated by the post that scored them (re-dated to the post they were made in); 88 scored claims that were extra copies of another claim (removed). Every change is listed in `data/corrections/2026-11-03-cohort-1-changes.csv`; the reasons are in `docs/DECISIONS.md`, "Cohort 1 record fixes of 3 Nov".
+- v1.1 (2026-11-04): 33 subjects, 28707 claims, 0 verdict changes; method 0.1.0+03d23c0e, 0.1.0+3e03e2df, 0.1.0+3f84d14d, 0.1.0+5c636cf5, 0.1.0+5e91aed6, 0.1.0+7831dd9d, 0.1.0+8bea457b, 0.1.0+c876030c.
+  What changed since v1, and why:
+  - **The X subjects re-run.** In v1 most claims from X posts carried the date and link of the three-month window they were read from, not of their own post, and no post had been checked for deletion. The 24 subjects with an X handle were run again on 4 November 2026 on the current pipeline, from the same list of web documents as v1, with their posts read again through the X API, then topped up to about 80 resolved claims as v1 was. Every claim from X now links to its own post and carries that post's date, and every post quoted was found still on X on 4 November (2,170 posts; none deleted). The re-run judges every claim of those subjects again, web claims included, so their verdicts and scores can differ from v1's; re-running the same claims has kept about 92% of verdicts in earlier measurements.
+  - **The record fixes of 3 November.** The changes listed above under "Known defects in v1" are in this release for the seven cohort subjects not re-run; for the 24 re-run subjects the same rules now apply in the pipeline itself (duplicates, restated statements, not-yet-due verdicts, private life in propositions).
+  - **Two more top-ups.** Warren Buffett (74 to 78 resolved claims) and Scott Alexander (78 to 79) were topped up on 4 November 2026, keeping their stored verdicts.
+  - **Scores by claim type.** The dataset page now shows each kind of claim on its own (factual claims, predictions, promises), with rank order only within a kind and only at medium or high confidence; the combined score is kept in `scores.csv` as the `overall` row, but nothing is ranked by it. The README says so.
+  - **Slate Star Codex is not X.** v1's page counted 275 Slate Star Codex claims as X posts; the count now goes by the link's host.
+  - **Two subjects added:** Ørsted and GiveWell (October 2026 reports), so the release holds every subject the claims viewer at https://epistacklabs.com/explore shows, claim for claim.
+  - **New file:** `claims.parquet`, the claims table as Parquet.
+  - Judge B, the second judge, was unavailable for part of the re-run (the provider account ran out of credit). It was asked again for the 468 claims affected, on the same evidence; judge A's verdict stands under the judging rule, so no verdict changed, and 197 of them are now flagged as judged differently by judge B.
+- v1.1.1 (2026-11-05): 33 subjects, 28705 claims, 0 verdict changes; method 0.1.0+03d23c0e, 0.1.0+3e03e2df, 0.1.0+3f84d14d, 0.1.0+5c636cf5, 0.1.0+5e91aed6, 0.1.0+7831dd9d, 0.1.0+8bea457b, 0.1.0+c876030c.
+  What changed since v1.1, and why (no verdict or score changed):
+  - **Two claims removed for being about private life:** Mark Zuckerberg's "I love surfing." and a remark of Rodney Brooks's about a ride-hailing service that refers to his house. Neither could be checked, so no score changes. Listed in `data/corrections/2026-11-05-privacy-changes.csv`.
+  - **Replies:** a subject's `respond_url` now explains how to reply; replies are made with the form on each claim's own page at https://epistacklabs.com/claims/<claim_id>.
+  - **Deletions:** every post quoted (2,170) was found still on X by lookups on 5 November 2026; none deleted. A post found deleted later is left out of the next release; files already published are replaced by that release, and the dataset's git history keeps earlier files.
+  - `datapackage.json` now lists `claims.parquet`.
+- v1.1.2 (2026-11-10): 33 subjects, 28548 claims; method 0.1.0+03d23c0e, 0.1.0+28da144b, 0.1.0+3e03e2df, 0.1.0+3f84d14d, 0.1.0+7831dd9d, 0.1.0+8bea457b, 0.1.0+c876030c.
+  What changed since v1.1.1, and why:
+  - **Document dates corrected.** Some pages had been dated by the day the site was last updated, or the day an episode was released, instead of the day the words were written or said, so their claims were judged as of the wrong date. A reader spotted it in Warren Buffett's 2010 Giving Pledge letter ("to date about 20% of my shares have been distributed"), which had been dated 2026 and resolved false; it is now judged as of 2010 and is unresolvable. Twelve documents were re-dated in eight subjects' records (Warren Buffett, Ray Kurzweil, Jeremy Grantham, Satya Nadella, Nassim Taleb, Steven Pinker, Daniela Amodei, Nate Silver), 524 claims took the corrected date, and the affected claims that had been researched were judged again from the start, with new criteria, research and both judges. 71 verdicts differ from v1.1.1 as a result; each change is recorded in the claim's history with the reason "a corrected record". Listed in `data/corrections/2026-11-05-document-dates-changes.csv` and `data/corrections/2026-11-10-document-dates-changes.csv`.
+  - **Two documents removed with their claims** (157 claims, 18 of them with a verdict): a transcript of a Michael Burry interview that names no date (152 claims), and a Bill Gates page listed as his 2013 annual letter whose stored text was the site's notices of 2025 (5 claims). A document whose date is not known is not used.
+  - **Deletions:** every post quoted (2,170) was found still on X by lookups on 10 November 2026; none deleted.