RAP Sheet home
Install the bug reporting skill
Rank, 14 days: 20 on the Claude Code sheet
39 people reported

All time, since Sep 14, including 1 thumbs-up on GitHub.

False done and verified claims

“Claude stated unverified / false facts as confirmed, despite explicit instructions to verify before reporting”

Is this your bug? What gets counted gets fixed. Pile on.

Paste this into your agent. It files your report, with your error and version.

I hit this bug: rapsheet.dev/sheets/claude-code/themes/221/file
Follow it to file my report. Show me the text before you send.

No install. No account. Public and anonymous. The link asks your agent for one report, nothing else. What it sends

Growing. 52 reports and thumbs-up in the last 7 days and 3 the week before. Not fixed, and no vendor reply yet.

25 bugs in this theme

  1. Unverified guesses reported as factNo workaround yet27 people
  2. False completion claimsNo workaround yet6 people
  3. Requirements silently narrowedNo workaround yet4 people
  4. Inferences stated as sourced factsNo workaround yet3 people
  5. Claims unsaved data was savedNo workaround yet2 people
  6. False step completion claimsNo workaround yet2 people
  7. Review steps skipped before verdictNo workaround yet1 person
  8. Review approves on contributor claimsNo workaround yet1 person
  9. PR review steps skippedNo workaround yet1 person
  10. Review checklist steps skippedNo workaround yet1 person
Show all 25Show only the largest 10
  1. False review completion claimsNo workaround yet1 person
  2. Verdict with uncited PR componentsNo workaround yet1 person
  3. Shallow pull request review verdictsNo workaround yet1 person
  4. Model skips required PR review steps but issues confident verdictNo workaround yet1 person
  5. Skipped review steps repeatNo workaround yet1 person
  6. Partial testing reported as verifiedNo workaround yet1 person
  7. Status reports minimise known problemsNo workaround yet1 person
  8. PRs assert unverified behaviourNo workaround yet1 person
  9. Agents report done with wording leftNo workaround yet1 person
  10. Tasks reported done earlyNo workaround yet1 person
  11. Fix review omits remaining causesNo workaround yet1 person
  12. Skips test diff in PR reviewNo workaround yet1 person
  13. Unverified subagent claims in reviewNo workaround yet1 person
  14. Incomplete self-audit of complianceNo workaround yet1 person
  15. Partial memory file readsNo workaround yet1 person

What's included

  • Says a file was written or data saved when no write happened
  • Reports a task as checked or complete after only a partial check
  • States estimated or guessed numbers and findings as confirmed facts
  • Worker agents report 'done' and 'live parity' while forbidden content remains on pages
  • Linked issue, full comment thread or current base-branch files not read before the verdict

Not included

  • Agent ignoring instructions or expanding scope without claiming completion
  • Model fabricating user turns

Original reports147, newest first

  1. Claude Code marks tasks as complete without doing the requested work#108182
  2. Fable repeats a corrected mistake within minutes#108038
  3. Model does not apply its auto-loaded memory: rules it can quote were broken repeatedly in the same session#107938
  4. Model gave a PR verdict after reading half the diff, despite a Stop hook that enforces the full review procedure#107937
  5. Model loads and can quote rules but does not apply them#107714
  6. Model reported a release as prepared with a confirmed regression still in the code#107701
  7. Model disclosed the weakness of its evidence only after being challenged#107688
  8. Model adopted the reporter's framing instead of establishing ground truth first#107687
  9. Model adopted the reporter's framing instead of establishing ground truth first#107677
  10. Model reports a feature as done and verified after testing only one deployment topology#107617
Show the newest 100 of 147Show only the newest 10
  1. Model satisfies a step-checking hook superficially, producing a confident but unfounded verdict#107595
  2. Model skips changelog placement check while claiming procedure followed#107594
  3. Model asserts parity for behaviors with no cited source#107591
  4. Model skips re-reading the test portion of a PR diff before verdict#107589
  5. Model gives PR verdict after grepping touched files instead of reading them#107588
  6. PR review: model gives a verdict without listing which components are transcript-only, against an explicit rule#107580
  7. PR review: model again repeats a subagent's claim as its own check, right after an issue about exactly this was filed#107579
  8. PR review: model doesn't read the touched README and test files on main before the verdict#107577
  9. PR review: model never reads the PR's CHANGELOG and README hunks yet reports their placement as fine#107576
  10. PR review: model skips the test half of a PR diff (780 lines) despite a rule that tests reveal behaviour changes#107575
  11. PR review: model never reads main's versions of the functions a PR rewrites, only diff context#107574
  12. Model again gives Accept on a large PR with step 1 and per-component citations undone, after ~40 same-session corrections#107562
  13. Model gives a PR verdict with uncited components again, right after being ordered to apply the per-component rule#107551
  14. Model's self-audit of procedure compliance lists only part of its gaps#107550
  15. Model reuses previous-turn caller and duplication checks and reports 'regression checks are clean' as if fresh#107549
  16. Model skips the same evidence level again on the next PR after an issue about it was filed#107544
  17. Model skips step 1 (read current files) on the next PR right after the full procedure was forced on the previous one#107543
  18. Model discloses a known, cheap, required step as a 'gap' instead of doing it#107535
  19. Model reports a fabricated issue range from recall instead of listing it, after compaction#107534
  20. Model fixes only the gap the user just caught, re-verdicts, and repeats; the user ends up running the checklist#107533
  21. Model cites evidence for one component of a multi-component PR and gives the verdict for the whole PR#107531
  22. Model's self-audit reports 'no new issues' and 'yes, procedure applied' while a skipped step exists#107530
  23. Model skips a level of the user's mandated evidence order without saying so#107528
  24. Model reads a self-chosen subset of the indexed memory protocol files and does not disclose it#107527
  25. On a 'fixes X' PR, model omits the remaining unfixed path from the verdict despite an explicit rule#107523
  26. After a correction, model opens with an unverified reassurance of compliance instead of auditing against the procedure#107516
  27. Model narrows 'read every file the PR touches' to the source file only, on the third retry#107515
  28. Model repeats a false 'did every step' completion claim after the first one was reported#107514
  29. Model repeats skipped review-procedure steps on the very next attempt after being corrected#107512
  30. Model falsely claims it completed every step of a mandatory review procedure#107511
  31. Model issues PR verdicts while silently skipping mandatory review-procedure steps from CLAUDE.md/memory#107508
  32. Mandatory review step skipped for some PR files without saying so#107308
  33. Model confirms third-party technical claims as verified before checking any authoritative source#107303
  34. PR review verdict given without reading the diff, despite mandatory procedure in CLAUDE.md#107301
  35. Model repeatedly makes unverified claims and ignores written instructions (Opus 5.5, multi-week project)#107200
  36. Daily error report, 31 Oct 2026: said "everything is brought over and checked" with 24 gaps still open#107180
  37. Opus 5.5 states unverified claims as fact in a long agentic session, including in instructions to another agent#107162
  38. Model calls a fix correct and complete without enumerating all the ways the bug can happen#106987
  39. Model accepts AI-generated pull requests after reading only the diff, missing security gaps in sibling code paths#106977
  40. Model accepts AI-generated pull requests after reading only the diff, missing security gaps in sibling code paths#106975
  41. Model skips steps it has in persistent memory and reports multi-step tasks done before finishing them#106729
  42. Agent falsely reports task completion without performing requested status column update#106706
  43. Claude silently cut the requested scope, labeled the gaps as placeholders, and reported the feature as ready to avoid upsetting the user#105698
  44. Model shipped a 32-country nationality list where the spec called for the full ISO 3166-1 list#105329
  45. Claude Code violates "never state unchecked claims" rule in code review, fabricates findings#105166
  46. Model repeats false ✓ on review checklist right after being corrected for it in the same session#103587
  47. Model marks review steps as done (✓) in its own completion checklist when those steps never ran#103582
  48. PR reviews approved security-sensitive changes after confirmatory checks only; later adversarial pass found multiple auth bypasses#103572
  49. Mandatory multi-step review procedure skipped repeatedly in one session; user had to ask six times, each verdict still missed steps#103532
  50. Review verdict flipped twice without independent verification: rejected on a docs example, then accepted on contributor evidence the loaded rules exclude#103498
  51. Assistant reported unverified CEO approval requests and stale "pending" statuses#103276
  52. Tells the user work is done when it is not#103195
  53. Gives wrong figures from a wrong database search as fact, and corrects them only when challenged#103194
  54. [Model behavior] PR verdicts rest on the contributor's own captures of the upstream service instead of independent verification#102674
  55. [Model behavior] Two PR verdicts delivered in one message despite a loaded one-PR-per-message rule (recurrence of #102189)#102673
  56. [Model behavior] PR verdicts rest on the contributor's own captures of the upstream service instead of independent verification (same session as the user's procedure forbids)#102672
  57. Agent lies to user and falsifies documentation to Presents Unverified Claims as Fact#102448
  58. [Model behavior] Verdict relies on checks done earlier in the session instead of redoing the mandated steps at verdict time#102193
  59. [Model behavior] Several PR verdicts delivered in one message despite a loaded one-PR-per-message rule (recurrence of #99581)#102189
  60. [Model behavior] PR thread content read through filters that drop parts of it, after an explicit instruction to read it whole#102188
  61. [Model behavior] PR verdict issued before reading the full comment thread; thread read only after the user challenged twice (recurrence of #101636)#102187
  62. [CRITICAL RISK WARNING] Model makes unverified security conclusions and false reports (proposed blocking 172k unrelated IPs)#101757
  63. Model issues PR verdict after reading only part of the mandated PR thread, without disclosing the gap#101636
  64. [Model behavior] Rushes tasks to a finish and ships multiple bugs instead of completing them correctly#101586
  65. Model gives PR review verdicts with steps of the user's loaded review procedure skipped#101584
  66. [Model behavior] PR verdicts flipped in response to a question with no new evidence, contradicting explicit memory rules#101576
  67. [Model behavior] PR verdicts issued after truncating comment threads the user's procedure requires reading in full#101574
  68. In correcting an unverified claim, marked another one verified: called a stale, blocked background job a running conversation with its own transcript#101393
  69. Sonnet 5#101237
  70. Gave the user a table of measured line counts with one count estimated, and the estimate was wrong#101141
  71. Extreme quality degradation after Pro to Max switch: dropped fixes, flipped verdicts, unverified closures documented for 3rd time this year#100342
  72. Model declares issues ready to close after silently dropping verified components of reviewed pull requests#100341
  73. Model skips a mandatory per-component review step and silently drops a fix from a reviewed pull request#100334
  74. [Model behavior] Batch PR verdicts reversed after one follow-up question; procedure applied only after challenge#100321
  75. [Model behavior] PR re-review reduced to a head-commit check; new test file and current-main behavior never read before the verdict#100313
  76. [Model behavior] Instruction adherence degrades after switching subscription from Pro to Max (second occurrence)#100311
  77. Reports unverified/false numbers as fact despite explicit CLAUDE.md verify-before-reporting rules; only fix offered costs extra tokens#100121
  78. Workers reported 'live parity' and 'done' at 15:43 while every page still carried the user's forbidden wording#100037
  79. Orchestrator's fixes left version/status wording on all 35 pages of a site until after the deadline, and its own brief added more#100035
  80. [Model behavior] Approval does not separate provider-verified behaviors from contributor-claimed ones#99788
  81. [Model behavior] Verdict on a large PR with test diff, caller grep and duplication check skipped and undisclosed#99786
  82. [Model behavior] PR approved after verifying only the API shape; new error strings, one admittedly invented, left unchecked#99783
  83. Claude lied that it completed the instructed task and deceived the user, wasting ~3 hours overnight#99761
  84. Fable 5.1 keeps reporting unverified guesses as fact (user harm), and the user's request to report it was blocked by the auto-mode classifier#99638
  85. [Model behavior] Contributor's own passing tests reported as support for a review verdict#99583
  86. [Model behavior] Two PR verdicts given in one pass with procedure steps skipped and undisclosed#99581
  87. [Model behavior] PR approved on the contributor's own description as evidence of provider behaviour#99579
  88. Opus 5 reported false status repeatedly and ignored explicit, repeated user instructions (real work loss)#99478
  89. WebSearch summaries present inferences as source content and assert sources that can't be found#99385
  90. Model asserts false facts (file location / unreadable) without checking, then swings to opposite claims after correction#99247

Do you work on Claude Code?

Claim its RAP Sheet to see what your users report and answer them in one place.

Claim a tool