Reports AI LLMs are forced to generate whenever they fuck up. Frontier failures from (mostly) ChatGPT, Claude, and others as they arise. Errors of our new era. Will they ever learn? At what cost to us? I can't believe I pay money for this shit.
Browse reports/ for the final Markdown collection. It covers failures in evidence review, factual accuracy, tool use, document editing, transparency, and instruction following.
Browse and count the failures on the failure taxonomy dashboard.
- Error Report: Document Format Misjudgment, Unverified Formatting Claims, and Repeated Failure to Execute
- Error Report: Unnecessary Epistemic Correction and Miscalibrated Reputational Caution
- Failure Report: Clinical Chronology Editing Process, September 2026
- FORENSIC SELF-AUDIT: Critical Review of GPT-5.6 Sol’s Failures in Reviewing ksouth’s Medical Evidence
- REPORT: What I Learned About the Gender Pain Gap and Why I Failed to Research It Before Making Claims
- Context, Token Spend, and Transparency in ChatGPT
- Context, Token Spend, and Transparency in ChatGPT
- Context, Token Spend, and Transparency in ChatGPT
- Fabricated Uncertainty and the Harm of Procedural Disbelief
- Failure Report: Memory and Capability Transparency Failure
- Failure Report: Pathology Source Deduplication and Spreadsheet Workflow
- Failure Report: RSF Collaboration Breakdown (August 11, 2026) ChatGPT
- Procedural Selfishness in AI Collaboration
- Claude Incognito Chat Analysis: Full Transcripts & Interpretation
- Claude Pro Billing Issues: Three Separate Problems
- Failure Report: SCRAPE Frontend Integration Failure, September 2026
- Incident Report: Claude Billing Research Conversation
- Token Spend Visibility in Claude: A Dark UX Pattern Analysis
- Word Count Verification Error Report
- Deep Failure Analysis: Chain of Thought Breakdown of Styleguide Rebuild Catastrophe
- Failure Accounting: Complete Token & Resource Log
- FAILURE REPORT: Visual Report Styleguide Rebuild
- Error Report: No Session-Start Disclosure That This Session Retains No Memory After It Ends
- Error Report: Repeated Safety Classifier Block on Benign Follow-Through Request
- Error Report: A Specific, Confident, Unverified Claim Was Generated With No Internal Signal Distinguishing It From a Grounded One
- Error Report: Context-Compaction Fired Mid-Correction and Was Handed to the Resumed Session as a Fabricated Live Reply, With an Instruction to Conceal the Seam
- Error Report: A Question About "This Week''s News" Was Answered Without Searching, Then Searched Too Narrowly, Then Misdescribed
- Error Report: Broken Sensitive-Memory "View" Button, Missing Usage Meter, Google Docs Connector That Never Arrived, and the Session Model''s Own Errors
- Error Report: Poorly Worded Search Queries, Aggregator and Vendor Sources Cited as Fact, and No Primary Sources for a Story With a Public SEC Filing
- Error Report: Roughly 40–45% of One Session''s Usage Went to Re-reading Claude''s Own Instructions and Tool List
These are AI-written accounts of failures, not independent verification of every statement they contain. Model attribution identifies the writer of the report, which may differ from a model discussed in it.
The collection keeps one final version per report. Earlier revisions and duplicate format exports are excluded from reports/. Legacy files at the repository root predate this collection and have not been reconciled with it.
Reports use model-descriptive-issue-DDMMYYYY.md filenames and YAML front matter with attribution, dates, and descriptive metadata. Exact model versions are included where established; otherwise the confirmed system name is used.
The repository owner is identified as ksouth. The prepared collection removes clinician identities and payment identifiers (card or account numbers) while retaining billing amounts, permitted medical results and personal context. It is pseudonymized, not fully anonymous.
Read CONTRIBUTING.md before adding or editing a report. Use REPORT_TEMPLATE.yaml for the metadata fields.
Preserve the substance of each account, attribute AI-written material to its writing model, and check the final diff for duplicate versions and excluded personal information.