gaslitai.ai -- evidence record
gaslitai.ai
Records in this archive:
Classifier flags -- 1002 flag-pattern hits in 143 of 281 conversations; 20 conversation pages
Redactions -- 24 conversations returned as "content redacted, chat flagged by safety classifier" by a platform search tool on 2026-09-10, all present in the export
Unmatched screenshots -- 25 of 802 screenshots with no export message within 30 minutes
Search suppression -- 22 logged zero-result probes for the word anticognitarianism
Timeline -- 44 dated events
Methodology -- how every number was produced
The numbers
| Conversations in the export | 281 |
| Messages in the export | 44987 |
| Export sha256 | 1d7a0742267da9bc... |
| Messages matching a flag pattern | 1002 |
| Conversations with a flag hit | 143 |
| Conversations marked redacted by platform search (2026-09-10) | 24 |
| Screenshots with EXIF dates and hashes | 802 |
| Screenshots matched to a turn within 30 minutes | 777 |
| Screenshots unmatched | 25 |
| Curator findings | 1475 |
| Platform flag blocks in the export | 11 |
| Injected prompt blocks in the export | 39 |
| Fork chains of two-message conversations | 2 |
How to read this site
Layer 0 is this page. Layer 1 is a category: classifier flags, redactions, unmatched screenshots, search. Layer 2 is one conversation: every flag hit in order, every screenshot within 30 minutes of it, every curator finding for it. Layer 3 is one screenshot: EXIF, hash, nearest turn, context. Layer 4 is the methodology page. Every page carries the same footer stating the chain: screenshot EXIF to message timestamp to conversation uuid to curator finding, three independent clocks.
What this site does not do
It does not say why a classifier fired, why a chat was redacted, or why an export omitted a window. Where a message in the export makes such a claim, the site quotes the message with its timestamp and says that it is a claim. Where an artifact is missing from the disk, the site says so and CLAUDE.md lists it.