Each bar includes all eligible prompts sent in that hour.
Prompt timestamps estimate T3 activity. They do not measure keyboard time or time spent coding.
Hours since the first prompt in the session.
Each dot is one day. Activity caps each prompt gap at 15 min.
Current word/phrase and category filters apply. Excerpts show matched text; context does not prove that a term was aimed at a person.
| Term | Uses | Share |
|---|
Scanned all 2,152 retained VPS thread records, including settled, archived and retained deleted records. Of 7,430 user-role messages, 3,970 had user attribution. After exclusions, the audit uses 3,381 submitted prompts across 569 threads, from 16 Sep to 10 Nov 2026, 18:01 UTC. This includes 471 prompts from retained deleted threads. All 2,069 older user-role records already occur in the newer store; they were not added again.
The linked Windows laptop did not answer. This report covers retained VPS history only. It cannot cover laptop-only records, messages erased from storage, other chat apps or text inside images/audio.
Excluded 362 test-fixture prompts, 194 automated MCP relays/continuations, 17 identifiable older agent handoffs, 13 relayed subagent answers and 3 exact text-and-time duplicates. A metadata filter excludes another 3,460 agent-authored user-role messages before this step. Older attribution can be imperfect.
Removed Obsidian delivery footers, paths, domains, code, tables, assistant quotations and recognizable copied response blocks. Kept the user comment attached to each T3 citation. Async-answer records have their initial question line removed. Repeated submissions at different times remain separate. Cleanup is conservative and rule-based; subtle copied text or indirect insults can be missed.
Words count case-insensitive whole-token matches against the audit lexicon, with observed spelling errors normalized (for example “fukin” → “fucking”, “cuntt” → “cunt”). “WTF”, “TF” and “FFS” stay separate. Word-family mode combines forms such as fuck/fucking/fucked/fuckups, shit/shitty/shitload/dogshit, and ass/arse/asshole/dumbass. A fused “whateverthefuck” counts as one “fuck”.
Insults include derogatory terms for people, software, ideas and the speaker, plus mentions of these terms. This is a language-use count, not a count of personal attacks, anger or abuse. Common criticism such as “bad” or “inefficient” is excluded. Context review removed 7 literal uses (trash icons, lazy unmounts, color transfer, “sucked into”). Positive profanity such as “bloody amazing” is included.
Phrases use a published set of 25 expressions found in the prompts. Variants are combined, and the longest overlapping phrase wins. “For fuck’s sake” includes “fuck sake”; “what the fuck” includes “what in the fuck” and “what in the genuine fuck”. Phrase counts describe selected multiword expressions, with a separate unit and denominator; word and phrase totals must not be added. They are not an exhaustive count of all possible offensive sentences.
Software-project activity excludes the novel, QLO2, No project, Saved conversation and issues projects. This leaves 2,896 prompts in software projects; some still concern planning or admin work. Sessions combine prompts across projects on the VPS and reset after the selected idle gap. Daily activity adds time between consecutive prompts, capped at 15 min per gap, and assigns each interval to its ending prompt’s day. Isolated prompts add zero time. Session duration includes waiting. These estimates cannot establish that long coding sessions cause language use.
Rates use all eligible prompts in the same bin as the denominator. No prompts means no rate. Faint bars/dots have fewer than 20 prompts. Hover a mark for uses, matched prompts and total prompts. The daily correlation uses days with at least 20 prompts and describes association only.