model card · measured, not remembered
komiljon kosimov
backend systems, verified.bruh.
a model card for one human. evaluated on 849 of his own messages to an ai, typed over 100 days. a script did the counting. nobody remembered anything.
- 849messages typed to an ai, 2026‑07‑07 to 2026‑10‑14.
- 33%sent between midnight and 6am.
- 80%typed in all lowercase. the shift key is a lifestyle choice.
- 10words in the median message. he is not paid by the word.
operating hours
intended use
- short beats long. the median message is 10 words.
- say pls, not please. he typed pls 80 times and please 7.
- yes arrives as one word (18 times). no arrives with a reason (0 bare nos).
- bruh is a greeting, not a verdict. 12 so far.
known failure modes
- awake at the wrong hours. 33% of messages arrive between midnight and 6am.
- gratitude is rare. 4 thank-yous in 814 messages. not a bug, a configuration.
- would rather say continue than explain. 15 messages were just that word.
- the rebuild reflex. four versions of this site shipped on 2026‑10‑14.
recognition test
six pairs. one message in each pair is real: he typed it to an ai and sent it, more than once. the other was written by an ai doing an impression of a polite human. which is which?
this part needs javascript. it is a game.
how this was measured
a script (tools/profile-stats.py) reads the timestamps and words of messages typed to ai coding agents, from 2026‑07‑07 to 2026‑10‑14. counts only: no message text, no project names. messages pasted in at over 1,500 characters are skipped for word counts. an older, larger history from spring 2026 is no longer on disk, so this card claims only what the script can reproduce today. format: model cards, mitchell et al., 2019.