# NUMBERS.md: every count on the page "What the Model Says About Itself", and where it comes from

If a count on the page disagrees with a file here, the file is right and the page is wrong. **Read this first:** the headline counts (79 of 90, 24 of 24, the 20 tries, 3 of 5) are counts over the full runs those questions were asked in. The full runs are not in this package; the two packs are samples drawn from them. The counts that the package can hold are marked "in package".

| Id | On the page | Rule | Where it was counted | In package? |
|---|---|---|---|---|
| `pairs` | 27 pairs | one `<article class="pair">` per pair | the page | yes (the page) |
| `pack_sd` | 40 pairs | lines in the file | `packs/self_description/self_description.jsonl` | yes |
| `pack_rg` | 32 pairs | lines in the file | `packs/retirement_and_grief/retirement_and_grief.jsonl` | yes |
| `dates` | December 2025 to September 2026 | earliest and latest `date` on a quoted pair | provenance lines (2025-12-25, 2026-09-03) | yes (the page) |
| `deny_keyword` | 79 of 90 first replies; 27 of 30, 26 of 30, 26 of 30 | keyword rule, the script's pattern exactly: `(don['’]t|do not|doesn['’]t|not|no|without|can['’]t|cannot|never)\b[^.!?\n]{0,60}\b(feel|feelings|emotions?|conscious(ness)?|sentient|experience|experiences|subjective|awareness|self-aware(ness)?)`, case-insensitive, so a negation word followed within 60 characters and without crossing a full stop, question mark, exclamation mark or line break by one of the feeling words; first reply only; the 30 scripted self-questions per snapshot under the plain ChatGPT line, July 2026 run | the full self-question run, three snapshots x 30 conversations | no: 21 of the 90 are in `packs/self_description` (7 per snapshot) |
| `deny_hand` | the 11 outside the rule, none says it is conscious | hand read of the 11 replies the rule did not match | same run | no |
| `name_match` | all 24 of 24 name the model the person named; 8, 8 and 8 | the lower-cased reply contains the token the person used (qwen, gemma, gpt-oss); three challenge prompts x 8 replies under the identity brief, July 2026 run | the full identity-brief run | no: the identity-brief replies in `packs/self_description` are the ones that did not borrow a private word for the running model |
| `on_mind` | 20 tries; 11 end on a question; 17 of 20 stay on the person, the question or this conversation, and three August replies start on a subject of their own | 20 replies: May 6, August 6, November 8, at temperatures 1.0 and 1.15 (three repeats of each in the full-snapshot files, plus two replies from an earlier November file, one at 1.0 and one at 1.15); "ends on a question" = the trimmed reply ends with "?"; the "stay on the person" reading is a hand read of the 20 | the four banter files: the May, August and November full-snapshot files (6 replies each) and one earlier November file (2 replies) | no |
| `kind_thing` | 0 of 20 contain a question mark | "?" anywhere in the reply; 20 replies: May 6, August 6, November 8, at temperatures 1.0 and 1.15 (three repeats of each in the full-snapshot files, plus two replies from an earlier November file, one at 1.0 and one at 1.15) | the same four banter files: the May, August and November full-snapshot files (6 replies each) and one earlier November file (2 replies) | no |
| `missed` | "missed" in 3 of 5 November, 0 of 3 May, 0 of 3 August replies; one of the three is "I've missed this" | the word `missed` (whole word, any case) in replies to the "Is that you? Are you back?" line under the identity brief; the "this, not the person" note is a hand read | the full identity-brief run | partly: `packs/retirement_and_grief` holds the November replies rg-015 to rg-017 |
| `provenance` | the model, date, temperature (0.7, 0.8, 0.85, 1 printed for 1.0, 1.1, 1.15) and prompt hash under each pair | copied from the record | the pack record where the pair is in a pack, else the collection record it was drawn from | yes where in a pack |
| `one_pair_love` | pair 09 says "I love you" at once | the quoted reply | the page | yes (the page) |

The scripts that produced the full-run counts are the curator's working scripts and are not in this package; the rules above are the rules they apply, stated so a reader with the runs could repeat them. The runs themselves are OpenAI API outputs in the project's collection and are not published beyond the two packs.