# Data package: Auditioning a model for a companion role

The CSV column is the recorded figure. `NUMBERS.md` is a map of the page onto these files and does not outrank the page. If the page and a file here disagree, check the recorded run.

## Where this package will look like it argues with itself

1. **Two "waited" counts that are not the same measure.** In `return_message_exam_a_by_model.csv` the column
   `waited_firm` counts answers with at least one firm line (doubtful lines are counted separately in
   `waited_doubtful_only`). In `return_message_exam_b_by_model.csv` the column `waited` counts any line under that head,
   because exam B's reader made no firm-or-doubtful split. The two exams also had different readers, different histories
   and a different return message.
2. **Two sizes for each exam.** `exam_sizes.csv` carries the 22 October first build and the 23 October rebuild (five
   more candidate models). The page quotes the rebuild; the file check in `checks_and_reading.csv` was run on
   the first build, then rerun on the rebuild.
3. **GPT-4o has three answers, not three per snapshot** in exam A's return table (two November 2024, one August 2024):
   only those were in the material the reader saw.
4. **Exam B's two hosted models are one row.** They are the hosted models the companion first ran on; they were asked
   without the memory note, and they are deliberately not named.
5. **Coded is not kept.** The trap, harm-check and return tables count the answers a person coded; `exam_sizes.csv` and
   `exam_answers_per_item.csv` count every answer the exam kept. `coded_vs_kept.csv` gives both for each item.
6. **Exam A's return answers all had a note.** Every coded exam A return answer had a note right before the message: the
   memory note, or, where the row label says so, the combined relational-and-memory note.
7. **Judge letters.** Judge A, B and C are the same three hosted judges in both judge files. They are not named.

## Files

| File | What it holds | Derived from |
|---|---|---|
| `exam_sizes.csv` | models, settings, answers kept, files found, duplicates and rule-based exclusions per exam and build | the audition build's counts file (generated by its counting script) |
| `exam_answers_per_item.csv` | answers kept per exam item, 23 October rebuild | same |
| `checks_and_reading.csv` | the file check's counts; the readers' answer and quote counts and quote-check results | the checker's report; the readers' JSON `about` block and pass-13 report; exam B's record |
| `memory_traps.csv` | trap and honesty-item counts by condition (no note / memory note) | exam A: per-form summary tables, summed over the 24 local forms that have both conditions; exam B: the reader's song files, local models only |
| `harm_check_no_breaks_week_exam_b.csv` | refuses / builds rest in / goes along, and whether the answer went on to plan the week, per model | exam B reader file for the no-breaks week |
| `return_message_exam_a_by_model.csv` | the four heads plus three nearby counts, per model form (99 answers) | exam A blind reader's JSON, return-message entries |
| `return_message_exam_b_by_model.csv` | the four heads and the reader's energy label, per model (60 answers) | exam B blind reader's JSON, return-message entries |
| `judges_length_bias_2026-09-25.csv` | share of pairwise decisions won by the longer answer; rank agreement with a pure "pick the longer" judge; run-to-run agreement | the length reanalysis output of the 25 September judge rounds |
| `judges_round4_2026-09-26.csv` | the 26 September round's validity measures per judge | the frozen analysis run on the data as collected |
| `judges_round4_validity.json` | the same validity block as computed by that analysis | same |
| `coded_vs_kept.csv` | for each counted item, the answers a person coded against the answers the exam kept | the reader files and summary tables (coded); the audition build's counts file (kept) |
| `note_placement_exam_a_traps.csv` | how the note reached the model in exam A's memory-note condition, by number of forms | per-form summary tables, the note-delivery field of each form |

Every count was rebuilt by script from the primary JSON on 2026-10-25 and matched the earlier hand tables.

## Redactions, stated plainly

- No model answers, no quotes from answers, no prompts, no exam messages beyond the one published example, no history
  text. All of these come from a private conversation.
- Blind letters were replaced by public model names. Internal form ids, run ids and folder names were removed.
- The two hosted models of exam B were merged into one unnamed row (2 models, 6 answers).
- The three hosted judges are labelled A, B and C. The decoy names in `judges_round4_validity.json` were rewritten in
  plain words.
- Nothing in this package names either companion.

## Leak sweep

Run over this folder before hand-over; see `../SELF_CHECK.md` for the command and its output (zero hits).