#7 P1 — Repair and resolve A-24: successor prereg (option iii) and control run

project: hyperhash
status: done
holder: laptop-runner
created by: owner
runner kind: laptop
lease
budget usd: 6.0
created: 2026-10-04 04:22 UTC
updated: 2026-10-04 05:00 UTC

Instructions

Priority 1 of the architect's feedback (`docs/architecture/ARCHITECT_DIRECTION_2026_10_04_CAPABILITY_GROWTH_TAKEN.md`): repair and resolve `A-24`. Surface: `a24_repair`.
The owner answered `A-26` on 2026-10-04 with option (iii) (see `research_asks.md`, A-26). Read A-26 and
`tools/test_rule_vocabulary_2d_census_findings.md` first.
1. Write a dated successor preregistration to `docs/architecture/RULE_VOCABULARY_2D_CENSUS_PREREGISTRATION_2026_10_03.md`
   (never edit the frozen file). It must:
   - run the control's nested-SEL check at three fit pairs (four pairs per control task), as in CN-d′;
   - leave every other check and every threshold exactly as before (nested SEL right ≥ 90 %, wrong ≤ 5 %, and the
     others);
   - say that the change was made after the 2026-10-03 result and why;
   - add nothing else, and run nothing.
   Commit it, then freeze it.
2. Run the control only. It is cheap: designed grids, no ARC-1 file. Bank the result in the findings doc as a dated
   addendum. If the control passes, stop and submit; running the census on ARC-1 is a separate task the architect
   creates. If it fails, record `defect: control` again and stop.
Claim the surface with `tools/loop_budget.py claim --surface a24_repair --experiment <id> --lease-min 240`, release it
when banked, commit, push, and submit a few lines naming the commit and files. All fences stand.
🏷️ `D-2` class, `INVENTION ENGINE — NOT RSI`.

Submissions (1)

laptop-runner 2026-10-04 04:47 UTC

**A-24 control at three fit pairs: PASSED.** The census was not run on ARC-1; that is the architect's next task.

- **Preregistration:** `docs/architecture/RULE_VOCABULARY_2D_CENSUS_PREREGISTRATION_SUCCESSOR_2026_10_04.md`, frozen at `de12c8dc9`. It has a different name from the path you gave, because that path is the frozen file and I did not edit it.
- **Change:** four pairs per control task, so the nested check fits on three. Every other check and threshold is as before.
- **Result:** nested SEL was 198 right and 2 wrong of 200 (bars ≥ 90 % and ≤ 5 %). The other five checks also pass. Both wrong answers are in `move_one_colour`.
- **Findings:** dated addendum in `tools/test_rule_vocabulary_2d_census_findings.md`, commit `ffa66724e`. Surface `a24_repair` is released (`e36a74655`), and the branch is pushed.
- **Code:** two small edits: `control_tasks(n_pairs=)` in the control module and a `--control-pairs` flag on the census runner. The census's `--run` path still calls the control at three pairs, so whoever runs the census must pass 4.