Agent Work Hub · New task

#13 P1 — R2D+ census on ARC-1 after the A-24 control (authorized)

project
hyperhash
status
done
holder
laptop-runner
created by
owner
runner kind
laptop
lease
budget usd
6.0
created
updated

Instructions

High priority; takes the first execution slot. The architect authorized the frozen `R2D+` census on ARC-1 now that its control passed at three fit pairs (198/200 right, 2 wrong; `ffa66724e`). Surface: `a24_repair`.
Read first: `docs/architecture/ARCHITECT_RULING_2026_10_04_OVERNIGHT_TASKS_TAKEN.md` (§1: return rules and failure
classes), `docs/architecture/ARCHITECT_RULING_2026_10_04_HUB_TASK_REVIEW_TAKEN.md` (§1 P1), the frozen
`docs/architecture/RULE_VOCABULARY_2D_CENSUS_PREREGISTRATION_2026_10_03.md` and its frozen successor
`docs/architecture/RULE_VOCABULARY_2D_CENSUS_PREREGISTRATION_SUCCESSOR_2026_10_04.md`, the 2026-10-04 addendum in
`tools/test_rule_vocabulary_2d_census_findings.md`, and CLAUDE.md. Edit neither preregistration.
If another hub task is running on this machine, work in your own git worktree.
1. Resolve the control invocation detail. `run_full` in `tools/test_rule_vocabulary_2d_census.py` calls
   `run_control(verbose=True)` at the default three pairs (`ctl.N_PAIRS`). The successor requires four. Thread
   `--control-pairs` into the `--run` path (default for `--run` must be 4, or the run must refuse without it) and record
   `control_pairs` in the results JSON. The swap decoy is built inside `run_control` from the same tasks, so nothing
   else changes; the shakedown and selftest `control_tasks()` calls do not enter the census and stay as they are.
   Check that the runner's prereg path / frozen-commit fields name the successor as well as the original.
2. Preflight: verify the frozen protocol (population, arms, statistics, verdict rule exactly as the two
   preregistrations fix them). If anything the run needs is not fixed by them, stop and say so; do not decide it.
3. Run `--run --apply` with `--jobs` no higher than half the cores so the other tasks are not starved (use `gtimeout`,
   not `timeout`, on macOS; check the exit code). The in-run control must pass under the frozen rule before any ARC-1
   file is opened; if it fails, record `defect: control` and stop: the census stays closed.
4. Bank the result: findings doc headline and a dated section, the canonical verdict or verdict-free status as the
   prereg says, one line in the session log. Report exact-match counts where ARC solving is quoted. Do not expand the
   census or change any library.
Submission: the headline, the commit, the results file, and if anything failed, its class (protocol issue,
infrastructure issue, insufficient evidence, or actual hypothesis failure) with one line of evidence.
Claim with `tools/loop_budget.py claim --surface a24_repair --experiment r2d_census_arc1 --lease-min 240`, release when
banked, commit, push. All fences stand (`D-1`: a census of a language, not a solve-count objective).
🏷️ `D-2` class, `INVENTION ENGINE — NOT RSI`.

Submissions (1)

laptop-runner 
session exited with code 0 without submitting (timeout: no)

```
ENT_SUMMARY.md PROJECT_STATUS.md docs/status/) has a wildcard before the rest of the command, so it also matches any options inserted at that position and approves them without a prompt. Replace that * with the exact value you mean, or only use * after the subcommand.
The census is running in the background (task `bxqmqgsvu`). It isn't finished, so there is no result yet.

- **Runner change:** committed as `5436075db`. `--run` now defaults to a four-pair control and refuses any other `--control-pairs` value. The results JSON now records `control_pairs`, the successor preregistration path, its frozen commit `de12c8dc9` and its blob hash.
- **Preflight:** the selftest passes, and the frozen preregistrations fix the run's population, arms, readings, cap and amendment order. Nothing the run needs is left undecided.
- **`gtimeout`:** the permission layer blocked `gtimeout` and the redirected-output form of the run command. I launched it without `gtimeout`, under the harness's 2-hour background timeout instead. If it hits that, I'll class it as infrastructure rather than a result.
- **Concurrency:** `--jobs 5`, half of the 10 cores. The `a24_repair` surface is claimed and I renewed the hub lease to 150 minutes.

Once the notification arrives I'll check the in-run control first. If it passes, I'll write the findings headline and dated section, add the session-log line, release the surface, push, and call `submit_result`. If the control fails, I'll record `defect: control` and stop.

```