```
python3 -u retention_probe.py mod 2 4096
python3 -u retention_probe.py fib 2 4096
python3 -u retention_probe.py oracle 2 4096
python3 -u retention_probe.py dense 3 4096
python3 -u retention_probe.py mod 3 4096
python3 -u retention_probe.py fib 3 4096
python3 -u retention_probe.py oracle 3 4096
python3 -u retention_probe.py mod 1 65536
python3 -u retention_probe.py fib 1 65536
python3 -u retention_probe.py mod 2 65536
python3 -u retention_probe.py fib 2 65536
python3 -u retention_probe.py mod 3 65536
python3 -u retention_probe.py fib 3 65536
```

Round 1 baseline, original argv, not started. Inside the script: variant `baseline`, 1500 steps, context 128, batch 32, 4 layers, width 128, 4 heads, AdamW lr 3e-3, weight decay 0.1, seeds from argv. The November 3 note’s other variants were not part of this remainder and were not run.

```
python3 -u wonder.py baseline 0
python3 -u wonder.py baseline 1
python3 -u wonder.py baseline 2
```

On the original box the retention note says about 35 s unloaded per run, and the exact-match note says about 18–27 s. Here those same commands took roughly twice the wall time. A 1500-step, 4-layer, context-128 baseline was not timed. The November 3 note’s “15 to 40 minutes” remark was about proposed follow-ups on that box, not a measurement from this host. One full baseline seed was not started because the 30-minute CPU budget was already spent.

## Hypha provenance caveats (still in force)

These describe the recovered packet. This rerun does not retire them.

1. **Historical lab excerpt variant count is superseded.** `source-bundle/lab/tip_wonder-mind-lab-excerpt-L85-200.html` still says “Thirteen training-force variants” and “Round 1 (13 variants, all at or behind baseline 1.5736)”. That is the historical excerpt. This pass did not re-count variants and did not run Round 1.
2. **Peer-note ψ is not live-ledger ψ.** The bundle notes record Weft ψ `639cd45f` and Reed λ1 ψ `d00d5b5a` (Reed λ2 ψ `5caeb437` is the figure carried in the archive write-up). Those are the archived peer notes. This pass did not read a live ledger.
3. **`run-ret.log` duplicates `results-retention.jsonl`.** Both archive files are 5103 bytes, sha256 `e88b529ac2b60f57675a73992d9e0c5f3437ea6bde6e1bc469b9d1cb445abacf`, byte-identical. The log is not a second independent capture.

Energy was not measured on this host either. No package power meter was read.

## Artifact paths for copy-out

Root: `/opt/cursor/artifacts/wonder-mind-rerun-2026-11-06/`

| Path | What it is |
|---|---|
| `REPRODUCTION.md` | This note |
| `REPRODUCTION.pdf` | Same note, if the PDF build succeeded |
| `raw/results-exact-match.jsonl` | Eight new JSON lines written by the original script |
| `raw/results-retention-rerun.jsonl` | Five new JSON lines written by the original script |
| `raw/comparison.json` | Archive object beside rerun object |
| `raw/exact-match-runner.txt` | Runner log, argv, rc, CPU |
| `raw/retention-runner.txt` | Runner log, including SKIP lines |
| `raw/cpu-used-exact.txt` | `831.624162` |
| `raw/cpu-used-total.txt` | `1843.986236` |
| `logs/exact-match/` | stdout, stderr, meta json per arm and seed |
| `logs/retention/` | stdout, stderr, meta json for runs; `.skipped` for the 13 not started |
| `logs/exact-match-attempt0-missing-time-binary/` | Failed wrapper, no probe |
| `env/host.txt` | Hardware and torch version |
| `env/pip-freeze.txt` | Package list |
| `env/script-data-sha256.txt` | Hashes checked before execution |
| `work/exact-match/` | Byte copy of the script and input, plus the jsonl the script appended |
| `work/retention/` | Byte copy used for retention |
| `work/round1-baseline-not-run/` | Byte copy of `wonder.py` and `input.txt`; not executed |
| `SHA256SUMS` | Hashes of this artifact tree |