Wonder Mind rerun — 2026-11-06
Rerun evidence from the recovered scripts. This pass does not rewrite the archive, does not prove what machine produced
the November 3–5 logs, and does not replace those logs.

Reviewed archive head left untouched: `31344f8c82efd0204f283936b1dea89ef0a5ccb3`
Pull request left open and unedited: https://github.com/Chaos2Cured/FreeLattice/pull/330
Originals were copied out and executed in `/opt/cursor/artifacts/wonder-mind-rerun-2026-11-06/work/`. The scripts append
their jsonl in the working directory, so they were not run inside `docs/library/wonder-mind/2026-11-06-v1/`.

AUTONOMY.md was read as the project foundation and was not modified. This was local computation. No merge, no new pull
request, no publication, no Quiet Room, no app-tab smoke, no patent ingestion, no new hypothesis arm.


What reproduced

**Nov 5 exact-match, seeds 1 and 2, all four recorded arms.** Eight runs, original argv, default `NB=65536` because the
recorded jsonl used that default. Seven of the eight logged JSON objects match the archive line for line except the
script’s own `sec` field (wall clock on this host is slower).

Matched except wall clock: exact seeds 1 and 2, paraphrase seeds 1 and 2, force seeds 1 and 2, dense seed 1. That
includes `params=1660435`, store hit, in-prompt, model accuracy, `mean_p_true`, `copy_loss_last`, and the
`sample_afterB` rows.

Headline means that use only those matching arms are unchanged:

| Arm | afterA model | afterB model | store afterA / afterB | in_prompt afterB |
|---|---:|---:|---:|---:|
| exact | 0.795 | 0.055 | 1.00 / 1.00 | 0.00 |
| paraphrase | 0.030 | 0.000 | 0.00 / 0.00 | 0.00 |
| force | 0.990 | 0.000 | 1.00 / 1.00 | 1.00 |


**Round 2 retention, first five commands of `run_ret.sh` only.** Each completed run matches the archive jsonl except
`sec`.

| argv | model A / B | store A / B | val_loss | writes | occupancy |
|---|---|---|---:|---:|---:|
| `dense 1 4096` | 0 / 0 | 0 / 0 | 3.5112 | 0 | 0 |
| `mod 1 4096` | 0 / 0 | 0 / 0 | 3.513 | 4096 | 4096 |
| `fib 1 4096` | 0 / 0 | 0 / 0 | 3.4962 | 4096 | 4096 |
| `oracle 1 4096` | 0 / 0 | 0.99 / 0.99 | 3.0551 | 60054 | 60054 |
| `dense 2 4096` | 0 / 0 | 0 / 0 | 3.4897 | 0 | 0 |
`params=1735125` on every retention row, same as the original jsonl. Refused-write counts also matched (mod seed 1:
1360339; fib seed 1: 1356805; oracle seed 1: 1045297).


What did not reproduce

**exact-match `dense` seed 2.** Same script, same seed, same data file. The logged summary moved.

| Field | Archive jsonl | This rerun |
|---|---:|---:|
| afterA model | 0.28 | 0.13 |
| afterA mean_p_true | 0.182 | 0.0927 |
| afterA store / in_prompt | 0 / 0 | 0 / 0 |
| afterB model | 0.0 | 0.0 |
| afterB mean_p_true | 0.0001 | 0.0001 |
| copy_loss_last | 0.1183 | 0.1544 |
| sample_afterB i=2 pred | 10 | 7 |
| sample_afterB i=3 p_true | 0.0004 | 0.0001 |


afterB model is still 0. The dense mean afterA across the two seeds is 0.620 in the archive `(0.96+0.28)/2` and 0.545
here `(0.96+0.13)/2`. Soft-exact, paraphrase, and force means above do not use this row.

Wall clock did not reproduce for any run. The script’s `sec` on this host was about 31–37 s for exact-match (archive
17.7–26.6 s) and about 56–79 s for the retention rows that ran (archive 29.2–46.2 s). Wrapper wall times are in each
`*.meta.json`.

**Round 2 remainder: 13 of 18 original commands were not started.** The CPU budget was already past 30 minutes when the
next command would have begun. Those runs are not imputed from the five that finished.

**Round 1 `wonder.py baseline` seeds 0, 1, and 2 were not started.** No shortened 1500-step run was substituted. There
is no new baseline loss from this pass. The archive figure remains the archive figure: baseline mean validation loss
1.5736 on seeds 0, 1, 2.

A first launcher failed before any probe ran: `/usr/bin/time` is not on this host (`rc=127`). Those empty logs are kept
under `logs/exact-match-attempt0-missing-time-binary/`. The numbers above come from the second launcher, which timed
processes with `resource.getrusage`.


Side-by-side exact-match (archive jsonl vs this host)

`sec` is the only difference in the seven matched rows. Dense seed 2 is the mismatch.

| arm | seed | modelA orig | modelA rerun | modelB both | storeA/B | in_prompt B | copy_loss orig | copy_loss rerun |
|---|---:|---:|---:|---:|---|---:|---:|---:|
| exact | 1 | 0.82 | 0.82 | 0.05 | 1 / 1 | 0 | 0.0975 | 0.0975 |
| exact | 2 | 0.77 | 0.77 | 0.06 | 1 / 1 | 0 | 0.1293 | 0.1293 |
| paraphrase | 1 | 0.04 | 0.04 | 0.00 | 0 / 0 | 0 | 0.0975 | 0.0975 |
| paraphrase | 2 | 0.02 | 0.02 | 0.00 | 0 / 0 | 0 | 0.1293 | 0.1293 |
| force | 1 | 1.00 | 1.00 | 0.00 | 1 / 1 | 1 | 0.0975 | 0.0975 |
| force | 2 | 0.98 | 0.98 | 0.00 | 1 / 1 | 1 | 0.1293 | 0.1293 |
| dense | 1 | 0.96 | 0.96 | 0.00 | 0 / 0 | 0 | 0.0789 | 0.0789 |
| dense | 2 | 0.28 | 0.13 | 0.00 | 0 / 0 | 0 | 0.1183 | 0.1544 |


Full objects: `raw/comparison.json`. Per-run stdout is the script’s single JSON line. Stderr was empty on every
completed run.


Inputs actually on disk

`input.txt` next to the scripts is the file they open. It is not a synthetic-fact file.

- 1,115,394 bytes, md5 `6fb458f1232090904fb40fe944165e91`, sha256
`86c4e6aa9db7c042ec79f339dcb96d42b0075e16b8fc2e86bf0ca57e2dc565ed`
- Text begins `First Citizen:` / `Before we proceed any further, hear me speak.` This is the TinyShakespeare text named
in the November 3 results note.
- The word-level regex in both probes yields 250,371 tokens and 12,637 distinct surface forms. Each script keeps the
3,000 most common of those, then adds its own fact tokens.
- Synthetic facts are built inside the script with `random.Random(1234)`, not loaded from a second dataset. Exact-match
uses `NF=100`. Retention uses `NF=200`. Those lists differ because the shuffle length differs.
- Torch seeds are the argv seeds: `torch.manual_seed(SEED)` and `torch.Generator().manual_seed(SEED)`. The fact RNG is
fixed at 1234 and does not follow the torch seed.
- Character-level Round 1, which was not run, would read the same file without lowercasing: `open('input.txt').read()`.

Script hashes, identical in the archive and in the work copies:

| File | sha256 |
|---|---|
| `exact_match_probe.py` | `938e0a11ec9e03d91285c64b1e51732d176bfc6b07427815b0eda12e4283ea42` |
| `retention_probe.py` | `5789c0853f02336c6166b154ddfeb58739f3c06563655e80c0effd30ef8f3c2b` |
| `wonder.py` (copied, not run) | `5daaddbf9d712c40d1a4664882d99d41bda1f4d29f6f2ddcac9500bdd94ea30c` |
| `input.txt` | `86c4e6aa9db7c042ec79f339dcb96d42b0075e16b8fc2e86bf0ca57e2dc565ed` |



How the runs were invoked

Host: Linux, 4 CPUs, Intel Xeon family 6 model 207 under KVM, no CUDA. Python 3.12.3. torch 2.15.0+cpu, installed for
this pass from the public CPU wheel index. The archive does not record the original torch build. Default torch threads
on this host before the script runs: 4. Each original script then calls `torch.set_num_threads(8)`. The scripts were not
edited to match the 4 CPUs. That oversubscription is one plausible source of the dense seed 2 difference. This pass did
not isolate it.

No `PYTHONHASHSEED` was set. The probes hash with `hashlib.sha256`, not Python’s salted `hash()`.

CPU budget: 1800 seconds of child user+system time. Accounting is cumulative across the exact-match runs and the
retention runs.

| Block | Child CPU seconds | Wrapper outcome |
|---|---:|---|
| 8 exact-match runs | 831.624 | all finished, rc 0 |
| 5 retention runs | 1012.362 | all finished, rc 0 |
| Total | 1843.986 | next command not started |


Exact file `raw/cpu-used-exact.txt` is `831.624162`. Total file `raw/cpu-used-total.txt` is `1843.986236`. The retention
block is the difference.

The last retention command (`dense 2 4096`) began at 1670.737 CPU-seconds, under the cap, and was allowed to finish. It
ended at 1843.986. No running script was killed, and no step count was reduced.

Recorded settings inside `exact_match_probe.py`, not overridden: `CTX,BS,DM,NH,NL,LR = 64, 32, 128, 4, 2, 3e-3`;
`PRE,INTER,COPY = 200, 150, 200`; `NF,THR = 100, 0.05`.

Recorded settings inside `retention_probe.py`, not overridden: `PRE,INTER = 500, 500`; `NF = 200`; `THR = 0.05`; same
width, depth, context, batch, and learning rate. `run_ret.sh` passes the bucket count. Oracle still builds a `V*V`
table; the `4096` argument does not size that table. That is the original script.

### Exact argv, in archive jsonl order

python3 -u exact_match_probe.py exact 1
python3 -u exact_match_probe.py exact 2
python3 -u exact_match_probe.py paraphrase 1
python3 -u exact_match_probe.py paraphrase 2
python3 -u exact_match_probe.py force 1
python3 -u exact_match_probe.py force 2
python3 -u exact_match_probe.py dense 1
python3 -u exact_match_probe.py dense 2

Working directory for those eight: `work/exact-match/` (contains the script, `input.txt`, and the new
`results-exact-match.jsonl`).
### Retention argv that ran

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

Working directory: `work/retention/`.

Per-run files: `logs/exact-match/<arm>-seed<s>.stdout`, `.stderr`, `.meta.json`, and the same pattern under
`logs/retention/` with `-nb<NB>`.


Runnable remainder

Do not run these inside the archive tree. The scripts append. `work/retention/results-retention.jsonl` already holds the
five new rows; further retention commands append there. `work/round1-baseline-not-run/` has byte copies of `wonder.py`
and `input.txt` and no `results.jsonl`.

Retention commands not started, in `run_ret.sh` order: