# Published Gooo 0.6.8 on macOS arm64

These observations execute the downloaded public release binary, whose source is
`86b182da4543efa24ea5bcb504e7b13ea1449c39`. All nine downloaded assets were counted,
all eight checksum entries matched, and all seven manifest payload hashes and
sizes matched. The annotated tag resolves to that exact source. The installed
binary has the same bytes as the downloaded macOS arm64 executable.

| Execution | Construction cases | Construction fields | Candidates tried | Native cases | New model calls |
| --- | --- | --- | ---: | --- | ---: |
| Retry, fixed order | 5/5 | 15/15 | 8 | 12/12 | 0 |
| Retry, own model | 5/5 | 15/15 | 1 | 12/12 | 1 |
| Retry, saved replay | retained 5/5 | retained 15/15 | 0 new | 12/12 | 0 |
| Retry, one fixed attempt | 0/5 | 8/15 | 1 | 0/12 | 0 |
| Partial retry, saved replay | retained 0/5 | retained 8/15 | 0 new | 0/12 | 0 |
| Division, fixed order | 4/4 | 12/12 | 8 | 8/8 | 0 |
| Division, own model | 4/4 | 12/12 | 1 | 8/8 | 1 |
| Division, saved replay | retained 4/4 | retained 12/12 | 0 new | 8/8 | 0 |
| Installed retry replay | retained 5/5 | retained 15/15 | 0 new | 12/12 | 0 |

Fixed and model construction produced the same generated Go per example. Complete
and partial saved replay preserved their respective generated code. Runtime
observations make zero new model calls. Replayed construction metadata retains
the original model prediction and must not be counted as a new prediction.

The unchanged shared QAT weights have SHA-256
`049081cc6dc62f3c4230f72e601c1fd421c514fce9cfdf7bc1eb412ca4c5514b`.
Prediction took 46.708 microseconds for retry and 27.125 microseconds for division;
model setup took 1.154208 and 0.273708 milliseconds. Retained tensors were 2,096
bytes. Whole commands took 0.34 and 0.35 seconds with maximum RSS 87,621,632 and
86,523,904 bytes. The raw BSD time files include user and system CPU time.
CPU/wall ratios are approximately 76.47% and 71.43% relative to one core, including
child work. Host CPU increase is unmeasured. These are single sequential runs
with uncontrolled cache and host load, not a general speed comparison.

## Recount or replay

With Go 1.27.1, recount exact actual/expected JSON values without converting large
integers to floating point:

```sh
go run observe-composition.go retry-fixed.json retry-model.json retry-replay.json \
  partial.json partial-replay.json division-fixed.json division-model.json \
  division-replay.json installed-replay.json
```

With the published Gooo executable:

```sh
gooo body-compose --source retry.gooo --cases retry-cases.json \
  --composition retry-model/composition.json
gooo body-compose --source partial.gooo --cases retry-cases.json \
  --composition partial/composition.json
gooo body-compose --source division.gooo --cases division-cases.json \
  --composition division-model/composition.json
```

Sources and cases are copied from the release source; `partial.gooo` changes only
the retry attempt budget from eight to one. The local installation and this
recount are same-author observations. Other inputs and model-training exposure
remain unmeasured. No model training was performed.