# Local Gooo 0.6.9 candidate execution

Observed 2026-10-08 on local macOS arm64. Compiler source
`e0d046503939883330eb37f998a2e0e8fc5154e5`, clean VCS, Go 1.27.1 and public
decision runtime v0.2.26-experimental. PR #1376 prepares the release. These files
record a local candidate; they do not establish publication or hosted readiness.
The receipt's `macos-15` is the fixed target profile, not this machine's OS claim.

The native platform witness built matching binaries and matching archives twice.
The actual candidate binary constructed and replayed division 8/8, retry 12/12
and filenames 12/12. The v3 source context matched source/context digests, decoded
and projected through the public SDK, with zero predictions and candidate tests.

## Existing own-model dogfood

The separately executed filename runs use the same 5 source selection cases
containing 15 fields and the separate 12-case native suite. `recount.json` counts
the actual/expected JSON values again and checks reported field/case flags.

| Run | Attempted candidates | Selection cases | Selection fields | Native cases |
| --- | ---: | ---: | ---: | ---: |
| Fixed order | 8 | 5/5 | 15/15 | 12/12 |
| Existing v1 model | 1 | 5/5 | 15/15 | 12/12 |
| Saved model composition | retained | 5/5 | 15/15 | 12/12 |
| Fixed order, one-attempt budget | 1 | 0/5 | 5/15 | 0/12 |
| Saved partial composition | retained | 0/5 | 5/15 | 0/12 |

Both saved compositions preserve their generated-program digests. Replay retains
the original construction receipt; its historical model call is not a new call.
Runtime and saved replay use zero new model calls. The partial source changes
only the attempt budget from 8 to 1 and records `PARTIAL_FINITE`.

Weights are unchanged from the public
[v1 shared-field model](https://huggingface.co/asketeddy/gooo-record-shared-field-tiny-v1/tree/5800946afb35506d66357ee3ea6f956f506795b9/models/qat_ternary).
Metadata SHA256 `f95925a277c45957722bc116213c61bee997cc07a94adae738a05d218d660059`;
weights SHA256 `049081cc6dc62f3c4230f72e601c1fd421c514fce9cfdf7bc1eb412ca4c5514b`.
Contract `triple_record_field_context_v1_shared_v1`; resident tensors 2,096 bytes.
This is no new v3 training. Earlier eight-arrangement testing found this same
v1 model accepted the complete combination first in 1/8 arrangements. This run
uses the original successful arrangement; it cannot establish general accuracy.

## Single-run cost

Model preparation: 0.58125 ms. Initial prediction: 12,334 ns (12.334 microseconds).
The complete model command, including native Go build/execution: 0.36 s wall,
0.16 s user CPU, 0.11 s system CPU, maximum RSS 85,901,312 bytes.
CPU time divided by wall time is about 75% of one core, including child work.
Host-wide CPU utilization change was not measured. Fixed-order command: 0.67 s
wall and 84,983,808 bytes maximum RSS. Cache warmth and background work were not
controlled, so these single observations do not establish a speedup.

## Reproduce

Use the exact compiler source above with Go 1.27.1. Use a new output directory.

```sh
gooo body-compose --entry Classify --source filename.gooo --cases cases.json --out fixed
gooo body-compose --entry Classify --source filename.gooo --cases cases.json --model /path/to/model.json --out guided
gooo body-compose --source guided/original.gooo --cases cases.json --composition guided/composition.json
gooo body-compose --entry Classify --source filename-one-attempt.gooo --cases cases.json --out partial
go run observe-composition.go fixed.json model.json replay.json partial.json partial-replay.json
```

The raw `.time` files use macOS `/usr/bin/time -l`; generated programs and source
contracts are retained inside the observation JSON. Run the recount command
inside this directory. Each finite result describes these declared inputs.