# Queue batch rule-change observation

See the [usage and adoption guide](https://github.com/kimjooyoon/meta-ontology-go/wiki/Small-Workflow-Pilot).
This directory contains a synthetic example, not customer or production data.
Sources, cases and protocol were committed before the six observed commands at
wiki commit `ce4413ddd01caf15b3dc8481f06cacd574adf8bc`.

## Producers

- Public Gooo `v0.6.26-dev`, source `8a759cba557b6c6e22aeea57040e8e1c69e6ec68`.
- macOS arm64 binary SHA256: `77c3f1a332df6ae265f66334e500213438025fd335638b463ee21622897f5ef5`.
- Native toolchain: Go1.27.2, Apple M4, 10 logical CPUs, 16GiB RAM.
- Released choice model: SDK0.2.37, SHA256
  `aa81790884cd21ac9dd435fe5cf14b604cdd658f514f6460a526fc858ae07ab1`.
- No new training, GPU, model download or remote model call. Weights are reused
  from the earlier published model, whose training scope is documented there.

## Exact command sequence

For limit16, then limit32, run these commands once each with `GOWORK=off` and
`GOTOOLCHAIN=go1.27.2`. `gooo` denotes the public binary above; `model-choice.json`
denotes the verified existing model file. Each command's stdout, exit and
`/usr/bin/time -l` stderr are retained under `results/`.

```sh
gooo body-construct --source batch16.gooo --entry Main \
  --construction-cases construction.json --cases evaluation16.json --attempts 4 \
  --out batch16-deterministic
gooo body-construct --source batch16.gooo --entry Main \
  --construction-cases construction.json --cases evaluation16.json --attempts 4 \
  --model model-choice.json --out batch16-model
gooo body-construct --source batch16-model/original.gooo \
  --construction batch16-model/construction.json --cases evaluation16.json
```

The second group substitutes32 for16 in filenames; the `attempts` budget stays4.
The public model runs' generated Go, source, full construction and evaluation
receipts are included. Standard output for all six commands includes full
candidate records. JSON uses deterministic gzip, and decompressed bytes were
compared with the original files. No native binaries or duplicate weights are
included. The local producer script used absolute paths only for the installed
compiler and existing weights; the command examples use portable names.

## Read the evidence without executing a model

From this directory in a downloaded wiki checkout:

```sh
shasum -a 256 -c SHA256SUMS
GOWORK=off GOTOOLCHAIN=go1.27.2 go run audit.go .
```

`audit.go` reads plain or gzipped saved JSON. It independently evaluates the
recorded masks and exact integer outputs, checks source and model identities,
compares each native output with clamp(input,0,limit), checks declared condition
observations, and distinguishes original model calls from replayed history.
It performs zero model predictions, compiler calls, candidate searches or fits.
Its output is `audit.json`. This is a small consistency audit, not a proof of
all runtime behavior or of arbitrary malformed evidence rejection.

Both limits complete10/10 native outputs in each mode. Both models initially
propose an incorrect body; local checking completes on the second candidate,
versus the third under the declared deterministic order. Saved replay makes
zero new predictions. The `recorded_local_attempts` in replay rows describe
the original construction history, not fresh candidate selection.

The ten native inputs overlap some source cases and are repeated across modes.
They must not be counted as60 independent evaluation problems. The intentionally
incorrect initial recipe makes the selection mechanism visible; manually
writing this fixed clamp in Go is a straightforward baseline. Whole-command
timings include compilation and are affected by fixed order, caches and load.
This observation establishes runnable composition, rule change and replay only.

`adoption-status.json` keeps business measurements explicitly unobserved. Future
observations should identify their workflow, baseline, sample, source revision,
timing boundary and actual cost assumptions before claiming improvement.