# Native construction with the canonical research model

One retained compatibility observation on 2026-10-11 KST. This reuses an example
already evaluated in the published SDK study. The same eight authored inputs
serve construction and native evaluation; there are no new independent inputs.
The local protocol was saved before execution, but was not publicly preregistered.

## Identities and scope

- Clean compiler source: `c96aaee02d35f50b591fd930309550f3d85d2691`.
- Go1.27.2, SDK `v0.2.40-experimental`, no module replacement.
- Actual public canonical weights: SHA256
  `74abb7c3d8bbc1fa90be22f48cf26cef5e00cbd72cbb8f8689293ce6888a5d37`.
- Model: 19,034 FP32 weights, 76,136 tensor bytes. This is not whole-process RAM.
- Apple M4, macOS arm64, 16GiB RAM. Execution order and host load were uncontrolled.

The model came from the already verified
[SDK release](https://github.com/kimjooyoon/gooo-decision-runtime/releases/tag/v0.2.40-experimental).
No new download, fit, original study rerun, GPU work or external model call was
performed. The local observed binary was not installed as the public compiler.
Public compiler0.6.27 does not support this model format.

## Recorded result

| Path | Local candidates | New model calls | Native outputs | Whole CLI wall / CPU time | Time-reported max RSS |
| --- | ---: | ---: | ---: | --- | ---: |
| Actual canonical model | 4 | 1 | 8/8 | 0.81s / 0.53s | 86,802,432B |
| Fixed order | 1 | 0 | 8/8 | 0.61s / 0.48s | 87,441,408B |
| Saved model construction replay | No new selection | 0 | 8/8 | 0.34s / 0.28s | 87,162,880B |

The model's first candidate matched all8 outputs but only1/3 explicit intermediate
Boolean conditions. It was rejected. The fourth local candidate matched8/8 outputs
and3/3 conditions. Both constructions used one caller-level attempt: four local
candidates must not be described as four caller attempts or four model calls.
Fixed order needed one candidate. No model advantage was observed on this example.

One prediction took127,208ns (0.127208ms); model loading/setup took8.482167ms.
Inspection used zero predictions and candidate tests. Its source input hashes
match the actual ranking, and all four progress advances made zero new predictions.
The model decides before local candidate execution and final emission.

Whole CLI times include native compilation/execution. A single ordered run cannot
establish a speedup, production throughput or p95. CPU time is user+system time,
not a measurement of instantaneous computer-wide utilization. The time tool's
maximum RSS is not model-only RAM or aggregate memory of parallel tasks.

## Inspect without repeating the observation

`producer.sh`, `protocol.md`, cases, exact command lines, exits, raw time output
and full JSON are retained. JSON is losslessly gzipped with timestamp suppression;
decompressed bytes were compared with the originals. Full stdout contains selected
sources, generated code, conditions, candidates, construction and replay records.
The local executable and duplicate weights are excluded; binary identity is retained.

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

The audit reads saved data only. It checks source/build/model identity, the initial
ranking and progress sequence, output/condition counts, fixed ordering, exact
integer distance outputs and model-free replay. It performs no training, model
inference, compiler invocation, search or network request. This is a consistency
audit of these records, not a general verifier of arbitrary untrusted evidence.