# Gooo-driven caller construction loop

2026-10-08. Workbench PR 30 merged as
`ce4cd166e886cfd0bd3c8f8557709ba18d5639f2` after native CI at
`d5d3e8eafae293d78bb94c596cb117d6d2af97e1`. Their file trees match.
Documentation PR 31 then passed CI and merged as
`e1e35fd88ec7805e5b9efc78b24460cdc9a401c4`, updating the first-run guide to
published Gooo 0.6.12. Its premerge and merge records are retained here.

Gooo source `recipes/joint-next.gooo` computes the next action and program budget.
The Go dispatcher freezes source/expectation bytes, executes the proposal inside
explicit ceilings, and retains every fresh run. Budget growth currently restarts
construction. A four-round fixed-order loop performs 1+2+4+8 = 15 whole-program
attempts; the eight-attempt final round is only part of that work.

## Captured runs

All runs reuse one existing three-choice example. Seven evaluation inputs contain
one consumed caller root and six other roots; zero is also a helper-local example.
The independently trained graph QAT model is unchanged. Its metadata SHA256 is
`3c68205a660695103712e2115ce90accd298b406dced8ad5ad5c9b84bab0a202` and weights SHA256
is `76f68845a03ed8c8bc261a57c919e96dcdd352d3c35bcdaeb3d9b51a88975c6f`.

| Environment | Fixed rounds / program attempts | Model rounds / program attempts | Final evaluation |
| --- | --- | --- | --- |
| Local clean compiler 2246ec35 | 4 / 15 | 1 / 1 | 7/7 both |
| PR CI compiler 4a0cfdd8 | 4 / 15 | 1 / 1 | 7/7 both |
| Published macOS arm64 0.6.12 / 4a0cfdd8 | 4 / 15 | 1 / 1 | 7/7 both |

The selected Gooo source is identical in all six loops. These repeated runs do
not add new program families. Each model loop uses one initial prediction; fixed
loops use zero. Saved model replay reproduces seven outputs with no new prediction.

On the published binary the summed construction-command wall time was 7.952 s
for fixed order and 0.665 s for model order. These include compilation/execution
and exclude separate diagnostic commands. The model prediction was 25.084 µs.
Workload, cache and platform affect the readings; this is a known-example
observation. Host CPU utilization was not sampled. The recorded tensor storage
is 2,096 bytes, separate from process/runtime memory.

## Original evidence and recounting

- `local-observations.tar.gz`: original local loop outputs, saved programs,
  diagnostics and model replay, using clean compiler 2246ec35.
- `ci-observations.tar.gz`: selected original `ci-joint-fixed`, `ci-joint-model`
  and compiler identity files from successful PR run 37789082485. Original
  artifact 11554854725 was downloaded, SHA256-checked, selectively extracted,
  and removed locally. Original ZIP SHA256:
  `20992cef3aae68c5d6c52d207d3182369fe1363aa5a090d1901e81562c5f17e5`.
- `published-observations.tar.gz`: fresh installed public-binary runs, saved
  replay and Gooo diagnosis of that replay.
- `ci-run.json`, `merged.json`: source/run/merge identities.
- `observe.go`: standard-library-only Go reader; exact JSON numbers, separate
  expected-case files, every round's output counts, proposed budgets, prediction
  counts and final source equality. `comparison.txt` is its captured output.

Extract the three archives into three distinct directories, then run:

```sh
go run observe.go /path/to/local /path/to/ci /path/to/published
```

This recounts captured values. The workbench's native tests and saved compiler
replay execute programs separately. Nine Gooo diagnostic branches include
remaining program choices, exhausted choices, unresolved local obligations even
when evaluation passes, evaluation counterexamples, missing expectations and
consumed-only evaluation roots. The bounded dispatcher executes budget growth;
other actions retain their reason for subsequent work.