# Gooo 0.6.15 public installation and execution

Observed on 2026-10-09 KST. Public release: [v0.6.15-dev](https://github.com/kimjooyoon/meta-ontology-go/releases/tag/v0.6.15-dev).

The source commit is `dc75f59fbee16e776dca13288efe1036d1bb5fab`, the normal squash merge
of PR 1396. Its tree equals tested head `9c5c1e37bf22da1ab8f6046d5b35c82085563438`.
Canonical PR CI 37836963810 passed all six checks. The original CI proof ZIP,
GitHub digest, provenance receipt and live PR tuple were checked before merging.
Readiness run 37840120443 executed that merged source on Linux amd64,
Windows amd64, macOS amd64 and macOS arm64; eligibility closed 7/7.
Publication run 37841064218 published the exact merged source.

Main promotion PR 1397 passed canonical CI 37840282154. The original proof ZIP
digest, provenance, exact six jobs, live dev/main tuple, source tree and sole
parent were rechecked. Expected-head normal squash merge produced
`2cdffde73b3792b93497301c32cef7103fa7988f`; its tree is the release source tree.
The original small main proof ZIP and extracted proof are retained here.

Workbench PR 39 passed CI 37842267646 at `3cd25a52`. Its original observation ZIP
(10,261,226 bytes, SHA-256 `f7e5031a7e2cde79c87fdbb2e27f86be1697f13663b0ebb3f4b11bea7df26564`)
was digest-verified. The exact compiler identity, five rounds, 13 attempts,
counterexample rows, all original/selected source files and construction cases
match the installed public-binary run. Final semantic evaluation is equal; the
hash of the whole raw input observation differs across independent executions
with their own timing/build observations. Those original hashes are preserved.
Normal expected-head squash merge produced workbench main `558b0158a6947cf2bf264599d271e43a8ed4be21`.
Its tree matches the tested head. Selected small CI records are retained here.

## Installed bytes

`release.json`, `tag-ref.json`, `tag.json`, the manifest and checksums bind nine
public assets to the source. The macOS arm64 archive and five metadata files
were downloaded and checked against GitHub asset digests. All eight checksum
entries were compared with public asset metadata; the other platform archives
were not downloaded again for independent byte verification. The archive contains
one `gooo` executable. Installed bytes match the extracted executable.
`archive-build.json` and `installed-build.json` match: Go 1.27.1, clean source,
version 0.6.15-dev, decision runtime v0.2.26-experimental.

## Actual public executable results

| Run | Saved program attempts | Rejected before native execution | Native combinations | Separate final cases |
| --- | ---: | ---: | ---: | ---: |
| Budget 3 | 3 | 2 | 1 | 1/4 |
| Budget 5 | 5 | 2 | 3 | 4/4 |
| Saved budget replay | 5 | 2 | 3 | 4/4 |
| Mixed, source order | 40 | 16 | 24 | 4/4 |
| Mixed, own model order | 5 | 2 | 3 | 4/4 |
| Saved mixed replay | 5 | 2 | 3 | 4/4 |

Native combinations are counted separately from child process executions; each
completed combination is executed twice. A replay preserves historical attempts
and executes them again. It makes zero new model predictions. The unchanged
model orders record choices once; source fills use deterministic source order.
The prediction observation is **38,167 ns** for this run. This is an in-process
prediction duration, not end-to-end build latency or a controlled speed benchmark.
The model's resident tensors occupy 2,096 bytes. Its all-data demonstration
training split does not establish accuracy on unseen programs.

The original final cases are retained. `observe.go` recounts the six compressed
outputs using the workbench reader and exact JSON numbers, including integers
above 2^53. `comparison.txt` also confirms equality with the frozen clean local
candidate: source IDs, plans, attempts, rejections, partial score, final values
and zero-inference replay. Both selected sources run as ordinary Gooo programs;
their two additional `body-compose` outputs are retained.

These are two related finite programs, four fresh constructions and two saved
replays. The model and fixed order choose identical final source. No expectations,
candidate space or failed attempts were removed. No new model was trained or
downloaded for these observations.

## Remaining limitation

0.6.15 continues after source-fill type failures and training-case evaluation
failures. A native execution error during caller evaluation still terminates
construction. The separate [caller-only failure record](../caller-native-failure-20261009/README.md)
preserves that counterexample and the next implementation plan.