# Gooo 0.6.10 local candidate observations

Candidate source: `a22eb0823ed0923e3bd73c99c2ab3ae83ab60d57`, clean checkout,
Go 1.27.1, SDK v0.2.26-experimental. The subsequent PR head `63a39565` changes
only `.github/workflows/gooo-release-publish.yml` release-note text. The local
observations belong to a22eb082; remote CI is tracked separately for the new head.

Local `go vet ./...` passed. Version/smoke tests passed, and the complete affected
packages passed in serial: toolchainrelease, toolchaincli, workspaceexecution
and cmd/gooo. These local checks are narrower than whole-repository CI.

The actual platform witness built the candidate twice, compared binary/archive
bytes, executed the CLI and native language examples, and passed. This was a
local macOS arm64 run of the fixed darwin-arm64 target profile, whose configured
runner label is macos-15; it was not a hosted GitHub runner observation.

`platform-observations.tar.gz` retains the receipt and raw runs, excluding the
temporary binary archive. Native division passed 8/8, retry 12/12 and filenames
12/12, with saved replay. Graph export performed no predictions or candidate
evaluations. The new package path returned the expected source/stem/byte values
for all three actual inputs; saved replay retained the same generated program.
Input-only receipts remain OBSERVED with finite_total=0 and finite_passed=0.
Runtime and saved-replay model calls were zero.

## Existing-model dogfood on this candidate

The public workbench at `8ed738d15e2d60bee21ee1d6d6ed29041989df61` reran the frozen
dependent Splice study with this executable. `splice-evidence.tar.gz` retains
all source variants, predictions, native values and replay. `splice-summary.json`
contains exact source/model hashes and build metadata.

At budgets 1/2/4/8, full output matches were:

| Mode | 1 | 2 | 4 | 8 |
| --- | --- | --- | --- | --- |
| Fixed | 321/542 | 321/542 | 321/542 | 542/542 |
| Unchanged graph QAT | 321/542 | 321/542 | 542/542 | 542/542 |

The model's complete combination remained fourth, versus eighth in fixed order.
First choice failed and all partial outcomes were retained. Every saved replay
matched its original actual values. Full programs matched 1,626/1,626 fields.
Both modes also edited a Gooo function from input + 1 to input * 2, compiled its
returned source and passed three separate integer inputs.

This is release regression evidence on an already observed program. It adds no
new program family or independent model-generalization result. Model weights
and the frozen study plan were unchanged. Publication and installation of the
release executable require separate evidence after CI and merge.