# 0.6.21 candidate: reproducible own model and a Windows timing failure

Candidate source `5e1127b6e9a55f47d1cc8d8220eb888879de2953` is submitted in
[PR 1419](https://github.com/kimjooyoon/meta-ontology-go/pull/1419), targeting dev
`832dbf3684b104b03101f41beb8f2c5efc517ec9`. It includes the existing own QAT
model's metadata and 446-byte weights, with their original provenance. No model
was downloaded or trained for this candidate.

The local Darwin ARM64 native release profile returned PASS/EXACT for the clean
candidate, Go 1.27.2. Eight original reports cover Korean/English entity names,
fixed/model assembly and saved replay. The reports preserve exact integers and
source/canonical identities. The full CLI and release race suites also passed.
The logical runner label in the local receipt describes the requested profile;
it is separate from GitHub's hosted platform executions.

[Original readiness run 37906275445](https://github.com/kimjooyoon/meta-ontology-go/actions/runs/37906275445)
completed with failure on original attempt 1. Linux AMD64, Darwin ARM64 and Darwin
AMD64 platform jobs succeeded. Windows AMD64 failed and release eligibility failed;
no successful aggregate, publication or release approval is claimed for this head.

The [Windows job](https://github.com/kimjooyoon/meta-ontology-go/actions/runs/37906275445/job/113740393314)
retained `model_calls: 1`, a prediction/ranking and `fields_passed: 12` /
`fields_total: 12`, with measured `predict_ns: 0`. The assembly checkpoint rejected
that report with `record model call observation differs`. Its replay verifier
required positive measured duration. The original failed report is retained here;
zero-duration timing is being addressed separately from prediction identity and
finite correctness. Original canonical CI 37906275575 completed successfully and
its original PASS proof was verified. PR 1419 was then closed with the native
failure retained; corrected PR 1420 carries the actual fix and new source-bound checks.

All values describe finite examples. Model training exposure of these runtime
cases is unknown. Candidate order and model timing do not establish general
accuracy or a speed advantage. Public installed compiler remains 0.6.20.