# Corrected 0.6.21 candidate: observed 0ns prediction now replays

Source `9870e644aa35f92bccdd5b5e6546a4dd39c8e0f3`, [PR 1420](https://github.com/kimjooyoon/meta-ontology-go/pull/1420).
Original readiness run [37908235016](https://github.com/kimjooyoon/meta-ontology-go/actions/runs/37908235016),
attempt 1, completed successfully. All four native platform jobs and the two
eligibility/aggregate jobs passed. Original final artifact 11605262551 is 1,047
bytes, SHA256 e401b2ef309451e673a21f89cd45dbe2f9ec9956e9483f26ab07e11006303ce4.
Its verified result is EVIDENCE_CLOSED / EXACT, 7/7 cells, unknown 0, refuted 0.
Canonical CI for this corrected source is still running; dev/main merge and
publication are not claimed. Installed public compiler remains 0.6.20.

The original Windows artifact 11606016184 is 14,278,743 bytes, SHA256
34d526e465e1d11c6a3da8cc20824eb09cd7278085fa355c77a568bd557c00ab.
The complete original ZIP was verified locally against its GitHub digest and size.
Only its eight scalar reports, platform receipt and metadata are copied here;
compiler archives remain in the Actions artifact.

The Korean Windows model construct report retains `predict_ns: 0`, model_calls 1,
a captured prediction/ranking, 12/12 authored field checks and native 4/4. Replay
also passes 4/4 with zero new model calls. English and fixed/model variants pass
the native profile. The correction retains the measured zero time; it requires
nonnegative duration and validates actual calls, prediction, source context and
probability/ranking evidence. The TDD red log reproduces the original rejection.
Five invalid inference mutations remain rejected by the new race regression.

These are finite case and platform observations, not general model accuracy or a
latency advantage. Training exposure of native cases is unknown. Existing model
metadata and 446-byte packed weights were reused without downloads or training.

## Subsequent closure

Original canonical run 37908235019 completed successfully. Proof artifact
11606635561, 11,185 bytes, SHA256 c58b8e5a91b7355d741ab06cb38d34e0c585bc865822129239d85bd0ee471e68,
was verified with its source/base/PR tuple and frozen body. PR 1420 merged normally
into dev d4d0966f on 2026-10-09 09:24:28 UTC; the merged tree equals candidate9870e644.
The accepted source's own push readiness then passed, and public0.6.21 was published
and installed. Its separate exact original evidence and installed-command reports
are in ../release-0621-published-20261009 . The initial candidate observations above
retain their producer source and original times.