# Typed model preflight: original CI, normal merge and installed use

Observed 2026-10-10 KST. PR1433 original corrected-source CI
[37995667254](https://github.com/kimjooyoon/meta-ontology-go/actions/runs/37995667254),
attempt1/check suite102952953017/sourcebff57e7e, completed SUCCESS normally.
All six canonical checks and CI proof belong to that exact suite/source. Its
original watcher closed exit0. Readiness37995667634 had already succeeded; its
original source-bound evidence is in the preceding bff readiness bundle.
No original run was canceled or rerun.

Original proof artifact11647174336 was 11,182 bytes. API ZIP SHA256
`d575b041ecb6c6ad1a57d488f925eb2bb4692b9395bf581a832fda833bbf3ad2`
matched before extraction. The v5 PASS proof bound headbff, basedev035,
PR1433, run37995667254/attempt1/checkoutbff. The local proof verifier with
--require-pass succeeded silently (its retained stdout is empty); the
set-e normal merge procedure finished exit0. Frozen PR body bytes were equal
before and after the expected-head normal squash merge.

Accepted dev source `426caecb711e47da26fe659237a117f301d112f4` has sole parent
035e4447 and exactly the bff candidate tree. A clean Go1.27.2/SDK0.2.26 build was
installed separately as gooo-typed-dev. Actual installed binary SHA256:
`9f9b66a1e7ed54f11679006aca1a581595c8558f005b568d9ff5fcc20cd34a2d`.
The public v0.6.23/source2b assets and tag were unchanged.

## Fresh installed-producer observation

The existing own 446-byte weights / 2,096-byte resident tensors were reused.
Inspection made 0 predictions/0 candidate tests; model identity, record input
text/digest and typed decline reason matched actual construction. Record
inference was 1 call/31,041ns, typed inference was 0. Initial source-local cases
were 2/2 while the first caller was 0/1. Program construction used 11 attempts
within budget16, then separate final inputs were 3/3. Saved replay was 3/3 with
0 fresh predictions. Exact input9007199254740993/output27021597764222979 were
checked as Go json.Number strings, not floating-point conversions.

Unary inspection model load was0.412708ms, total0.791ms, with 0 repository writes
and no selected emission. External CLI wall time/maximum RSS were6.13s /
87,457,792 bytes for construction, and3.42s /88,014,848 bytes for replay.
These are single observations; host CPU utilization is UNMEASURED and training
exposure is UNKNOWN. They do not establish model accuracy or an efficiency
improvement. Local cases, initial caller and separate final inputs have distinct
denominators. Earlier C1/C3 data retain their original producers.

The Go verifier checks actual model metadata/weights/schema/feature/arithmetic,
explicit counters, input parity, runtime producer and exact integer delivery.
A fabricated rounded-input copy was rejected with exit1, without any generation
or model calls. Only its failure text/exit are retained; the temporary copy was
removed. The actual original records were unchanged.

Accepted push readiness37997779793/attempt1/suite102958674669/source426
completed SUCCESS normally across four platforms; its original watcher closed
exit0 and terminal metadata/jobs/watch are retained. Accepted push CI37997779830
also completed SUCCESS normally: attempt1/suite102958674785/source426.
All six canonical checks and CI proof bundle passed in that exact suite; its
original watcher closed exit0. Terminal API/jobs/check suite/watch were appended
after the earlier publication. No original run was canceled or rerun.
This bundle contains the completed candidate and accepted CI, installed
observation and accepted readiness. Public v0.6.23/source2b remains unchanged;
the next release identity is separate source work.

Every gzip -n payload was decompressed and byte-compared with its original.
The ZIP and Go verifier bytes were copied unchanged. FILES.sha256 binds the
saved files. No model weights, executable or large platform ZIP was added.