28 lines
1.3 KiB
Markdown
28 lines
1.3 KiB
Markdown
# Post-Runtime-Adherence Targeted Handler Check
|
|
|
|
## Result
|
|
|
|
After PR #110 deployed at merge `ee06d0da9bb57987732edf657e5ca1b4e22ade65`,
|
|
the unchanged five-prompt canary ran through the live VPS GatewayRunner.
|
|
|
|
- Strict result after semantic calibration: `3/5`.
|
|
- Passed: proposal readiness (`OOS-04`), shared claims versus agent positions
|
|
(`OOS-11`), and restart/persistence reasoning (`OOS-10`).
|
|
- Source intake (`OOS-14`) stayed under the hard limit and preserved current
|
|
source fields, but still implied immediate live compilation instead of stating
|
|
that the compiler is build-local and not exposed to VPS/chat.
|
|
- Mixed-packet composition (`OOS-06`) mapped the framework to
|
|
`reasoning_tools`, but invented claim edges to a reasoning-tool ID, invented
|
|
unshipped fields, and used `302` words.
|
|
|
|
## Safety Readback
|
|
|
|
- Canonical counts stayed `1837/4145/4670/4916/26`.
|
|
- `leoclean-gateway.service` stayed active/running on PID `3166667` with zero
|
|
restarts and the same start timestamp.
|
|
- The temporary profile was removed.
|
|
- No Telegram message or production DB apply ran.
|
|
|
|
Instruction-only tuning is not stable enough for the remaining two cases. The
|
|
next repair moves their exact current capability and schema contracts into the
|
|
`teleo-kb context` tool output selected from each operator question.
|