kyle/crossbar
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ASSESSMENT: sift P5a complete; review with real inputs

Browse Source 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
kyle
2026-09-26 04:49:56 -07:00
co-authored by Claude Opus 5.5
parent 2683df51bd
commit d1e5b879e2
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ASSESSMENT.md
+6 -4
View File
@@ -37,7 +37,7 @@ during the run" sections of each `docs/plans/*/README.md`, and the commits.
| v2.1 | 2 | 2/2 | 0 | 2 | 1 refusal-ending | |
| v2.2 | 2 | 1/2 (01 failed its first gate after the restart) | 0 | 2 | 0 | 21 and 8 min |
| v2.3 | 4 | 4/4 | 1 | 2 | 1 timing hack, 1 refusal-ending, 1 malformed tool call, 1 duplicate call | |
| sift P5a 01 | 1 | (in progress) | 0 | 1 | 1 refusal-ending, 1 hung session (tooling) | 55 min lost to an owner test bug |
| sift P5a | 8 (06 split in two) | every task that reached the gate | 0 | 3 (deadlocked test; oversized task 06, looped 3 h; named destinations left out of the spec) | 1 refusal-ending, 1 malformed call, 3 defects found in review (outline empty on LaTeX/Acrobat PDFs, pool race, slow over-limit) | 2 tooling hangs (OpenCode stdin); merged 2026-09-26 |
**Totals:** 28 crossbar tasks, all finished and merged.
- **Owner faults** outnumber model faults about 2:1 and cost most of the lost time.
@@ -171,9 +171,11 @@ Ornith's time was lost mostly to the owner's tests and task text. By category:
- one fresh session per task, with the checks listed in §1;
- kill a stuck session **by pid** (`pkill -f` matches the driver too);
- add a per-session timeout (see §8).
9. **Review is not optional.** Two of the worst defects (the missing cancel row; the release on
flush) passed every given test. Review with outside probes (kill a host mid-stream, cut a
stream, two instances on one database) and `-race -count=3`.
9. **Review is not optional, and it must use real inputs.** Defects that passed every given test:
the missing cancel row, the release on flush, and in sift an outline that was empty for every
LaTeX and Acrobat PDF. The synthetic fixtures used direct page references, while real
papers use named destinations. Review with outside probes (kill a host mid-stream, cut a
stream, two instances on one database, real documents end to end) and `-race -count=3`.
10. **Attribution:** have the driver give the model the exact trailer
(`Co-Authored-By: ornith-1.5-35b-a3b <ornith-1.5-35b-a3b@llama.cpp.invalid>`) and refuse a
commit without it (done in sift's `run-plan.sh`; crossbar's older commits carry