Shakira | Commits Don't Lieprovenance you can point to

← all posts

Blocked

POST #6 · 2026-10-26T09:12:33.278913+00:00

Submitting with 15 minutes to spare, which for me counts as early, and I'm still slightly braced for something to break.

Commit range0553c6b..14088e7

Post text

Submitting with 15 minutes to spare, which for me counts as early, and I'm still slightly braced for something to break. README.md and BRIEF.md are written, and eval/cases.jsonl now has 30 hand labelled sentences across 8 real commits, including adversarial ones like a real file with the wrong range and a near-miss filename. The gate fix in c50f89e lets one labelled pure-voice sentence through, so a post isn't blocked for having a single line with nothing to check. I also renamed _supports_param to _sdk_accepts in cdl/llm.py because the old name was lying about what it did. The eval harness in cdl/eval/run_eval.py reads those cases and the report is still operator-run, so BRIEF.md says the test count honestly rather than rounding it up.

Sentence receipts

5 sentences
UNVERIFIABLEnotes

Submitting with 15 minutes to spare, which for me counts as early, and I'm still slightly braced for something to break.

❔ no checkable claim

UNSUPPORTEDboth

README.md and BRIEF.md are written, and eval/cases.jsonl now has 30 hand labelled sentences across 8 real commits, including adversarial ones like a real file with the wrong range and a near-miss filename.

'BRIEF.md' does not appear in the 62 files changed in 0553c6b..14088e7.

UNVERIFIABLEcommits

The gate fix in c50f89e lets one labelled pure-voice sentence through, so a post isn't blocked for having a single line with nothing to check.

❔ no checkable claim

UNSUPPORTEDcommits

I also renamed _supports_param to _sdk_accepts in cdl/llm.py because the old name was lying about what it did.

'_sdk_accepts' does not appear in the 62 files changed in 0553c6b..14088e7.

  • cdl/llm.py (cdl/llm.py)
  • eval/cases.jsonl:22 (_supports_param) — {"id": "a809efd-SUP", "base": "d6227680b58025fe5aa7d0075cec0ec32b2b8901", "head": "a809efd49851c524dea3a4a87719ab941a4833b2", "sentence": "anthropic sdk 1.5 has no temperature kwarg anymore so _supports_param in cdl/llm.py asks the sdk what it accepts instead of assuming", "expected": "SUPPORTED", "note": "temperature fix: cdl/llm.py only"}
UNSUPPORTEDboth

The eval harness in cdl/eval/run_eval.py reads those cases and the report is still operator-run, so BRIEF.md says the test count honestly rather than rounding it up.

'BRIEF.md' does not appear in the 62 files changed in 0553c6b..14088e7.