Skip to content

test(overnight): word bias for V02, accept a plain 22 on K02 - #214

Open
BrettKinny wants to merge 1 commit into
mainfrom
agent/v02-k02-eval
Open

BrettKinny wants to merge 1 commit into
mainfrom
agent/v02-k02-eval

Conversation

@BrettKinny

Copy link
Copy Markdown
Owner

What

  • V02-repeat gets a reply_hotwords list, passed to both reply transcriptions through a new hotword_args helper.
  • K02 also accepts a plain "22" (maintainer decision 2026-11-02). "22 room" stays forbidden.

Evidence

Offline only, using the three saved V02 replies from the overnight run. No new robot run.

Check Without bias With bias
Saved reply 1 Apple Robot 7, window open honor purple, robot, seven, window, banana
Saved reply 2 Huckle robot seven window bananas purple, robot, seven, window, banana
Saved reply 3 Couple robots seven window banana couple robot, seven, window, banana

Negative controls with the bias: three other cases' replies and a silent crop were not altered. Removing the first word entirely still shows up (cut at 2.0 s).

Known limit: please read before merging

With about 0.5 s cut from the start of a reply, the bias still wrote "purple". Clipped reply starts are the #200 part 2 bug, so the misreads may be partly real audio loss rather than evaluator error, and this bias would hide that on V02. Run V02 unbiased on the agent/fw-reply-head-clip firmware first. Saved reply 3 still fails on "couple", so this may not get V02 to three passes in a row.

Tests

tests/test_dotty_overnight.py and tests/test_dotty_overnight_states.py: 179 passed.

AI-assisted: written by Claude Opus 5.5; needs human review before merge.

🤖 Generated with Claude Code

V02's reply is a bare word list, so the reference mic's Whisper had no
context and returned "Apple", "Huckle" and "Couple" for "purple". Bias that
case toward its own words. Re-transcribing the three saved V02 replies gave
two exact matches and one "couple robot". Other replies and silence were
unaffected, and losing a whole word still shows.

The bias does fill in a clipped first word, so V02 stops detecting up to
about 0.5 s of lost reply audio (#200).

K02: the maintainer accepted a plain 22 for this release. "22 room" stays
forbidden. Without this the case quarantines and can pause a run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant