Fix crashes and hangs in seven evaluation scripts
Seven evaluation scripts in evaluation_scripts/2025_10_23/ raise an exception, or never return, in situations that real evaluations reach, and the evaluation then records no result for the answer. In six of them the trigger is an answer that lists fewer items than the task asks for. With this PR those answers are scored: the missing items score 0 through a convention the scripts already use, placeholder objects whose failed existence checks skip the dependent verifications. In research_car.py the trigger is the ground truth, which the script extracts from the automakers' websites while it runs; the script now stops with an error that names the brand and the page.
| Script | Failure before this PR |
|---|---|
buy_monitor.py (dev set) |
TypeError on every answer: a loop iterates over an int instead of a range |
oscar_A24.py |
Never returns when the answer lists fewer films than expected: the padding loop does not advance its counter |
apple_watch.py |
ValidationError when fewer items than required are extracted: Optional fields without a default are required in pydantic v2 |
coursera_uk.py |
IndexError when the answer has no item of a category |
esport_player.py |
AttributeError for padding placeholders, which lacked the nested objects that extracted items have |
largest_city_flights.py |
AttributeError for padding placeholders, which have no expected date |
research_car.py |
AttributeError when a brand's ground-truth page yields no sedan models, or a model without a name |
buy_monitor.py is also part of the public dev set in the GitHub repository, which receives the identical one-line fix.
oscar_A24: placeholder films
The padding loop now advances, so an answer with fewer than the 11 expected films gets one placeholder film per missing film. A placeholder fails its identification check without a judge request, and since each film node is SEQUENTIAL, its other checks are skipped. Without that guard, each placeholder would ask the judge whether a film named "None" is each of the 11 expected films and then verify it on the web: with a judge that rejects every claim, an answer that lists no films would take 264 verification requests, where it now takes none.
research_car: missing ground truth
research_car.py extracts each brand's sedan lineup from the automaker's website during the evaluation, and a page that cannot be captured yields no models. The released script then raises AttributeError, and so does a model extracted without a name. With this PR, models without a name are left out of the ground truth, and a brand without any named model stops the evaluation with:
RuntimeError: No Mazda sedan models could be extracted from the ground-truth page https://www.mazdausa.com/vehicles, so the answer cannot be scored; evaluate it again once the page can be loaded.
The answer is then left without a result, as with the released script, and the evaluation runner evaluates it again on its next run. The error is raised before any verification of the answer, so no judge request is spent on it.
Decision point. The alternative is to score the answer without that brand's ground truth, through a skipped placeholder node, as the script already does when the ground truth lists no price to compare. That caps the answer's score. The brand node is SEQUENTIAL, so its first requirement failing skips the brand's other three, and each comparative-post check requires every brand's cheapest-sedan check, so all three posts are skipped as well. An otherwise correct answer then scores 0.5 with one brand missing (measured with the offline harness) and 0.25 with two, depending on whether a live capture succeeded during that answer's evaluation. This PR raises instead, so that a failure on the evaluator's side is visible and does not lower the agent's score. Switching to the placeholder is a small change if preferred.
Verification
All 130 scripts were run offline with the harness proposed in OSU-NLP-Group/Mind2Web-2#12, which uses a fake judge and a synthetic webpage cache, under four synthetic answer policies: every verification passes; a deterministic mix of passes and failures; every extracted list is empty; every extracted list has 12 items. Rubric trees were compared before and after this change:
- Before: 18 of the 520 runs crashed or were killed after 30 seconds.
- After: 519 runs complete, and the 502 runs that completed before produce identical rubric trees and final scores. The remaining run is
research_car.pyunder the empty policy, which also empties the ground-truth extraction and therefore stops with the error above.
Targeted runs with a modified fake judge:
| Script | Input | Released | This PR |
|---|---|---|---|
research_car |
Mazda's page yields no models; every verification passes | AttributeError after 22 verification requests |
The RuntimeError above, before any verification |
research_car |
Mazda's only model has no name | AttributeError |
The RuntimeError above |
oscar_A24 |
No films; the judge rejects every claim | Never returns | Score 0 with a single request, the extraction of the film titles |
For apple_watch.py, the JSON schema sent to the judge for structured output is also unchanged, because the OpenAI SDK drops None defaults when it builds the strict schema.
The fixes are applied in place in 2025_10_23 because they change no score that these scripts could produce before. No changelog entry is included; add one if the fixes should be announced as a script update.
🤖 Generated with Claude Code