Three reproducible measurement issues in these archives (cost double count, ungraded Generalist episodes, grader-injected actions)
We re-derived parts of the HAL leaderboard from these archives alone, on CPU, with no model calls. Three findings, all reproducible from public data:
- Cost is double counted.
get_total_costsumssummary.usageover every Weave call withtrace_roots_only=False, and alitellm.completionparent and itsopenai.chat.completions.createchild carry the same usage block. In 70 of 134 runs with logs the reported token count is exactly 2.000x the root-only count; over the 133 runs with usage data the overstatement is 36.5%. - HAL Generalist episodes on TAU-bench Airline are mostly ungraded. tau-bench computes the reward only when an episode is
done, and the Generalist's single closingrespondrarely draws a###STOP###from the user simulator. It was graded in 39.3% of episodes against 98.4% and 98.7% for the task-specific scaffolds. Re-grading offline moves its pooled accuracy from 0.300 to 0.468. - The archives record grader-written actions as agent actions. tau-bench's
calculate_rewardreplays the gold sequence throughenv.step, which appends it toenv.actions, and HAL serialises that list astaken_actions. The suffix is present in 1,523 of 2,334 airline records (65.3%); in 58 records the full gold tool sequence appears although the agent made zero tool calls.
If you use these traces for action counting, tool-use analysis, or fine-tuning, item 3 affects you directly: strip the trailing gold before analysis. The detection rule is that the last len(gold) actions equal the task's gold actions with terminating actions excluded.
The full write-up, with line references, per-run numbers and the reproduction scripts, is at https://github.com/YangzeLiu/eval-instrument-disclosure; the same report has been sent to the paper's corresponding authors. The harness repository is archived, so there is no issue tracker to file this on. Corrections very welcome, here or by email.
Thanks — we've reproduced all three on our own archives and can confirm them.
On the frontend we are adding disclosures to the affected leaderboard pages:
- Cost: the Total Cost caveat now notes that nested trace spans counted the same API response twice, inflating affected models by exactly 2×, and that the overstatement is not uniform across rows.
- TAU-bench Airline: the HAL Generalist rows are marked as a lower bound on that scaffold's accuracy and not directly comparable to the task-specific scaffolds, since episodes that never terminate are recorded as 0.
- Traces: the download notice tells users that
taken_actionsincludes grader-replayed reference actions, and to strip the trailing gold sequence before analysis.
The archives themselves are unchanged, so your note here remains the right pointer for downstream users. Item 3's root cause is upstream in tau-bench, so we'll follow your filing there rather than duplicate it.