Datasets:

---

agent-evals

/

hal_traces

Like 7

Follow

hal 24

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:

1. Cost is double counted. get_total_cost sums summary.usage over every Weave call with trace_roots_only=False, and a litellm.completion parent and its openai.chat.completions.create child 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%.
2. 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 closing respond rarely 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.
3. The archives record grader-written actions as agent actions. tau-bench's calculate_reward replays the gold sequence through env.step, which appends it to env.actions, and HAL serialises that list as taken_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.