resume-or-prevent-traces

Raw evaluation and training traces for "Resume or Prevent?", an individual project for the GStar Bootcamp 2026 that builds on the paper Evaluating Agents Across Runtime Contracts: When Mismatch Costs Efficiency or Quality (IAEval 2026, the NeurIPS 2026 Workshop on Evaluation of Interactive Agents). This dataset is not produced by the paper's own authors; it covers only the new experiments this project adds on top of the paper's own released evaluation data and teacher traces (see the paper's own runtime-contracts Hugging Face organization for those).

Full methodology and the numbers computed from these traces are in the project report.

Contents

Directory	What it is
rq1_knapsack_ladder/	The main L0-L3 carryover-level comparison on Opaque Knapsack, 50 held-out instances (Table 2 of the report).
rq1_knapsack_rerun/	A plain rerun of the L2/L3 comparison above, nothing changed, used to check whether the apparent win was a wording artifact or run-to-run noise (Table 3, "Original, rerun" row).
rq1_knapsack_wording_added_rule/	Table 3's "added rule" wording variant.
rq1_knapsack_wording_inline/	Table 3's "shown only when relevant" wording variant.
rq1_knapsack_wording_fewshot/	Table 3's "worked example" wording variant. Its L2 cell stopped at 46/50 instances (cut off by a cluster time limit, never resubmitted); the report's footnote on that row explains how the comparison is computed on the 46 matched pairs.
heldout50_tasks_knapsack/	The held-out-50 task pool itself: each item's real weight, value, and class, for all 50 Knapsack instances used by every directory above plus rq2_eval/rq2_eval_rerun. This is what the fabrication checks (Table "Fabrication at test time" and the mechanism table) compare the agent's claims against; it is not agent output.
rq1_navigation_n50/	An independent 50-instance replication of the full L0-L3 carryover-level comparison on the Navigation task (including the L0 reset baseline), plus the two prompt ablations (worked example removed/added, last_action blanked).
rq1_mechanism_knapsack/	The 2x2 (training regime x carryover level) trap-rate and fabrication mechanism check on Opaque Knapsack, two training seeds. Also uses heldout50_tasks_knapsack/ as ground truth.
rq2_training_traces/	Teacher traces (Gemini 3 Flash) for the prevention and recovery strategies, both scales (56 and 103 valid pairs), before and after pairing/filtering.
rq2_eval/	Evaluation traces for the trained prevention and recovery adapters, both scales, both tool-call caps (25 and 80), on the same 50 held-out instances.
rq2_eval_rerun/	A plain rerun of the n=103 prevention/recovery evaluation (both caps), nothing changed, to check whether that result holds up under eval noise.

Each task instance has a .json (the agent's outcome: score and task-specific metrics, not ground truth) and a .trace.json (the full agent transcript). The one directory that holds actual ground truth is heldout50_tasks_knapsack/, called out separately above. Cluster-internal bookkeeping (SLURM job IDs, local filesystem paths) is not included, since it carries no reproducibility value off that cluster.

Reproducing the report's numbers from this release

The repo's scripts/reproduce_report_tables.py, scripts/analyze_phase1.py, and scripts/reproduce_fabrication_table.py were run directly against this cluster's own local directory layout, which uses different folder names than this public release does. To run them against a local download of this dataset instead, recreate the layout they expect by renaming the downloaded top-level folders:

This dataset's folder	Rename to (what the scripts expect)
rq1_knapsack_ladder/	heldout50_ladder/
rq1_knapsack_rerun/	heldout50_ladder_rerun/
rq1_knapsack_wording_added_rule/	heldout50_ladder_a/
rq1_knapsack_wording_inline/	heldout50_ladder_c/
rq1_knapsack_wording_fewshot/	heldout50_ladder_fewshot/
heldout50_tasks_knapsack/	heldout50/tasks/knapsack/
rq1_navigation_n50/	nav_heldout50_ladder/
rq1_mechanism_knapsack/	addition_phase1/
rq2_eval/n103/*	eval_180/*
rq2_eval/n56/*	eval/* (the n=56 scale uses a differently-named parent folder in the scripts, not eval_180)
rq2_eval_rerun/	eval_180_rerun/

Then point INDIVIDUAL_PROJECT_RESULTS at the folder containing the renamed heldout50*/nav_heldout50_ladder/addition_phase1 directories, and INDIVIDUAL_PROJECT_DATA at the folder containing the renamed eval_180* directories, and run the scripts as documented in their own docstrings. Cell-level names (e.g. PREV_cap25) are unchanged between this release and the scripts, only the top-level container folders were renamed for public clarity.

License

Apache License 2.0, matching the base paper's own released datasets.