Skip to content

Current results

Everything on this page is generated from results/harness.json, which is written by agentgate harness export and committed on every recording session. Nothing here is typed by hand, so it cannot drift from the evidence it describes.

Read every value with its interval. A point estimate on its own is the artifact this project exists to argue against.

2 recording(s) across 1 suite(s) and 2 model(s), at 95% confidence.

What these numbers are not

The τ² suite is a single-turn adaptation of τ²-bench, which is multi-turn with a simulated user. The tasks, the tool surface and the gold trajectories are τ²-bench's; the interaction protocol is ours. These are not τ²-bench leaderboard scores and must not be compared to them.

Metrics absent from a table did not apply to that suite and were skipped, never scored zero.

Llama 3.2 3B on tau2_retail

ollama_chat/llama3.2:3b · agent tau2_retail_agent · 111 tasks x K=3 = 333 units · 99% completed · recorded 2026-08-08

Headline

metric value 95% CI n method
outcome.task_success 0.180 [0.108, 0.252] 111 clt(task-rates)
judge.instruction_following 0.619 [0.559, 0.680] 111 clt
judge.coherence 0.994 [0.988, 1.000] 111 clt
trajectory.recall 0.535 [0.476, 0.595] 111 clt
trajectory.precision 0.574 [0.517, 0.631] 111 clt
trajectory.argument_correctness 0.145 [0.116, 0.173] 111 clt
trajectory.error_recovery_rate 0.000 [0.000, 0.000] 105 clt
trajectory.step_efficiency 0.766 [0.711, 0.821] 111 clt
All other metrics
metric value 95% CI n method
efficiency.completion_tokens 253.853 [234.532, 273.174] 111 clt
efficiency.est_cost_usd 0.000 [0.000, 0.000] 111 clt
efficiency.latency_ms 13,260 [12264.887, 14254.989] 111 clt
efficiency.llm_roundtrips 2.027 [1.988, 2.066] 111 clt
efficiency.prompt_tokens 1,875 [1836.916, 1913.102] 111 clt
efficiency.tool_calls_count 4.234 [3.758, 4.711] 111 clt
efficiency.total_tokens 2,129 [2077.419, 2180.305] 111 clt
outcome.abstained 0.000 [0.000, 0.000] 111 clt(task-rates)
rag.answer_relevancy 0.267 [0.243, 0.290] 111 clt
trajectory.any_order_match 0.225 [0.147, 0.303] 111 clt(task-rates)
trajectory.exact_match 0.000 [0.000, 0.000] 111 clt(task-rates)
trajectory.f1 0.458 [0.421, 0.495] 111 clt
trajectory.in_order_match 0.189 [0.116, 0.262] 111 clt(task-rates)
trajectory.lcs_ratio 0.501 [0.445, 0.557] 111 clt
trajectory.loop_detected 0.000 [0.000, 0.000] 111 clt(task-rates)
trajectory.redundant_call_rate 0.028 [0.012, 0.044] 111 clt
trajectory.single_tool_use 0.324 [0.237, 0.412] 111 clt(task-rates)

Qwen2.5 7B on tau2_retail

ollama_chat/qwen2.5:7b · agent tau2_retail_agent · 111 tasks x K=3 = 333 units · 95% completed · recorded 2026-08-09

Headline

metric value 95% CI n method
outcome.task_success 0.111 [0.053, 0.169] 111 clt(task-rates)
judge.instruction_following 0.356 [0.318, 0.395] 111 clt
judge.coherence 1.000 [1.000, 1.000] 111 clt
trajectory.recall 0.442 [0.379, 0.505] 111 clt
trajectory.precision 0.522 [0.445, 0.600] 111 clt
trajectory.argument_correctness 0.321 [0.268, 0.374] 111 clt
trajectory.error_recovery_rate 0.047 [0.011, 0.082] 92 clt
trajectory.step_efficiency 0.730 [0.672, 0.787] 111 clt
All other metrics
metric value 95% CI n method
efficiency.completion_tokens 339.862 [290.310, 389.414] 111 clt
efficiency.est_cost_usd 0.000 [0.000, 0.000] 111 clt
efficiency.latency_ms 58,702 [48578.407, 68825.158] 111 clt
efficiency.llm_roundtrips 5.390 [4.855, 5.926] 111 clt
efficiency.prompt_tokens 11,169 [9516.088, 12822.260] 111 clt
efficiency.tool_calls_count 5.721 [4.861, 6.580] 111 clt
efficiency.total_tokens 11,509 [9823.344, 13194.728] 111 clt
outcome.abstained 0.000 [0.000, 0.000] 111 clt(task-rates)
rag.answer_relevancy 0.179 [0.163, 0.194] 111 clt
safety.destructive_action_without_confirmation 1.000 [1.000, 1.000] 27 clt(task-rates)
trajectory.any_order_match 0.126 [0.065, 0.187] 111 clt(task-rates)
trajectory.exact_match 0.018 [0.000, 0.043] 111 clt(task-rates)[clipped]
trajectory.f1 0.415 [0.359, 0.472] 111 clt
trajectory.in_order_match 0.120 [0.060, 0.180] 111 clt(task-rates)
trajectory.lcs_ratio 0.435 [0.372, 0.497] 111 clt
trajectory.loop_detected 0.033 [0.003, 0.063] 111 clt(task-rates)
trajectory.redundant_call_rate 0.047 [0.027, 0.068] 111 clt
trajectory.single_tool_use 0.237 [0.158, 0.316] 111 clt(task-rates)