Leaderboard v1: Computer / Kimi K3

[c-suvamsa]

Summary

• Agent / Model: DevRev Computer / Kimi K3 (Moonshot AI).
• Source Harbor job URL: https://hub.harborframework.com/jobs/d66ff9d0-e6f0-480e-b1c7-fb3534e3a785.
• Run: 2026-10-16; enterprise-bench/tasks/v1 via Daytona; 14 tasks × 10 attempts.
• Results: 132/140 rewards of 1.0 = 94.3% accuracy; 44,280,478 recorded input + output tokens, excluding unavailable usage in two failed attempts.
• Cost: Not fully recorded. Saved SDK outputs total $89.18 for 138 trials; the other two have no cost output. The row's numeric 0 is a schema-required placeholder, not a measured $0 total, and the row is hidden pending review.

PR checklist – leaderboard row entry

Provenance and access

• Source job is public and owned by the submitter; entry header and URL match its job ID.
• Entry includes all 140 distinct trial IDs, including failed attempts; n_trials matches.
• Accuracy, pass@k, and the recorded token breakdown reconcile with the run.
• Agent, model, and provider are shown on the job.
• Every trial has an ATIF trajectory and token/cost usage: two failed trials do not.
• Reviewer to confirm logged-out replay of available trajectories.

Run conditions and review

• Task content digests for all 14 tasks match enterprise-bench main. The run used an absolute local tasks/v1 path rather than a registry-pinned dataset version.
• Scored timeout multiplier is 1.0 and no CPU/memory/storage override was recorded.
• Reviewer to confirm benchmark-answer access boundaries, applied model route, and independently replay the score.

Validation

• Main-branch leaderboard validator passed locally.
• Repository validate and validate-leaderboard CI checks pass.

Notes for reviewers

Draft – do not merge as a display row without reviewing these exceptions. sales-l2-a__zM65nZ5 and sales-l2-a__pjjPQaG both received reward 0 after an SDK streaming socket closure; neither has an agent trajectory or usage output. Both remain in the score. The run-owner-supplied model label is not independently resolved from the generic gateway alias in the saved traces. This is a new October run, separate from the earlier Computer / Kimi K3 submission using Fireworks AI.