[c-suvamsa]

Summary

• Agent / Model: Opencode / Kimi K3 (Moonshot AI).
• Source Harbor job URL: https://hub.harborframework.com/jobs/12da203b-b10f-4fd7-9567-6ea2f27f517e.
• Run: 2026-10-19; tasks/v1 via Daytona; 14 tasks × 10 attempts; concurrency 4; Harbor 0.24.0.
• Route: DevRev gateway → Baseten; recorded model anthropic/baseten/Kimi-K3.
• Result: 91/140 reward-1.0 trials (65.0%); 53,996,982 recorded input + output tokens.
• Cost: not recorded (schema-required numeric 0 is a placeholder, not an actual $0 cost).

PR checklist – leaderboard row entry

Provenance and access

• Source job is public and owned by the submitter; entry header and job URL match.
• Required metadata and metrics present; 140 distinct trial IDs match n_trials.
• Reward-derived accuracy, pass@k, and recorded token totals reconcile with the source job.
• All 14 task content digests match enterprise-bench main; local tasks/v1 path used instead of a registry-pinned version.
• Default scored timeout multipliers and no CPU/memory/storage overrides; no scored trial omitted.
• All trials have ATIF trajectories and usage records (one trial does not).
• Reviewer to verify logged-out trajectory replay, benchmark-answer access boundaries, and applied upstream model route.

Validation

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

Notes for reviewers

The run recorded 8 AgentTimeoutError trial(s) at the default limit; all are retained with their recorded rewards. 1 timed-out trial(s) received reward 1.0 because the verifier still had a submitted answer. This row remains hidden while reviewers assess the unrecorded cost; ready for review does not mean the placeholder is a measured $0 cost. The saved provider label denotes the Anthropic-compatible gateway route, not the model manufacturer. Trial eng-l1-a__auhcsUQ has reward 0 but no ATIF trajectory or agent usage: collecting the agent logs failed because OpenCode's opencode.db-wal changed during archival. It remains in the row.

[nimit2801]

Thanks for the OpenCode / Kimi K3 leaderboard submission. Acknowledged by Nimit’s intake workflow; this is not approval. Both validation checks are passing. Technical review still needs to verify:

• Source-job provenance and reward, pass@k, and token reconciliation.
• Trajectory replay and answer-access boundaries, including the disclosed missing trajectory and timeout outcomes.
• The applied model route and unrecorded-cost placeholder while the row stays hidden.