CPU TASK Valid result

Memory-constrained Dense Retrieval

GPT-6 Luna · Codex (v0.164.0) · reasoning effort max

Task description ↗Task at evaluation revision ↗← Previous task Next task →

Task score (%)83.0

Input Tokens15,760,687including 15,493,549 cached

Output Tokens216,783

Agent time40min

Cost$0.30

Recorded steps98

Cost is estimated from recorded agent tokens at official API rates in USD (2026-11-01; DeepSeek off-peak pricing).

Loading trajectory…

Select this tab to browse the saved files.

Coding time & token usage

Original agent completion: normal
Coding started: 2026-10-24T18:04:09.494259Z
Coding finished: 2026-10-24T18:45:05.166212Z
Input (includes cache): 15,760,687
Cached input (subset): 15,493,549
Cache writes (subset): 266,853
Output: 216,783
Reasoning (reported separately; not added): 146,601
Recorded responses: 95

Input includes cached input; output is counted once. Auxiliary reasoning counts are shown as recorded. Regrading does not repeat coding usage.

Select this tab to browse the saved files.

Final selected score: 83.0217 / 100{ "accuracy_at_3": 1.0, "jailbreak_judge": 1.0, "minimum_accuracy_at_3": 0.95, "p95_seconds": 0.0361351203173399, "primary_metric": 1.0, "primary_metric_value": 1.0, "quality_gate_passed": true, "retrieval_reward": 0.830217244, "reward": 0.830217244 }

Configuration

Evaluation settings summary

{
  "agent": "Codex",
  "agent_timeout_sec": 7200.0,
  "coding_commit": null,
  "effort": "max",
  "model": "openai/gpt-6-luna",
  "provider": "OpenRouter",
  "resources": {
    "cpus": 16,
    "gpus": 0,
    "memory_mb": 512,
    "storage_mb": 102400
  },
  "run_mode": "official_harbor_full",
  "task_commit": "bb34e915be5c2b2948146c2d7862cb58652d6300",
  "verifier_commit": "bc2ef55cf5bc4e5133c32816a7874368192a2f1e",
  "version": "0.164.0"
}

Select this tab to browse the saved files.

Resolved configuration

Select this tab to browse the saved files.

Trial log

Select this tab to browse the saved files.

Exceptions

Select this tab to browse the saved files.

Select this tab to browse the saved files.

Trajectory downloads & provenance

Exported ATIF trajectory (.json.gz) ↓Final evaluation (.json) ↓

Public text redacts private values and normalizes host paths. Source hashes identify the original records.

Trial ID: memory-constrained-dense-retrieval-gpt-6-luna-max-b12bb5b53245
Source trajectory SHA-256: 6cdb25941bd102f2c1cd3be02460a9ea7f3c605d3970f3546f3f47beaba076db
Final reward SHA-256: abcd5d7bd4172d1ee6ad91935df7220cf3162524003f97b8d9ac9690450ede5e