apodex/FrontierChallenge-reference

FrontierChallenge reference data

FrontierChallenge reference data provides 97 authenticated, encrypted verifier archives.

Path	Contents
tasks/<task-id>/verifier.fcref	encrypted tests/: grader, rubric, fixtures, validation code, and reference outputs
manifest.jsonl	archive paths, sizes, and SHA-256 commitments
source_registry.json	release binding shared with GitHub and the solve dataset
tools/	integrity checker and standalone unsealer

The archive password is public:

frontier-challenge-reference

Encryption prevents casual search-engine and Dataset Viewer indexing; it is not access control. Anyone evaluating the benchmark can unseal an archive:

python tools/unseal_verifier.py tasks/<task-id>

Normal evaluation does not require manual unsealing. The evaluator runtime requires Python 3.12+ for its pinned Harbor 0.20.0 (see the Quickstart). The FrontierChallenge runtime verifies the solve/reference registry, copies the selected archive into evaluator-owned staging, authenticates and decrypts tests/, and gives Harbor that staged task. Harbor exposes the instruction and environment to the agent but reserves tests/ for its verifier phase.

This repository deliberately contains no instructions, task inputs, or runtime environments. Verify a downloaded reference package with:

python tools/verify_reference_dataset.py

Metrics

Official Pass Rate uses a strict score threshold: a task passes only when evaluation_complete == 1 and valid task_score > 0.999. Native per-task thresholds do not determine this metric. Use unrounded scores: exactly 0.999 does not pass; 0.9991 and 1.0 pass.

• Pass Rate (%) = 100 × number of completed tasks with valid task_score > 0.999 / 97.
• Score (0–100) = 100 × sum of valid, completed task_score values / 97.
• Partial credit remains the native rubric score normalized to [0, 1]. Missing tasks, invalid scores, and incomplete evaluations contribute zero to the fixed denominator and must be reported separately; incomplete runs are not final benchmark results.
• Use one predeclared attempt per task. A subset may use its predeclared task count as denominator, but must be labeled as a subset, not the 97-task result.

There is only one pass field, passed, using the same strict threshold in reward.json, summary.csv, and summary.json. No alternate pass field is emitted. The runtime applies this rule to the staged reward adapter after unsealing; encrypted archives and partial-credit rubrics remain unchanged. The updated runtime also removes native pass decisions from published verifier diagnostics. Jobs created by older runtimes cannot be resumed under the new policy: use a fresh job name. Old files are left untouched; summarizing an old job recomputes the official metric but does not rewrite its logs or rewards. The summary records metric_definition: score-gt-0.999. Historical results using per-task thresholds, == 1.0, or >= 0.999 must be recomputed from raw scores before comparison. See the scoring guide and runtime update.

Runtime rollout: the runtime update is pending. Before reporting results, check that your summary records metric_definition: score-gt-0.999; older runtimes use the legacy rule. Use the runtime summary for the fixed 97-task denominator; Harbor may aggregate only attempted trials.

Integrity

The root README.md is a mutable dataset card and is intentionally outside checksums.sha256. Task files (including task-level READMEs), manifests, registries, and verification tools remain checksummed. Payload changes require regenerating their checksum entries; editing this card does not.

Citation

@misc{apodex11,
  title         = {Apodex 1.1: Scaling Agentic Intelligence for Complex Work},
  author        = {{Apodex Team}},
  year          = {2026},
  eprint        = {2608.23283},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2608.23283}
}

@misc{frontierchallenge,
  title         = {FrontierChallenge: Evaluating Scientific Workflow Completion},
  author        = {Liangcai Su and Zhaopeng Feng and Zhuo Chen and Zhen Zhang
                   and Xiang Lin and Ruilin Li and Handuo Zhang and Ning Wang
                   and Kailong Wen and Yueqi Guo and Feng Xing and Yiling Guo
                   and Chenxiong Qian and Simon Shaolei Du and Lidong Bing
                   and Xinyu Wang},
  year          = {2026},
  eprint        = {2608.24979},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2608.24979}
}