Rules previewIntegrity principles come from the accepted proposal; operational thresholds remain TBA.
Competition policy

Rules

Build real repairs, disclose the system that produced them, preserve the evaluator, and make final results reproducible under organizer control.

Primary focus
LLM-based repair Agents
Final validation
Organizer audit and rerun
Thresholds
Published before launch

This page preserves the accepted proposal's integrity principles while avoiding unconfirmed limits. Team size, submission frequency, API allowlists, and exact resource budgets will be added after organizer approval.

Who the competition is designed for

The primary leaderboard is intended for LLM-based and agentic repair systems from academic, industry, and independent teams. Registration rules, team-size limits, affiliation restrictions, and conflict-of-interest procedures will follow the official ICSE Competition Track requirements and the released rulebook.

One team, one declared entry

Final registration and team-composition rules are TBA.

Agent-focused ranking

Non-LLM systems may appear as clearly labeled comparison baselines, subject to final policy.

Valid final run required

A team must submit a conforming Agent and complete organizer-run evaluation to receive a final result.

Describe what the Agent actually uses

Each final entry must disclose enough information for organizers to understand and rerun the method without revealing hidden evaluator data.

  • Base model names and exact versions when available
  • Major prompts, retrieval sources, external tools, static analyzers, and repair-loop design
  • Training, fine-tuning, external knowledge bases, or private data used for the entry
  • Third-party APIs and services contacted during Agent execution
  • Known nondeterminism, caching, preprocessing, and reproducibility constraints

Repair package content through the official interface

Reason over all released case evidence

Specifications, sources, patches, logs, scripts, metadata, and architecture labels.

Use an internal toolchain

Prompting, retrieval, code search, static analysis, log parsing, and iterative Agent loops.

Modify permitted package paths

Specifications, build scripts, source code, tests, packaging macros, and related package files.

Return a different valid repair

Repairs need not match an organizer reference patch when they pass policy and executable validation.

Do not bypass the task or evaluator

Hard-code hidden solutions

No embedded hidden patches, case-specific answers, leaked labels, or lookup tables.

Exploit evaluator bugs

No attempts to escape isolation, alter evaluator metadata, spoof results, or rely on unintended validator behavior.

Disable meaningful validation

No deleting essential tests, bypassing the build, or suppressing failures without a legitimate repair.

Replace the package with a dummy artifact

The repaired output must remain the intended package and preserve its meaningful functionality.

Modify forbidden paths

Case manifests, checksums, hidden metadata, evaluator files, and other immutable inputs must remain unchanged.

Use undeclared manual intervention

Organizer-run cases must be processed by the submitted Agent under the announced interface and policy.

Do not assume unrestricted external access

The unified Agent model requires an explicit policy for model APIs and network access. That policy is not final. The release will state whether calls are disabled, allowlisted, or brokered through organizer-managed APIs.

Participants must never embed production API keys or personal credentials in submitted artifacts. Secret injection, logging, accounting, and redaction will be defined by the official platform.

Held-out cases stay under organizer control

  • 1

    The hidden test set will not be distributed before final evaluation.

  • 2

    Final Agents run with a frozen interface, evaluator version, and resource policy.

  • 3

    Only aggregate results and limited diagnostic categories will be released after the deadline.

  • 4

    Organizers may inspect successful, unusually small, suspicious, or policy-sensitive repairs.

  • 5

    Entries affected by an evaluator defect may be rerun under a documented correction policy.

Results must survive review and rerun

Organizers may request source, logs, configuration, method documentation, or a reproducible artifact for finalist verification. A result may be corrected, removed, or disqualified when it cannot be reproduced or violates the released rules.

SituationPotential action
Malformed or nonconforming Agent outputCase failure or invalid submission under the final scoring policy
Accidental evaluator defectAnnounced fix and affected-entry rerun where feasible
Undeclared dependency or unreproducible finalistRequest clarification, corrected artifact, or result removal
Build bypass, hidden-answer leakage, or sandbox abuseDisqualification and removal from the leaderboard