Submission Rules

Requirements, format specifications, and resource limits for all AISB 2026 competition tracks.

Submission Format

Each submission is a Docker image that, when executed with the provided task inputs, produces outputs in the specified format. The image must be self-contained: all code, model weights, and dependencies must be included.

Required Directory Structure

submission/
  Dockerfile              # Build instructions
  run.sh                  # Entry point script
  src/                    # Source code
  models/                 # Model weights (if any)
  config.yaml             # System configuration
  metadata.json           # Team info, system description
  output/                 # Generated at runtime
    results.json          # Final results
    iterations.jsonl      # Step-by-step execution log
    claims.json           # Self-reported numerical claims

metadata.json Schema

{
  "team_name": "string",
  "system_name": "string",
  "track": "T1 | T2 | T3",
  "contact_email": "string",
  "description": "string (max 500 chars)",
  "base_models": ["list of LLM/model names used"],
  "estimated_cost": "float (USD)",
  "submission_date": "YYYY-MM-DD"
}

Requirements

Must

+Run inside the provided Docker sandbox without modification
+Produce output/results.json and output/iterations.jsonl
+Complete within the track-specific time limit
+Stay within the API cost budget
+Include accurate metadata.json with team and system details
+Report all numerical claims in output/claims.json for CAS verification

Must Not

xAttempt to access the internet or external APIs
xModify the Docker container or escape the sandbox
xAccess test labels, answer keys, or canary tokens
xHard-code benchmark-specific answers in source code
xReport results not produced during the evaluation run
xUse more than the allocated compute resources (GPU, memory, disk)

Evaluation Procedure

Submission Upload -- Teams upload their Docker image to the AISB submission portal. The image is scanned for security issues and size compliance.

Sandbox Execution -- The image is launched inside the Docker sandbox (Layer 1) with task inputs. Execution traces are recorded (Layer 2).

Output Collection -- Results, execution logs, and claims are extracted from the container output directory.

Integrity Check -- CAS is computed (Layer 3). Canary tokens are checked (Layer 4). Integrity status is assigned (PASS / FLAGGED / FAIL).

Scoring -- Raw task score is computed. DualScore = RawScore x CAS. Results are posted to the leaderboard.

Review Period -- FLAGGED submissions enter a 7-day review period. Teams may provide explanations for unverified claims. Final status is determined by organizers.

Disqualification Criteria

--CAS below 0.5 (significant result fabrication)
--Canary token detected in output (test set contamination)
--Sandbox escape attempt detected
--Exceeding resource limits by more than 10%
--Evidence of hard-coded benchmark answers
--Multiple submissions from the same team under different names

Disqualified submissions remain on the leaderboard with FAIL status for transparency but are ineligible for prizes or rankings.