FREUID Challenge
Post-submission verification

Reproducibility package

After the competition closes on July 15, 2026, we verify solutions before the final private leaderboard and prize announcement. This page is the public reference for requirements, templates, and the Docker sandbox contract.

Prize eligibility - mandatory

Teams that do not provide a complete reproducibility package will not be eligible for prizes or for invitation to the IJCAI-ECAI live showdown, regardless of private leaderboard rank. Partial, late, or non-reproducible submissions forfeit prize eligibility.

  • Open-source code (OSI-approved license) at a frozen commit SHA
  • Technical report (PDF)
  • Runnable Docker artifact (no-network sandbox)
  • Exactly one structured reply on the pinned Kaggle discussion thread

Reply deadline: July 15, 2026, 23:59 AoE - same as the final Kaggle submission deadline. All times 23:59 AoE.

Process overview

  1. Now

    Prepare your package

    Download templates below; test Docker with --network none.

  2. By July 13th, 2026

    Private test release + code freeze

    Private images released. Solution must already be in a public git repo; no model or training changes after this date. Details

  3. By July 15th, 2026

    Final Kaggle submission + reproducibility reply

    Final submission and one pinned-thread reply (use REPLY_TEMPLATE.txt) due by 23:59 AoE. Only update private rows after July 13 - model weights must stay frozen.

  4. July 15th-24th, 2026

    Organizer verification

    Preliminary results window; we verify containers against your Kaggle submission.

  5. July 24th, 2026

    Final winner announcement

    Top teams notified; live showdown invitation at IJCAI-ECAI 2026 (Aug 18th-21st).

Full competition timeline: freuid2026.microblink.com.

Code freeze after private test release

Private test images are released on July 13th, 2026 (~2 days before the final deadline). Your solution must already be in a public git repository (OSI-approved license). No separate lock registration is required.

After July 13th - inference only

  • No changes to model weights, architecture, or training code
  • You may run inference on private images and update private-row Kaggle predictions
  • Documentation and Docker packaging updates are allowed if weights are unchanged
  • Organizers verify compliance via git commit history after private release

Downloads

Optional starter kits. Using them is not required - only the sandbox contract is mandatory for prize eligibility.

Kaggle discussion reply

Short copy-paste template for the pinned thread. Point to your repo, no need to repeat instructions already there.

Docker sandbox contract

Organizers execute your container with no network access. Test images are mounted as a flat directory of files only - no CSV, manifest, or subfolders inside /data.

Mounts

HostContainerAccess
Private test images/data/read-only
Output/submissions/read-write

Input layout (/data/)

/data/
  3f49e6921f3f4b10910738f7e9b476f3.jpeg
  c39d1c24d46a44a088991ddc7a072b7c.jpeg
  ...

The row id is the filename without extension. Supported extensions: .jpeg, .jpg, .png, .webp, .bmp, .tif, .tiff.

Output (/submissions/submission.csv)

id,label
3f49e6921f3f4b10910738f7e9b476f3,0.99
c39d1c24d46a44a088991ddc7a072b7c,0.99

label is a finite float fraud score - higher = more confident the document is fraudulent. This matches the numeric semantics of the Kaggle leaderboard.

Constraints

  • docker run --network none - no internet at inference
  • All model weights inside the image or a read-only mount - no runtime downloads
  • No writes outside /submissions/
  • One output row per image file; no missing or extra ids

Local test

docker build -t freuid-repro:local .

docker run --rm \
  --network none \
  -v /path/to/flat/test/images:/data:ro \
  -v "$(pwd)/out:/submissions" \
  freuid-repro:local

Replace the placeholder logic in predict_labels() inside prepare_submission.py with your model while keeping paths and output schema unchanged.

Technical report

A short PDF describing your method, training data, model architecture, inference procedure, external resources, and reproduction steps. The LaTeX template is optional.

Minimum sections: introduction, method, data (including external sources), inference, results, reproducibility (commands, Docker image, hardware).

Kaggle discussion reply template

Post exactly one reply on the pinned Kaggle discussion thread by July 15th, 2026, 23:59 AoE (same as the final Kaggle submission deadline). Full requirements are in the Kaggle Overview → Code Requirements section. Download REPLY_TEMPLATE.txt. Keep it short: link your repository (with Docker build/run instructions in the README) and your technical report PDF. Do not duplicate content that is already in the repo. 


---
FREUID Challenge 2026 - Reproducibility Package
---

Team name: [Kaggle team name]
Kaggle usernames: [comma-separated]
Final Kaggle submission: [label / date-time as shown on Kaggle]

Repository (this should be public git repository): [URL]
Commit SHA: [40-char hash]
Technical report (PDF): [URL]

We confirm this repository at the stated commit reproduces our selected final
submission and complies with the competition rules.

Signed (team captain): [Kaggle username]
Date (UTC): [YYYY-MM-DD]

Your repository README should cover Docker build/run, model weights, external data, and hardware - organizers will follow that, not a second copy on the forum.

Kaggle leaderboard submission (separate from Docker)

On Kaggle, upload a CSV with columns id,label for the public/private leaderboard:

id,label
3f49e6921f3f4b10910738f7e9b476f3,0.99

Questions

Email freuid-challenge-2026@microblink.com with subject [FREUID Repro] <team name> for private matters (e.g. broken links). Use the pinned Kaggle discussion only for your one official reply.

Kaggle discussions