06On-device

Reproducibility

Same inputs, same bytes. A digest-pinned image, pinned dependencies, frozen parameters and a deterministic decoder: anyone can rebuild our exact output file in three commands.

  • < 10 sfull decode on a laptop CPU
  • 15behaviour tests, no data needed
  • ~15 sto rerun the seeded Pareto search
  • 2architectures: arm64 and amd64

AIn motion

Clone to identical file

  1. Clonepublic repo, no data
  2. Builddigest-pinned image
  3. Test15 behaviour tests
  4. Runoffline, under 10 s
  5. Same filevalidated, byte for byte
Run 1submission.csv
sha256 fingerprint · LF line endings
Match
Run 2submission.csv
same fingerprint, every run

Deterministic decoding: the same data gives byte-identical output on every run, and the official validator checks every file.

BTry it

Reproduce in three commands

Only Docker is needed. DATA is your local folder with the match bundle, placed as DATA.md describes. It is never part of the repository.

# git clone https://github.com/Shrewdies/MTAH2027
# cd MTAH2027/initial-challenge
$ scripts/run.sh build
$ scripts/run.sh test
$ scripts/run.sh submit "$DATA" outputs 
# same clone, then from
# initial-challenge
> .\scripts\build.ps1
> .\scripts\test.ps1
> .\scripts\run_protected.ps1 -DataDir "$DATA" 

COpen the files

The container and the scripts

Docker

The digest-pinned image and the complete dependency list.

Linux · macOS

One script: build, test, submit.

Windows

PowerShell: build the image, run the tests, generate the file.

Full reproduce guide

The README's Reproduce section: every OS, plain Docker, and offline builds.

DPinned

Every layer is pinned

  1. Base imagepython:3.11.11-slim-bookworm @sha256:081075da…fd38e5
  2. numpy2.2.6
  3. pandas2.2.3
  4. python-dateutil2.9.0.post0
  5. pytz2024.2
  6. tzdata2024.2
  7. six1.16.0
Offline wheelsWith the pinned wheels in wheels/, the build installs with no network at all.
Multi-archOne image for linux/arm64 and linux/amd64: Windows, macOS and Linux alike.
Both paths testedThe offline wheel build and the pinned PyPI build were both tested.

EFrozen

Nothing left to chance

  • Frozen profiles

    Decoder parameters live in possession/profiles.json: macro, balanced, spell and reasoned, each one argument away.

  • Seeded search

    The Pareto search behind the profiles is seeded and runs on simulated matches; it reruns in about 15 seconds.

  • Tested and validated

    15 behaviour tests on fabricated scenes run with no match data. The official validator runs on every output file.

NextOur future plans join this page as the Innovative Challenge takes shape.