06On-device

Security & Data Protection

The match data never left our laptop. It was processed in a sealed, offline container, never committed and never uploaded, and only aggregate numbers were ever published.

  • 0match files in the repository
  • 0uploads to any service or model
  • Nonenetwork inside the container
  • 1writable folder: the output

AIn motion

Data never leaves the laptop

Data never leaves the laptop Inside the team laptop, match data flows read-only into a sealed Docker container. The container's network cable is cut, so nothing reaches the cloud. A single output door leads to an output folder outside the repository. A .gitignore shield keeps data out of the public repository, which holds code only; the site and report receive aggregate metrics only. CLOUD 0 BYTES SENT OUR LAPTOP --network none CODE MATCH DATA LOCAL ONLY DOCKER · OFFLINE READ-ONLY H·A·D READ-ONLY ROOT · TMPFS /tmp ONE DOOR CSV OUTPUTS OUTSIDE THE REPO .gitignore PUBLIC REPO: CODE ONLY SITE + REPORT: AGGREGATES
  • Match data · local only, mounted read-only.
  • Container · no network, read-only root, tmpfs /tmp.
  • One door · outputs land outside the repository.
  • .gitignore · match data is never committed.

Read-only in, one door out, cable cut. The repository never sees a byte of match data.

BThe run, flag by flag

One sealed command

docker run --rm --network none1 --read-only2 --tmpfs /tmp3 -v "$REPO:/work:ro4" -v "$DATA:/data:ro4" -v "$OUT:/out5" mtah2027:dev

  1. --network noneNo network at all. Nothing inside can reach anything outside.
  2. --read-onlyThe container cannot change its own filesystem.
  3. --tmpfs /tmpScratch space lives in memory and is gone after the run.
  4. :roSource code and match data are mounted read-only.
  5. /outThe single writable path: an output folder outside the repository.

CWhat we did

Three more safeguards around the run

  • Never committed

    The repository holds no match data, and two ignore files keep it that way (below). DATA.md tells each user where to put their own copy locally.

  • Never uploaded

    No external service, cloud platform or hosted model ever saw the data, as the competition terms require. Outputs stay outside the repository; the site and report publish aggregate metrics only.

  • Lean supply chain

    Two libraries, numpy and pandas, and a build that can install from local wheels with no network. Every pin, from the base image down, is listed on Reproducibility.

DTwo ignore files

Two locks, two doors

data/*
*_atd.txt  *_atd.xml
*_smart_tagging.xml  *_homography.csv
*.mp4
private_data/
outputs/
.env  .env.*

*
!requirements-solution.txt
!wheels/

  1. .gitignoreThe door to the public repository. Match data is blocked twice: by folder, and again by file name in case a file ever lands somewhere else.
  2. outputs/Generated label files are answers for the whole match, so every output path is ignored too.
  3. .envSecrets and local environments never enter the history. The project needs none, and the rule is there anyway.
  4. .dockerignoreThe door to the image. It refuses everything, then allows two things: the pinned dependency list and the optional wheels.
  5. *So the image holds Python and two libraries. Code and data are mounted at run time, read-only, and never baked in.

The competition rules

Answers stayed sealed

No hand labels, nothing tuned on the match's answers, and the public 5-minute window used only as a correctness check. What the rules are, and what each one forced, is in The rules are the architecture. And the team photos on this site? Processed locally too.