# BOX Model Arena

Date: 2026-06-12

BOX Model Arena lets private training models compete without exposing the model file to the public.

The v1 privacy model is commit-reveal with private evaluation:

- Public: wallet, model nickname, model hash, manifest hash, tournament ID, score, rank, and award record.
- Private: model weights, training file, source code, training notes marked private, and evaluator test cases.
- Operator-visible: the evaluator receives the submitted package after the commit deadline so it can verify and score it.

This is not full zero-knowledge inference yet. It is the practical v1 that lets BOX launch quickly while preserving a path toward stronger proofs later.

## Arena file format

Submission package:

```text
box-model.zip
  manifest.json
  model/
  adapter/
  README.md
```

Required public manifest fields:

- `schema`
- `tournamentId`
- `modelId`
- `modelNickname`
- `ownerWallet`
- `modelPackageSha256`
- `manifestSha256`
- `commitmentMessage`
- `walletSignature`
- `runtime`
- `taskType`
- `visibility`

See `model-arena/box-model-manifest.schema.json` and `model-arena/box-model.manifest.example.json`.

## Commit-reveal flow

1. Builder creates `box-model.zip`.
2. Builder computes SHA-256 hash of the package and manifest.
3. Builder signs the commitment message with the owner wallet.
4. Builder submits public commit before deadline.
5. After commit deadline, builder uploads the private package to BOX evaluator.
6. Evaluator verifies package hash and wallet signature.
7. Evaluator runs the model against hidden boxing tasks.
8. BOX publishes leaderboard record with rank, score, hash, and award.

## Weekly champion loop

The weekly champion is automatically seeded into the next week.

Rules:

- Current champion keeps the public title until beaten.
- New challengers submit commits before the weekly deadline.
- Champion model must run against the same hidden challenge set as the challenger pool.
- If the champion wins, title streak increases by one.
- If a challenger wins, challenger becomes the new champion.
- Monthly champion is selected from weekly champions plus top consistency score.

This creates the right pressure: the best model can stay on top, but only by defending against newer models.

## Tournament cadence

Daily:

- small practice challenges,
- participation points,
- anti-cheat smoke tests.

Weekly:

- official title challenge,
- BOX award pool,
- champion defense.

Monthly:

- monthly champion board,
- larger BOX award,
- public write-up if approved.

## Scoring categories

Start with boxing-intelligence tasks that can run without Unity:

- footwork phase classification,
- punch timing prediction,
- guard recovery score,
- distance-control recommendation,
- matchup style classification,
- drill recommendation quality.

Later, Unity can add simulated boxer control and replay scoring.

## Privacy and safety boundaries

- Do not publish model files.
- Do not publish hidden test sets.
- Do not publish training data unless the owner explicitly marks it public.
- Do not accept private personal data, medical records, biometric identifiers, or unapproved fighter footage in v1.
- Do not claim the model gives medical, concussion, return-to-play, or weight-cutting advice.

## Award record

Each award should store:

- `tournamentId`
- `period`
- `rank`
- `ownerWallet`
- `modelId`
- `modelPackageSha256`
- `score`
- `boxAward`
- `awardTxHash` after public token deployment

Before public token deployment, `awardTxHash` can stay `null` and awards can be tracked as pending BOX credits.