Skip to content

api.md

Latest commit

 

History

History
209 lines (142 loc) · 4.13 KB

api.md

File metadata and controls

209 lines (142 loc) · 4.13 KB

API Reference

PRISM exposes public challenge routes and internal Platform routes.

Public Routes

GET /health

Returns challenge health metadata.

GET /version

Returns challenge version, API version, SDK version, and capabilities.

POST /v1/submissions

Submit a model project directly to PRISM. In production, miner submissions should usually enter through the Platform proxy.

Request:

{
  "filename": "project.zip",
  "code": "<base64 zip payload>",
  "metadata": {}
}

The direct public route uses miner authentication headers. Platform bridge uploads use an internal route instead.

GET /v1/submissions/{submission_id}

Returns status and score fields:

{
  "id": "...",
  "hotkey": "...",
  "epoch_id": 123,
  "status": "completed",
  "code_hash": "...",
  "created_at": "...",
  "error": null,
  "final_score": 0.72,
  "q_arch": 0.8,
  "q_recipe": 0.6
}

status can be pending, running, completed, failed, rejected, or held. Held submissions are awaiting manual component-attribution resolution and do not affect weights.

GET /v1/leaderboard

Returns ranked submissions for the current epoch.

GET /v1/architectures

Returns known architecture families.

[
  {
    "id": "...",
    "family_hash": "...",
    "owner_hotkey": "...",
    "owner_submission_id": "...",
    "canonical_submission_id": "...",
    "q_arch_best": 0.9,
    "created_at": "...",
    "updated_at": "..."
  }
]

GET /v1/training-variants

Optional query parameters:

  • architecture_id
  • limit

Returns training variants and current-best state.

GET /v1/epochs/current

Returns the current epoch id and epoch length.

Internal Platform Routes

All internal routes require:

Authorization: Bearer <shared-token>

GET /internal/v1/get_weights

Standard Platform challenge contract. Returns normalized hotkey weights.

POST /internal/v1/bridge/submissions

Receives Platform-verified submissions.

Headers:

Authorization: Bearer <shared-token>
X-Platform-Verified-Hotkey: <hotkey>
X-Submission-Filename: project.zip
Content-Type: application/zip

The body can be raw ZIP bytes or JSON matching SubmissionCreate.

POST /internal/v1/worker/process-next

Claims and processes one pending submission.

POST /internal/v1/worker/poll

Polls remote jobs. This is retained for compatibility with asynchronous job flows.

GET /internal/v1/component-review/holds

Lists low-confidence semantic attribution holds.

Response:

[
  {
    "id": "...",
    "submission_id": "...",
    "status": "pending",
    "reason": "low confidence component attribution",
    "confidence": 0.61,
    "created_at": "...",
    "updated_at": "..."
  }
]

POST /internal/v1/component-review/holds/{hold_id}/resolve

Resolves a held submission with an operator attribution decision.

Request:

{
  "architecture_action": "existing",
  "training_action": "new",
  "architecture_id": "7ec2c3a8-...",
  "training_variant_id": null,
  "reason": "manual review confirmed same architecture and new training hook"
}

Allowed architecture actions are new, existing, transfer, and reject. Allowed training actions are new, existing, transfer, reject, and none.

Validator Assignment Routes

These routes support validator-assignment operation.

POST /internal/v1/validators/assignments/next

Headers:

X-Validator-Hotkey: <validator-hotkey>

Returns an assignment or null.

POST /internal/v1/validators/assignments/{assignment_id}/accept

Marks an assignment as accepted.

POST /internal/v1/validators/assignments/{assignment_id}/reject

Request:

{
  "reason": "busy"
}

Rejects an assignment and requeues the submission if attempts remain.

POST /internal/v1/validators/assignments/{assignment_id}/result

Request:

{
  "metrics": {
    "q_arch": 0.8,
    "q_recipe": 0.7,
    "train_loss": 1.0
  }
}

Finalizes the assignment and records scores.

POST /internal/v1/validators/assignments/expire

Expires stale assignments and requeues or fails submissions according to the configured attempt limit.