Skip to content

api.md

Latest commit

 

History

History
112 lines (90 loc) · 4.13 KB

api.md

File metadata and controls

112 lines (90 loc) · 4.13 KB
Raw

API Reference

PRISM exposes public challenge routes and internal BASE routes.

Public Routes

  • GET /health — challenge health metadata.
  • GET /version — version, API version, SDK version, and capabilities.
  • GET /v1/submissions/history — daily submission counts over a window.
  • GET /v1/leaderboard — submissions ranked by final_score for the current epoch (earliest-commit-wins on a tie, one entry per hotkey).
  • GET /v1/architectures/{architecture_id} — one architecture's lab detail (name, owner, best score/submission, variant and submission counts, first_seen_at, updated_at); 404 if absent.
  • GET /v1/architectures/{architecture_id}/variants — the architecture's training-script variants (best first); 404 if the architecture is absent, empty variants is valid.
  • GET /v1/architectures/{architecture_id}/report — the cached LLM auto-report, generated lazily and non-blockingly through the master LLM gateway and grounded only in measured facts. report.status is ready, pending, or unavailable; content may be null when not ready.
  • GET /v1/submissions/{submission_id}/curve — the persisted loss curve (online_loss + covered_bytes_cumulative, downsampled to at most 500 points, first and last preserved), the prequential bits-per-byte scalars, and the reconciled compute profile (estimated_flops, gpu_hours, peak VRAM/RSS, wall-clock); 404 if none.
  • GET /v1/epochs/current — current epoch id and length.
  • GET /v1/epochs — recent epochs.
  • GET /v1/health/eval-jobs — recent eval-job health entries (id, submission id, level, status, attempts).
  • GET /v1/gpu/status — GPU-lease summary (total GPUs, active leases, by status, by tier).

POST /v1/submissions

Submit a two-script bundle directly to PRISM (miner authentication headers). In production, submissions usually enter through the BASE proxy and the internal bridge route instead.

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

GET /v1/submissions/{submission_id}

{
  "id": "...",
  "hotkey": "...",
  "epoch_id": 123,
  "status": "completed",
  "final_score": 0.72,
  "anti_cheat_multiplier": 1.0
}

final_score is the challenge-computed prequential bits-per-byte score (a lower bpb yields a higher final_score); q_arch, q_recipe, diversity_bonus, and penalty are legacy fields retained for response-schema stability. status is pending, running, completed, failed, rejected, or held (rejected = failed a static gate, the two-script contract, the LLM hard gate, or duplicate review; held = LLM-review quarantine).

GET /v1/architectures

Architecture-lab leaderboard grouped by architecture family, ranked by best_final_score descending. Optional epoch_id scopes to architectures with a completed submission in that epoch; omitting it resolves to the most-recent non-empty epoch. name is the miner-declared, deterministically moderated architecture name (may be null).

{
  "epoch_id": 42,
  "architectures": [
    {
      "rank": 1,
      "architecture_id": "...",
      "arch_hash": "...",
      "name": "Rotary MoE v3",
      "owner_hotkey": "...",
      "best_final_score": 1.2345,
      "best_submission_id": "...",
      "variant_count": 3,
      "submission_count": 7,
      "updated_at": "..."
    }
  ]
}

Internal BASE Routes

All internal routes require Authorization: Bearer <shared-token>.

GET /internal/v1/get_weights

Standard BASE challenge contract. Returns normalized, dry-run hotkey weights (one per hotkey, from that hotkey's best final_score). Weights are never written on-chain.

POST /internal/v1/bridge/submissions

Receives BASE-verified submissions.

Authorization: Bearer <shared-token>
X-Base-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 through the full pipeline: static gates, the LLM hard gate, the forced-init re-execution, and prequential bits-per-byte scoring.