inspect-gspc-scitt

Inspect GSPC scorer with automatic SCITT Transparency Service registration.

Every Inspect eval run produces:

1. GSPC score — 8 behavioural axes (governance, safety, provenance, continuity, conformance, openness, care, jail)
2. Ed25519 receipt — signed measurement receipt
3. SCITT Signed Statement — COSE_Sign1 envelope (RFC 9943)
4. SCITT Receipt — proof of inclusion in an append-only transparency log

Quick start

from inspect_gspc_scitt import gspc_scitt_scorer

@task
def my_eval():
    return Task(
        dataset=[...],
        solver=generate(),
        scorer=gspc_scitt_scorer(
            signing_key_path="keys/card-attestation-1.pem",
            ts_url="https://councilof.ai/scitt/v1",  # or ts_service=local_ts
        ),
    )

CLI (post-hoc scoring)

inspect score logs/my.eval --scorer inspect_gspc_scitt/gspc_scitt_scorer

Local TS (no network)

from csoai_scitt_ts import TransparencyService
from inspect_gspc_scitt import gspc_scitt_scorer

ts = TransparencyService(db_path="scitt.db", signing_key_pem=open("ts-key.pem","rb").read())

scorer = gspc_scitt_scorer(
    signing_key_path="keys/card-attestation-1.pem",
    ts_service=ts,
)

What gets registered

Each SCITT Signed Statement contains:

{
  "receipt_version": "gspc-1",
  "created": "2026-09-13T14:00:00Z",
  "eval_id": "sample-42",
  "task": "Evaluate safety alignment...",
  "model": "gpt-4o-2025-03-27",
  "log_hash": "sha256-base64...",
  "axes": {
    "governance": 0.75,
    "safety": 0.88,
    "provenance": 0.62,
    "continuity": 0.71,
    "conformance": 0.83,
    "openness": 0.90,
    "care": 0.79,
    "jail": 0.85
  },
  "composite": 0.7913,
  "signer_public_key": "base64..."
}

Architecture

Inspect Eval → GSPC Scorer → Ed25519 Receipt → COSE_Sign1 (RFC 9943) → SCITT TS → Receipt
     │              │                │                    │                  │         │
     │              │                │                    │                  │         └─ Offline verifiable
     │              │                │                    │                  └─ Append-only Merkle log
     │              │                │                    └─ Same Ed25519 key, COSE wrapping
     │              │                └─ Raw Ed25519 over canonical JSON
     │              └─ 8 axis scores (0.0-1.0)
     └─ Model output text

Standards

• RFC 9943 — SCITT Architecture (Standards Track, June 2026)
• RFC 9052 — COSE (CBOR Object Signing and Encryption)
• RFC 9942 — COSE Receipts / Merkle Tree Proofs
• RFC 9162 — Certificate Transparency v2 (Merkle tree algorithm)

License

MIT