Skip to content

Architecture

Current modules

contracts.py validates input models; integrity.py owns bounded I/O and research hashes; data.py checks IDs and cross-split groups/transcripts; reference.py executes the integer floor; training.py performs optional float64 fitting/PTQ; evaluation.py calibrates and counts; experiment.py writes/replays bundles; cli.py exposes local commands. No network service or security verifier exists. load_records checks uniqueness/leakage; admit additionally requires all splits, data identity and configured source/licence allowlists. Those allowlists are declarations, not approval.

Target boundaries

Boundary Responsibility Inputs → outputs RFC
wire Strict versioned codecs, commitments and envelopes bytes → admitted immutable types 0001
registry Signed snapshots, history, enrollment, retirement trusted roots + snapshot → enrolled artifacts 0002
classifiers Exact byte/token transforms and integer graph execution admitted manifest + text → ordered chunk scores 0003
research Approved lineage, fitting, calibration and held-out reporting frozen protocol + data → candidate/evidence 0004
proofs Bounded real backend adapters enrolled program + witness → proof/result 0005
attestation / gateway Measured loading, genuine evidence and release control admitted request → signed receipt + gated output 0006
receipts / verification Lifecycle, composition and typed claims complete bundle + external trust → result 0007
diagnostics Stable errors, redacted events and costs stages → bounded local metadata 0008
conformance / independent verifier Cross-implementation checking published vectors + genuine artifacts → agreement report 0009
packaging / documentation Private reproducible release and maintenance reviewed source + evidence → review packet 0010

Names specify ownership, not a demand for empty packages. Extend existing modules where the interface remains clear. Keep core verification free of training runtimes. Heavy integrations are optional extras and run in separate processes with explicit budgets.

Data flow

flowchart LR
    E[External trust enrollment] --> R[Registry verification]
    U[Request bytes and nonce] --> G[Measured gateway and model]
    R --> G
    G --> I[Exact classifiers and release gate]
    I --> S[Signed receipt and private proof spool]
    S --> P[Real proof workers]
    S --> B[Offline bundle]
    P --> B
    E --> V[Relying party verifier]
    B --> V
    U --> V

External enrollment fixes the trusted identities before a request. The measured gateway binds the actual model execution and complete classifier inputs, signs the receipt, and stores the private witness for proofs. The assembler collects genuine proofs and hardware collateral. The verifier takes its trust policy and request expectation from outside the bundle and checks the bundle without fetching URLs.

INV-ONE-WRITER: one process owns each receipt's state transition; workers use exclusive job claims. Publishing a completed bundle uses a same-filesystem atomic rename after durable writes. Retries never allocate a new receipt for the same nonce or change the signed body.

INV-FAIL-CLOSED: unavailable keys, collateral, storage, classifier execution or policy never releases a response. Proof failure after release yields an unverified receipt; it does not retroactively make the response verified.

Only the gateway holds epoch signing keys. The spool is encrypted with per-receipt keys sealed to the authorized host, inaccessible to the untrusted operator under the declared hardware assumptions. The offline verifier opens only caller-supplied regular files; it follows no archive paths or embedded URIs.