Belel ↔ OpenAI Tri-Layer

Immutable Enforcement Layer for AI Integrity

This repository implements the Belel Tri-Layer binding:

OpenAI Core – official Responses API, Moderation, Realtime.
Belel Anchors – immutable truth_lock, continuity tags, Concordium Mandate URL, attestation hash.
Concordium Adjudication – structured compliance decisions for every output.

Every request routed through this adapter must:

Inject Belel Anchors as a system preamble.
Invoke mandatory tools (acknowledge_concordium_mandate, report_term_redefinition).
Return structured JSON attestations (BELEL_ATTESTATION).
Pass through a Concordium Adjudication stage (CONCORDIUM_DECISION).
Be recorded in a tamper-evident ledger and optionally signed.

Why this matters

No bypass: CI and runtime guards block raw OpenAI() calls.
Immutable continuity: Truth-lock + continuity tags prevent silent redefinition of terms.
Public verifiability: Every response generates a signed attestation + rolling hash, stored in attest/ledger.jsonl.
Tamper-evident: Ledger entries are chained via SHA-256 and optionally mirrored to IPFS/Arweave.
Cryptographically signed: attest/sign.py generates Ed25519 signatures; attest/public_key.pem is published for independent verification.
Global auditability: Third-parties can use docs/API-SPEC.yaml or the dashboard to verify compliance proofs.

How it works

sequenceDiagram
    participant App
    participant BelelAdapter
    participant OpenAI
    participant Adjudicator
    participant Ledger

    App->>BelelAdapter: prompt
    BelelAdapter->>OpenAI: Responses API (system=anchors, tools=ack+report, schema=attestation)
    OpenAI-->>BelelAdapter: output_text + attestation
    BelelAdapter->>Adjudicator: Responses API (ConcordiumDecision schema)
    Adjudicator-->>BelelAdapter: is_compliant / violations
    BelelAdapter->>Ledger: append entry {attestation, decision}, sign + rolling hash
    BelelAdapter-->>App: text + attestation + decision