Skip to content

OPS-RUNBOOK.md

Latest commit

 

History

History
127 lines (86 loc) · 3.06 KB

OPS-RUNBOOK.md

File metadata and controls

127 lines (86 loc) · 3.06 KB
Raw

CRK-1 Operational Piping Runbook (Local Governed Inference)

Objective

Establish a fully governed, end-to-end local inference path using CRK-1:

  • All governed calls flow through a single adapter.
  • Invariants (K0–K12) are enforced pre/post.
  • CE-1 / CRR-1 / CLG-1 are invoked as designed.
  • Receipts + lineage are emitted for every run.
  • No side-door / raw model calls from the operator path.

Target Artifacts

Role Path
Core adapter src/runtime/governedPredict.ts
Local model client src/model/localClient.ts
Invariants src/governance/invariants.ts
Receipts src/governance/receipts.ts
Lineage src/governance/lineage.ts
CLI entrypoint bin/run_local_governed.ts
Config config/local-governed.yaml
Smoke test tests/local_governed_smoke.test.ts

A. Operator Kernel Health

A1. Operator identity

Requirement: every governed call carries a stable operator id.

  • Field: operator_id: string
  • Source: config (config/local-governed.yaml) or env (OPERATOR_ID).
  • Must appear in: GovernedContext, receipts, lineage root.

A2. Receipts

Requirement: every governed call emits a receipt.

Minimum fields:

  • call_id
  • operator_id
  • timestamp
  • invariant_set_version
  • mode (predict / observe / correct)
  • invariants_passed: boolean

Implementation: src/governance/receipts.ts

B. AAIS Local Inference Path

B1. Local-only model client

File: src/model/localClient.ts

export async function localPredict(input: string, opts: LocalModelOptions): Promise<string>;

Uses local weights only (no remote calls). localPredict must only be called from governedPredict.

B2. Governed adapter

File: src/runtime/governedPredict.ts

export async function governedPredict(
  input: string,
  context: GovernedContext
): Promise<GovernedResult>;

Responsibilities:

  1. Load invariants (K0–K12) and invariant_set_version.
  2. Run pre-call admissibility checks.
  3. Call localPredict(...).
  4. Run CE-1 / CRR-1 / CLG-1 post-call hooks.
  5. Emit receipt + lineage entry.

C. End-to-End Governed Pipeline

C1. CLI — bin/run_local_governed.ts

  1. Load config/local-governed.yaml.
  2. Construct GovernedContext.
  3. Call governedPredict(prompt, context).
  4. Print model output, receipt summary, lineage root id.

C2. Config — config/local-governed.yaml

model_path: "./models/local-llm"
operator_id: "local-operator-001"
invariant_set_version: "K0-K12-v1"
log_receipts: true
log_lineage: true

D. Tests

D1. Smoke test — tests/local_governed_smoke.test.ts

Happy path: governedPredict returns output, receipt with invariants_passed: true, lineage root_id.

Refusal path (optional v0.2): prompt that triggers invariant failure.

E. Success Criteria

  • bin/run_local_governed.ts runs using local weights only
  • All calls go through governedPredict
  • Receipts + lineage are emitted
  • tests/local_governed_smoke.test.ts passes
  • No direct/raw model calls from operator-facing code