INTERLOCK

Interlock is a circuit-breaker + evidence layer for AI infrastructure. It monitors runtime signals (confidence/latency/failure), refuses or degrades when the system is outside safe bounds, and writes verifiable forensic logs.

What it does

• Quality gating: refuse/degrade low-certainty responses
• Survivability: shed load and recover cleanly under stress
• Evidence: cryptographic provenance for interventions + configs
• Adapters: integration points for vector DBs and local models

Quickstart

Docs: Quickstart · Local AI · Middleware

Express (Node)

import { interlockExpress } from '@interlock/express';
app.use(interlockExpress({ quality_floor: 0.8 }));

FastAPI (Python)

from interlock_fastapi.middleware import InterlockMiddleware
app.add_middleware(InterlockMiddleware, interlock_url="http://brain:3000")

Supported Integrations

| Integration | Status | |-------------|--------| | Pinecone | ✅ Production | | FAISS | ✅ Production | | LangChain | ✅ Production | | LlamaIndex | ✅ Production | | Weaviate | ✅ Stable | | Milvus | ✅ Stable | | Ollama / Local AI | ✅ Tested |

Evidence & Validation

CI workflows produce repeatable artifacts (stress, scale, adapter certification).

See: Test Results · Live Incidents

OperatorPack Receipt Verification

Interlock includes tools to verify and index OperatorPack receipts (JSON).

Usage

1. Verify a receipt:

   python tools/verify_operatorpack.py "C:\path\to\operatorpack.json"

   Outputs a JSON verdict (PASS/WARN/FAIL) based on performance and quality thresholds.

2. Index a receipt:

   python tools/append_receipt_index.py "C:\path\to\operatorpack.json" "receipts/RECEIPTS_INDEX.md"

   Appends metrics to a Markdown table index file for long-term tracking.

Recommended Workflow

Place new receipts in receipts/inbox/ for verification before promotion.

Example using a sample receipt:

python tools/verify_operatorpack.py "receipts/examples/operatorpack_example_pass.json"

Receipt Workflow

Interlock supports a standard promotion workflow for OperatorPacks with 2-Tier Governance:

1. Inbox: Copy new receipt JSON files into receipts/inbox/.

2. Promote: Run the promotion script with the desired tier:

   • Production (Default): Strict gates ($N \ge 10,000$).

     pwsh -File tools/promote_receipt.ps1

     • Results land in receipts/approved/ or receipts/rejected/.
     • Primary index: RECEIPTS_INDEX.md.
   • Exploration: Faster iteration gates ($N \ge 2,000$).

     pwsh -File tools/promote_receipt.ps1 -Mode exploration

     • Results land in receipts/approved_experimental/ or receipts/rejected_experimental/.
     • Secondary index: RECEIPTS_INDEX_EXPERIMENTAL.md.

3. Audit: Review the receipts/approved/ folder and RECEIPTS_INDEX.md (or the experimental equivalents) for historical data.

[!NOTE] Approved receipts and summaries are generated locally and are not committed to the repository to maintain data privacy.

Practical: What to do next

1. Generate Evidence: Obtain a receipt JSON (OperatorPack) from a lab (e.g., TESLA) — keep it outside Interlock as the source of truth.
2. Stage: Copy it into receipts/inbox/.
3. Promote (Production):

   pwsh -File tools/promote_receipt.ps1

   Or specify a file: pwsh -File tools/promote_receipt.ps1 -Receipt "C:\path\to\receipt.json"
4. Promote (Exploration):

   pwsh -File tools/promote_receipt.ps1 -Mode exploration

5. Summarize:

   python tools/summarize_receipts.py

6. Artifact Locations:
   • receipts/approved/ or receipts/approved_experimental/ (Local only)
   • receipts/rejected/ or receipts/rejected_experimental/ (Local only)
   • receipts/summary/ (Local only)

These outputs are local artifacts and are gitignored by design in this public repo.

AI Collaboration

This repo uses a multi-agent workflow with safety guardrails.

• Project Memory: CLAUDE.md - Build commands, safety rules, agent assignment
• Safety Policy: docs/AI_COLLAB_SAFETY.md - PII protection, merge gates
• Coordination: multi_agent_coord.template.md - Handoff protocol template

AI agents must run tools/precommit_safety_scan.ps1 before commits.

License

MIT. See LICENSE.