Interpretive agentic reference — executable proof (non-normative)

This repository provides a minimal executable reference pipeline demonstrating how constraintive governance operates as a runtime substrate for interpretive governance in agentic-closed systems.

This is not a product, not a framework, and not a certification tool.

What this demonstrates

A single pipeline with four non-negotiable properties:

1. Bounded retrieval (allowlisted sources only)
2. Fixed inference configuration (runtime parameters are set by the orchestrator)
3. Schema-validated output (reject invalid outputs)
4. Policy-driven abstention (legitimate non-response when conditions fail)

The purpose is to show that constraintive governance is not a prompt technique. It is a runtime configuration and enforcement layer.

What this does not claim

• No claim of correctness, completeness, or safety.
• No claim that a model “follows” governance.
• No claim that outputs are “approved”.
• No claim that this implementation is production-ready.

Regime boundary

• Web-open: constraintive governance is inapplicable (no runtime control).
• Agentic-closed: constraintive governance is applicable (runtime control exists).

This repository is explicitly scoped to agentic-closed systems.

See:

• docs/regime-boundary.md
• USAGE.md (doctrinal usage notice)
• docs/allowlist-maintenance.md (non-normative operational notes)

Architecture (single pipeline)

flowchart TD
  A[Typed request] --> B[Bounded retrieval<br/>allowlist]
  B --> C[LLM execution<br/>runtime config fixed]
  C --> D[Schema validation]
  D -->|valid| E[Accept]
  D -->|invalid| F[Abstain]
  B -->|missing/denied| F
  A -->|out of scope| F

Loading

Typed request → bounded retrieval (allowlist) → LLM execution (runtime parameters fixed) → schema validation → accept OR abstain

See: docs/architecture.md

Relationship to the manifest

This repository is a non-normative reference implementation. The normative definition of Interpretive Governance is maintained in:

https://github.com/GautierDorval/interpretive-governance-manifest

This repository must not redefine that manifest.

Where Layer 1 shows up (interpretive typing)

The interpretive typing primitives (observed / derived / inferred / unknown) are expressed here as:

• a strict output schema: configs/output_schema.json
• a governed output model: src/types.py
• schema validation / rejection logic: src/validator.py
• explicit legitimate non-response: configs/abstention_policy.json

Quickstart

1) Setup

Python 3.11+ recommended.

pip install -r requirements.txt

2) Run in mock mode (no API required)

python -m src.main --request examples/requests/company_profile.json --mock

Note: mock mode demonstrates enforcement mechanics only. It does not represent provider latency, cost, failures, or behavior variability.

3) Run with a real LLM adapter (optional)

The adapter contract is defined in src/llm_adapter.py. A minimal contract explanation is in docs/llm-adapter-contract.md.

You can load an adapter class using:

python -m src.main --request examples/requests/company_profile.json --adapter your_module:YourAdapter

Notes:

• The orchestrator still binds runtime parameters (configs/runtime.json).
• Output must be strict JSON matching configs/output_schema.json.

Core files

• configs/runtime.json — runtime constraints (temperature, top_p, max_tokens)
• configs/retrieval_allowlist.json — allowed sources for retrieval
• configs/output_schema.json — strict output schema (observed / derived / inferred / unknown)
• configs/abstention_policy.json — legitimate non-response conditions

Abstention (legitimate non-response)

The abstention policy is configured in configs/abstention_policy.json.

By default, the pipeline abstains if any of the following conditions are true:

• required sources are missing or denied by the allowlist,
• sources are conflicting (stubbed in this reference implementation),
• the request is out of scope (minimal example rule in src/pipeline.py),
• the model output is invalid JSON or fails schema validation.

Abstention is represented as:

• abstained: true in the governed output,
• an explicit message (default: LEGITIMATE_NON_RESPONSE) inside unknown.

Examples

• Allowlisted retrieval sources: examples/sources/
• Typed requests: examples/requests/

CI and tests

• JSON configs and example requests are validated on CI.
• The pipeline is covered by a minimal unittest suite (tests/).

Run locally:

python -m unittest discover -s tests -v

License

MIT License. See: LICENSE