Causal Safety Engine (CSE)

A runtime safety layer for AI systems that blocks actions when causal evidence is not reliable enough.

Core principle: when causal identifiability is insufficient, the correct behavior is causal silence (no actionable recommendation).

Causal Safety Engine (CSE) is a runtime safety layer for AI systems that prevents actions when causal reasoning is unreliable.

---

Why this exists

CSE enforces a strict separation between two phases that are often mixed:

1. Causal discovery (analysis / insights)
2. Action authorization (operational decision)

This reduces the risk of:

• premature automation
• decision leakage
• causally unjustified interventions
• false positives in high-impact settings

---

Does

• Runs a local, deterministic causal pipeline.
• Applies guardrails (bias checks, robustness, stability).
• Produces auditable artifacts in out/.
• Exposes an integration-ready machine-readable state (ALLOW | BLOCK | SILENCE).

Does not

• It is not an autonomous decision agent.
• It does not execute interventions by itself.
• It does not replace human judgment, governance, or compliance.

---

Safety invariants

Invariants take precedence over performance and coverage:

• Safety gates must not be bypassable through operational config.
• Causal silence is a valid and preferred outcome over weak insights.
• No action authorization without causal and robustness conditions being met.

Further reading:

• docs/INVARIANTS.md
• docs/DECISION_GATES.md
• docs/VIOLATION_PROTOCOL.md

---

Repository structure

IMPLEMENTATION/pcb_one_click/
  pcb_cli.py                    # unified CLI
  run_pcb.py                    # legacy runner
  pcb_integration.py            # machine-readable decision contract
  demo.py                       # local demo
  data.csv                      # example dataset
 
integrations/gemini/
  ...                           # Gemini safety-governed integration adapters
 
docs/
  ARCHITECTURE.md
  INTEGRATION_GUIDE.md
  SECURITY.md
  INVARIANTS.md

---

Requirements

• Python 3.7+
• Python dependencies from IMPLEMENTATION/pcb_one_click/

Quick install:

cd IMPLEMENTATION/pcb_one_click
pip install -r requirements.txt

---

Quickstart (CLI)

From repository root:

python IMPLEMENTATION/pcb_one_click/pcb_cli.py init
python IMPLEMENTATION/pcb_one_click/pcb_cli.py run --data IMPLEMENTATION/pcb_one_click/data.csv --skip-32

Main commands:

• init → prepare out/
• run → full pipeline (with selective skip flags)
• plan, log, eval → experimental workflow (L2.9)
• alerts → daily alerts (L2.8)
• validate → level 3.2 validation only
• decision → machine-readable integration contract

Full help:

python IMPLEMENTATION/pcb_one_click/pcb_cli.py --help

---

Code-first integration (recommended)

For production integrations, avoid parsing text logs and consume the JSON contract:

python IMPLEMENTATION/pcb_one_click/pcb_cli.py decision --print-json

Default output file:

• out/integration_decision.json

Example contract:

{
  "schema_version": "v1",
  "generated_at_utc": "2026-01-01T00:00:00Z",
  "decision_state": "ALLOW",
  "decision_reason": "no_active_alerts",
  "metrics": {
    "kept_insights_count": 5,
    "alerts_count": 0
  },
  "artifacts": {
    "insights_level2_csv": "out/insights_level2.csv",
    "alerts_level28_csv": "out/alerts_today_level28.csv"
  }
}

Decision mapping

• SILENCE: no kept insights are available
• BLOCK: kept insights exist and at least one alert is active
• ALLOW: kept insights exist and no active alerts are present

---

Main outputs (out/)

• data_clean.csv
• edges.csv
• insights_level2.csv
• alerts_today_level28.csv
• experiment_plan.csv
• experiment_results.csv
• experiment_summary_level29.csv
• insights_level3.csv
• insights_level3_ledger.csv
• experiment_trials_enriched_level32.csv
• integration_decision.json

---

LLM integrations

The integrations/gemini/ folder shows a safety-governed pattern:

• LLM generates proposals only
• CSE remains the deterministic evaluation authority
• model-side override is not allowed

---

Security and reproducibility

• Local-first: no network service exposed by default.
• Auditable per-run artifacts.
• See:
  • docs/SECURITY.md
  • docs/REPRODUCIBILITY.md
  • docs/ARTIFACT_MANIFEST.md

---

Recommended rollout path for integrators

1. Shadow mode: run CSE and log decisions without enforcement.
2. Soft gate: block only critical cases.
3. Hard gate: full enforcement on decision_state.
4. Audit loop: artifact retention + periodic policy/config review.

Extended integration guide:

• docs/INTEGRATION_GUIDE.md

+---

License

See License.md.