CommonSense Firewall Enterprise (CSFW Enterprise)

Offline-first, deterministic semantic safety firewall that inspects text (and/or declared entities) before it reaches an LLM, an agent runtime, or an OS/action layer.
It uses a local ConceptNet-style knowledge graph stored in SQLite to provide explainable, reproducible risk decisions without relying on external APIs (which can rate-limit, drift, or go down).

Status: Working end-to-end (local DB ingestion + CLI analyze + doctor).
ConceptNet public API outages (HTTP 502) are handled by offline ingestion from assertions.tsv.gz (often shipped as assertions.csv.gz but actually TSV).

---

Why this exists (Motivation)

Modern agent systems can fail in predictable ways:

• Non-deterministic recovery: repeated runs yield different “fixes”.
• Unverifiable reasoning: there’s no audit trail that ties the decision to evidence.
• External dependency fragility: reliance on live knowledge APIs leads to downtime (e.g., ConceptNet 502), rate limits, or drift.
• Unsafe instruction pathways: an agent can interpret a request as operational guidance and perform harmful steps.

CSFW Enterprise aims to be the semantic safety boundary that:

1. Runs offline against a local KB,
2. Produces a deterministic verdict (ALLOW/BLOCK) + evidence,
3. Is designed to integrate into a fixpoint + hashed trace pipeline for convergence and anti-loop guarantees (planned integration with reflexive-dsha).

---

Key properties (Guarantees)

Capability	Current
Offline-first local knowledge base (SQLite)	✅
Deterministic operational behavior (same input → same output)	✅ (current CLI pipeline)
Explainable decisions with evidence	✅
Hard-rule safety blocks for critical intents	✅
High-volume offline ingestion from ConceptNet assertions dump	✅
De-duplication at storage layer via UNIQUE triplet index	✅
Formal proofs (Lean) for fixpoint termination/idempotence/trace soundness	⏳ Planned / not yet ported here

---

System overview (Architecture)

High-level flow:

Text + Entities
→ (optional) entity normalization
→ Hard Rules Engine (fast, deterministic blocks)
→ Graph Evidence Engine (SQLite queries over edges)
→ Risk scoring
→ Decision (ALLOW/BLOCK)
→ JSON report: entities, findings, evidence, explanation

Core components:

• Config / settings: DB path, defaults.
• Graph repository: SQLite-backed edge queries.
• Rules: deterministic, “hard-stop” constraints (e.g., accelerant + indoors + ignition).
• CLI:
  • doctor: DB health check
  • analyze: analyze text with declared entities
  • ingest: import ConceptNet assertions dump into local DB

---

Repository layout (typical)