FAQ
What problem does CFA solve?
CFA addresses the governance gap in AI-native data systems: who decides what an AI agent or data pipeline is allowed to do, under what constraints, and with what evidence? Instead of "which tool should I call?", CFA asks "what state transition is being requested, under what constraints, and can it be executed safely?"
How is CFA different from Great Expectations or Soda?
Great Expectations and Soda validate data quality — they check if data meets expectations. CFA validates execution governance — it checks if an intent (what you want to do) complies with policy before execution. CFA operates at the intent level, not the data level.
How is CFA different from ASSERT or RAMPART?
ASSERT and RAMPART evaluate AI agent behavior — they test if agents produce correct/safe outputs. CFA governs data operations — it validates that any agent or pipeline writing to data systems complies with PII, schema, cost, and partition policies. CFA has a hash-chain audit trail, state projection, and lifecycle indices that none of these tools offer.
How is CFA different from ACS (Agent Control Specification)?
ACS is a policy engine for AI agents — YAML-based, MIT-licensed, backed by KPMG/IBM/Zscaler. Like CFA, it acts as a governance gate before agent actions. The key differences:
| Aspect | ACS | CFA |
|---|---|---|
| State model | Stateless (allow/deny) | Stateful (StateSignature + ContextRegistry) |
| Decision | Binary (pass/fail) | Ternary (approve/replan/block) |
| Audit trail | Standard logs | SHA-256 hash chain (tamper-evident) |
| Lifecycle | Not present | IFo/IFs/IFg/IDI indices |
| State projection | No | Yes (post-execution) |
| Backend targets | Agent actions | Data pipelines + agents (PySpark, SQL, dbt) |
ACS excels at agent safety in enterprise environments. CFA adds stateful governance across both agents and data pipelines with cryptographic auditability.
Does CFA require an LLM?
No. The default IntentNormalizer uses a deterministic mock backend (keyword matching). The Policy Engine, Planner, CodeGen, Sandbox, and Validators all operate without LLMs. LLM backends are optional plugins.
What languages/runtimes does CFA support?
CFA core is Python 3.11+ with zero mandatory runtime dependencies. The default code generation target is PySpark (Delta Lake). The backend registry supports plugging in any target (DuckDB, BigQuery, SQL, REST APIs).
Can I use CFA in CI/CD?
Yes. cfa evaluate --exit-code returns exit code 1 when an intent is BLOCKED, making it suitable for GitHub Actions, GitLab CI, or any CI pipeline. The repository includes an example workflow at .github/workflows/governance.yml.
What is the REPLAN state?
REPLAN is CFA's unique intermediate state between APPROVE and BLOCK. When a policy rule fires with action: replan, the kernel automatically applies corrective interventions to the signature (e.g., adding partition filters, enabling PII anonymization) and re-evaluates. This happens up to 3 times before terminal BLOCK.
What are IFo, IFs, IFg, and IDI?
Four quantitative lifecycle indices:
- IFo (Fluidez Operacional):
(1 - norm_latency) × (1 - norm_cost) × success_rate - IFs (Fidelidade Semântica):
schema_match × (1 - replan_rate) × fault_free_rate - IFg (Governança): binary — 1.0 if no violations, 0.0 otherwise
- IDI (Intent Drift):
1 - (replanned / total)over a time window
These feed the Promotion Engine: ACTIVE (IFo ≥ 0.75, IFs ≥ 0.90, IFg = 1.0), WATCHLIST (IDI < 0.75), DEMOTED (IDI < 0.50).
How does the audit trail work?
Every governance event (normalization, confirmation, policy evaluation, replan, code generation, validation, execution, projection, decision) is recorded as an AuditEvent with a SHA-256 hash. Each event links to the previous event's hash, forming a tamper-evident chain. cfa audit verify checks the chain integrity.
What is State Projection (I4)?
Invariant I4: after every execution (success, partial, or failure), the StateProjectionProtocol updates the ContextRegistry with the new state of all affected datasets. This means the system always knows what happened and can make informed decisions about subsequent intents.
Can CFA handle streaming data?
CFA v1.0.0 targets batch and micro-batch execution. Streaming support is planned for a future version. The architecture's concurrency model assumes single_active_intent_per_target_scope.
Is CFA production-ready?
CFA v1.0.0 has 534 passing tests, a full CLI, MCP server, rich reporting, and framework adapters. It is designed for production governance gates. The core has been battle-tested through multiple refactors. As with any governance tool, deploy with monitoring and start with cfa evaluate in CI before moving to runtime gates.