v1.0.0 ships five patterns
- DEFCON State Machine — 5-level risk degradation with hysteresis
- Sovereign Veto — human-only kill switch
- Audit Chain — tamper-evident SHA-256 hash-chain logging
- Autonomy Ladder (A0→A4) — governance classification framework
- EU AI Act mapping — Article 9–15 control mapping
What's being considered for v1.1 and beyond
The following patterns have been designed and documented internally. Publishing priority will be determined by community demand:
Shadow Mode Rollout (patterns/shadow_mode.py)
Run a new agent version in parallel with production, comparing outputs without executing. The go/no-go decision is made by a human reviewing the divergence log. Essential for regulated environments where you cannot A/B test with real capital.
Circuit Breaker (patterns/circuit_breaker.py)
Auto-suspend a specific agent (not the whole system) when its individual error rate, latency, or output anomaly score crosses a threshold. Finer-grained than the system-level DEFCON machine.
Regulatory Reporting Hook (patterns/reg_reporting.py)
Standard interface for capturing the data required by SEC Rule 17a-4, MiFID II Article 17, and similar regulations. Not the reporting itself — the hook that ensures the right data is captured at the right moment.
FCA / MAS / APRA Compliance Mappings (docs/)
Same structure as the EU AI Act mapping, extended to UK FCA, Singapore MAS, and Australian APRA AI governance requirements.
Multi-Agent Consensus Pattern (patterns/consensus.py)
When multiple agents must agree before an action executes — preventing a single compromised or malfunctioning agent from driving the system into a bad state.
Your vote counts
Which pattern would your team reach for first? Comment below — the top requests ship first.
Also open to patterns not listed above. If you've hit a governance failure mode in a regulated environment that isn't covered here, describe it. That's exactly how this library grows.
v1.0.0 ships five patterns
What's being considered for v1.1 and beyond
The following patterns have been designed and documented internally. Publishing priority will be determined by community demand:
Shadow Mode Rollout (
patterns/shadow_mode.py)Run a new agent version in parallel with production, comparing outputs without executing. The go/no-go decision is made by a human reviewing the divergence log. Essential for regulated environments where you cannot A/B test with real capital.
Circuit Breaker (
patterns/circuit_breaker.py)Auto-suspend a specific agent (not the whole system) when its individual error rate, latency, or output anomaly score crosses a threshold. Finer-grained than the system-level DEFCON machine.
Regulatory Reporting Hook (
patterns/reg_reporting.py)Standard interface for capturing the data required by SEC Rule 17a-4, MiFID II Article 17, and similar regulations. Not the reporting itself — the hook that ensures the right data is captured at the right moment.
FCA / MAS / APRA Compliance Mappings (
docs/)Same structure as the EU AI Act mapping, extended to UK FCA, Singapore MAS, and Australian APRA AI governance requirements.
Multi-Agent Consensus Pattern (
patterns/consensus.py)When multiple agents must agree before an action executes — preventing a single compromised or malfunctioning agent from driving the system into a bad state.
Your vote counts
Which pattern would your team reach for first? Comment below — the top requests ship first.
Also open to patterns not listed above. If you've hit a governance failure mode in a regulated environment that isn't covered here, describe it. That's exactly how this library grows.