Articles

Notes on software engineering, leadership, and the craft of building things that work.

Explored through a single lens: engineering rigour as a form of responsibility.

Filtered by:Analysis
Clear filter
Analysis||5 min read

Guardrails over Gates: A Decision Architecture for AI

Why vigilance degradation creates bottlenecks in high-volume loops, and how to replace manual gates with deterministic, auditable guardrails.

Read full article →
Figure 0: Operational Models: The Gate vs. The Guardrail
Operational Models: The Gate vs. The GuardrailA comparative diagram showing two workflows. Left: 'The Gate', where all decisions pass through a human bottleneck, causing high latency. Right: 'The Guardrail', where routine decisions flow automatically through a safe-to-fail zone, and only exceptions are escalated to a human.THE GATE (All Decisions)HumanReviewer(Delayed)High Latency / FatigueTHE GUARDRAIL (Exceptions Only)HumanSafe-to-Fail ZoneEscalateManagement by Exception
Operational Models: The Gate vs. The Guardrail
Analysis||6 min read

Why Your AI System Looks Fine, But Is Not

Operational Decoupling is the symptom; missing Reasoning Integrity is the root cause. How to trace AI behaviour back to human intent to prevent silent drift.

Read full article →
Figure 0: The Core Failure Mechanism
Comparison of traditional vs. AI-augmented system failure modesPanel A shows a traditional system where human intent is coupled to an explicit, encoded constraint, leading to observable failures. Panel B shows an AI-augmented system where optimization metrics bypass un-encoded implicit constraints, leading to silent failures.Panel A: Traditional System (Coupled)Intent: Reduce latencyEngineer encodes constraint:blast radius ≤ XCode + TestBehaviour Aligns(Failure is LOUD)Panel B: AI-Augmented (Decoupled)Intent: Reduce latencyAI Optimises for MetricConstraint: blast radius ≤ X(NOT encoded)Behaviour Violates Intent(Failure is SILENT)BYPASS
Visualising the Decoupling: In Panel B, the optimisation loop bypasses the implicit constraint because it was never encoded as an executable artifact, leading to silent alignment failure. In agentic systems, the failure deepens: not only is the constraint unencoded, but the agent dynamically selects tools and contexts that bypass it, without human visibility.