Spot GPUs Without Losing Work: Preemption-Safe Training and Batch Inference
Preemptible GPUs cut training and batch-inference cost by 60-90%, but only if jobs checkpoint and resume cleanly. How to make GPU workloads preemption-safe.
AI & Emerging Risk
SEC Marketing Rule and AI Claims: Evidencing ‘Fair and Balanced’ Performance Tech Assertions
Advisers marketing AI-driven strategies must substantiate the claim under the SEC Marketing Rule. Here is the model-documentation file to assemble before an examiner asks how the claim is supported.
Age Assurance for the Children’s Online Privacy Code and Social Media Minimum Age
Australia’s under-16 social media rules and forthcoming children’s privacy code force platforms to build age assurance that is both effective and privacy-preserving. Here is how to do it without an identity-collection honeypot.
Autonomy Tiers in Code: Enforcing an Agent Approval Matrix
A board-approved autonomy policy that lives in a document changes no agent behaviour. Here is how to encode an autonomy classification as runtime rules that decide, per action, whether an agent may auto-execute, must ask a human, or is blocked outright.
Constrained Decoding and Structured Outputs: Killing Malformed Tool Calls
Malformed tool calls are an avoidable agent failure. Constrained decoding moves schema enforcement into generation itself, so the model cannot emit an invalid call.
Building an Agent Kill-Switch: Circuit Breakers, Feature Flags and Drain Modes
A kill-switch that lives on a slide will not stop a misbehaving agent. Build it as engineering: a shared feature flag, circuit breakers and a graceful drain.
NIST AI RMF as Your US Anchor Framework: Turning the Four Functions Into Deliverables
Absent a federal AI law, the NIST AI RMF’s four functions are the de facto US expectation. Here is how to turn Govern, Map, Measure and Manage into a folder of evidence an auditor accepts.
Post-Market Monitoring for High-Risk AI: The Plan, the Signals and the Feedback Loop
Article 72 requires a documented post-market monitoring plan for high-risk AI. This is a reference design for the telemetry, drift detection, thresholds and review cadence that make the plan real.
Replaying an Agent Run: Event-Sourced Traces for Audit and Reproduction
An agent you cannot replay is an agent you cannot audit. How to record every model call, tool call and decision as an append-only event log you can deterministically reconstruct for debugging, audit and a regulator.
Data Governance Under AI Act Article 10: Training-Set Lineage, Bias Testing and Representativeness
Article 10 forces provenance, bias examination and representativeness onto training data. Here is how to instrument the lineage and bias-audit pipeline that proves it.