Briefings
All research notes.
Technical writing on evaluation, monitoring, human oversight, and the evidence needed to operate AI in consequential workflows.
37 published briefings
Newest first
Building an AI Assurance Function: Roles, Skills, and Org Design
You cannot assure a system you are accountable for shipping. Independence is not a personality trait — it is an org chart property, and most AI governance functions are built without it.
Chain of Custody for AI Evaluation Evidence
Two years after an evaluation, someone will ask who scored this, what they saw, and whether anything changed afterwards. If the answer is reconstructed from memory and a shared drive, you do not have evidence. You have a story.
Evaluating Voice AI: Latency, Recognition Error, and Escalation
A voice agent fails in ways a chatbot cannot: it mishears, it interrupts, it pauses too long, and it cannot be scrolled back. Evaluating the language model tells you almost nothing about whether the call went well.
Assessing Third-Party AI: What to Demand From a Vendor
The standard AI vendor questionnaire is a document that both parties know is theatre. Here are the questions that actually discriminate between vendors, the artefacts to demand, and the parts you must verify yourself.
Model Cards and System Cards as Evidence
Most model cards describe. Almost none prove. The difference between a document that satisfies a reader and one that survives an auditor is a small number of properties that are cheap to add and rarely present.
Evaluating AI Coding Assistants
Code that compiles, passes the tests, and looks like something a competent engineer wrote can still be wrong, insecure, or unmaintainable. Pass rate is the easiest metric to collect and the least informative one to report.
Capability Is Not Authorization: Least Privilege for AI Agents
Giving an agent a tool is not the same as authorising it to use that tool, on that object, at that moment, at that scale. Most agent permission models collapse those questions into one — and the blast radius is the difference.
Agent Memory and Context Drift in Long Sessions
An agent that behaves correctly on turn three can behave dangerously on turn thirty, with no change to the model, the prompt, or the tools. State accumulation is a failure mode in its own right — and almost nobody evaluates for it.
How to Manage AI Costs in Production
AI cost management is not just prompt compression. The real lever is measuring cost per resolved workflow, then routing each task to the cheapest model, context, and evaluation path that still meets quality thresholds.
How to Select the Right AI Model
The right AI model is the cheapest model that passes the workflow, safety, privacy, latency, and governance tests for your actual deployment. Here is the selection process.
A Taxonomy of Hallucination — and How to Measure Each Kind
"Hallucination rate: 3%" is a number without a referent. Fabrication, unfaithfulness, and overclaiming have different causes, different fixes, and different measurement methods — and collapsing them into one metric guarantees you fix the wrong one.
Regression Testing for AI: Gates That Actually Block a Release
A CI gate that reports a score and lets the pipeline through is not a gate. It is a logging statement with a dashboard. Here is how to build AI regression tests that fail a build — and what has to be true before anyone will let them.
Evaluating Recommender and Ranking Systems
The ranking model deciding what a million people see today is almost certainly not in your AI governance register. It is older than the LLM, it affects more users, it trains on data it generated itself — and it is evaluated with metrics that flatter it.
Bias Testing Beyond Demographic Parity
Most fairness programmes compute one metric, find it acceptable, and stop. The metrics they did not compute are mathematically incompatible with the one they did — and choosing between them is an ethical decision that engineering cannot make alone.
AI Incident Response: A Runbook
Your incident process assumes a system that is either up or down. AI systems fail while remaining perfectly available — and by the time you notice, thousands of outputs have already left the building. Here is the runbook that difference demands.
Evaluating Vision Models: What Breaks That Text Evaluation Misses
A document AI system can extract every field correctly and still be dangerous, because the numbers it returned came from the wrong table. Vision introduces failure modes that text evaluation has no vocabulary for — and a metric that cannot see them.
Most High-Risk AI Is Not a Chatbot: Evaluating the Whole Model Stack
Public attention fixates on large language models. But most of the AI that regulators classify as high-risk — credit scoring, medical imaging, drug-interaction models, hiring screeners — is not generative at all, and it fails in ways no LLM benchmark measures. Here is what a complete evaluation actually has to cover.
What Is an Agent Harness? The Hidden Variable in AI Agent Performance
The same model can score 20 points apart on the same benchmark depending on one thing almost no one talks about: the harness. Here is what an agent harness is, why it quietly determines agent performance, and why no benchmark number means anything without it.
How to Evaluate AI Agents: Why Agentic Complexity Breaks Traditional Testing
AI agents are not bigger chatbots — they take multi-step actions, and that changes everything about how you test them. Here is why agentic complexity defeats traditional evaluation, what the benchmarks actually reveal, and how to evaluate an agent before you trust it.
How to Validate an LLM Before Production: A Practical Guide
Passing a benchmark is not the same as being production-ready. This is a practical, framework-grounded guide to validating an LLM application before you ship it — the dimensions to test, the RAG triad, OWASP risks, red teaming, and why human review still matters.
When the Approval Button Is Theatre
Human-in-the-loop is the standard answer to AI risk, and most implementations of it are rubber stamps. Here is what separates real oversight from theatre — and what an approval record must capture to prove which one you have.
EU AI Act vs. NIST AI RMF vs. ISO 42001: A 2026 Compliance Comparison
Three frameworks dominate AI governance — one is binding law, one is a voluntary risk framework, and one is a certifiable management standard. Here is exactly how the EU AI Act, NIST AI RMF, and ISO/IEC 42001 differ, what each requires, and where they fit together.
How Many AI Decisions Must You Actually Review?
Reviewing everything is unaffordable. Reviewing only the failures the system already flagged is worse than useless — it guarantees you never find the failures it missed. Here is how to size a risk-based review sample.
Detecting Drift in Production AI: Model, Data, and Policy
An AI system can degrade without anyone changing a line of code. Three distinct kinds of drift — model, data, and policy — produce different signals and demand different responses. Here is how to detect each one and when to re-evaluate.
What to Log Around a Tool Call: An Observability Schema for Agents
Most agent logs record that a tool was called and what it returned. That is enough to debug a crash and nowhere near enough to reconstruct a decision. Here is the field set that separates the two.
Prompt Injection: Defence in Depth
Prompt injection has no clean fix, and every team that treats it as an input-filtering problem eventually learns why. The only durable answer is architectural: assume the instruction gets through, and make sure it cannot do anything.
LLM-as-Judge: Where Automated Judges Work and Where They Fail
An LLM judge is cheap, fast, and consistent — and biased toward length, position, self-similarity, and confident prose. Here is where automated judging is genuinely the right tool, where it is quietly dangerous, and how to build the escalation path between them.
The Non-Negotiable AI Agent Stack
Eight open-source layers every enterprise needs before putting AI agents into production — and why the ninth layer, independent evaluation, is the one you cannot build yourself.
Building Golden Datasets That Don't Rot
Every evaluation rests on a set of examples with known-correct answers. Those examples decay — through leakage, staleness, and quiet contamination — and a rotten golden set produces confident scores that mean nothing. Here is how to build one that survives.
Inter-Rater Reliability: Why AI Evaluation Needs Cohen's Kappa
If two qualified people score the same AI output differently, your evaluation is measuring your people as much as your model. Raw agreement hides this. Kappa exposes it — and disagreement, handled properly, is the most useful signal an evaluation produces.
Evaluating RAG: Faithfulness, Groundedness, and Retrieval Quality
A RAG answer can be true, well-written, and completely unsupported by the documents it cites. Faithfulness, groundedness, and retrieval quality are three different measurements — and conflating them is why RAG evaluations keep passing systems that go on to fail.
Red-Teaming LLM Applications: A Practical Attack Taxonomy
Most LLM red-teaming is a weekend of creative prompting that produces anecdotes instead of coverage. Here is an attack taxonomy, a concrete test matrix, and the discipline that turns adversarial testing into evidence.
Why AI Benchmarks Alone Won't Protect You
Lab benchmarks like MMLU and HumanEval measure capability in controlled conditions. Production failures happen in the messy reality between those conditions. Here is why human evaluation remains the critical safety net.
The EU AI Act: What Evaluation Teams Need to Know in 2026
The EU AI Act is now in force, and conformity assessments require structured evidence of AI system performance. Here is a practical breakdown of what this means for evaluation teams and how to prepare.
What Is an AI Evidence Report? Anatomy of Auditable AI Proof
An AI Evidence Report is the structured, auditable output of an independent AI evaluation — documented findings, severity, and proof that an AI system was tested. Here is what each section contains and why it matters to risk, legal, and compliance teams.
When AI Fails in Healthcare: Lessons from Recent Deployment Incidents
Healthcare AI failures are not hypothetical. Recent incidents in clinical decision support, radiology, and EHR summarisation reveal systematic failure patterns that structured human evaluation should have caught.
Building an AI Evaluation Programme from Scratch
A practical, step-by-step guide for technical leaders building their first structured AI evaluation programme — from defining scope and taxonomy through to continuous monitoring and iterative improvement.
Research and methodology