Proof Before Production
Evaluation systems that can stop a release
A probabilistic system cannot be shipped on taste, so evaluation has to be built like a build system: versioned datasets, trajectory scoring, judges you have tested, and gates wired into CI.
The demo passes because someone chose the inputs. The product fails quietly because production did not. This book builds the apparatus that closes the gap: a definition of done for open-ended output, a first dataset drawn from real traffic, metrics separated into outcome, trajectory and system, a judge calibrated against human annotation that carries its own error bar, and a suite with the authority to fail a deploy. It also draws the line where the OpenTelemetry GenAI conventions stop and your own evaluation data has to start.
Written for: Engineering leads who can demo an AI feature and cannot defend releasing it.
No chapter is published yet. The full contents below is the finished outline; chapters go live as they are written, and none of them will sit behind a signup.
What it makes operable
Outcomes, not topic coverage.
- 01Define done for an open-ended output before writing a single eval
- 02Build the first dataset out of real traffic and keep the seed set uncontaminated
- 03Score trajectories, then calibrate the judge against a human-annotated reference set
- 04Wire the suite into CI as a gate with the authority to fail a deploy
Complete contents
12 chapters, in order.
Each chapter settles one decision, so it can be read on its own. Read in sequence to build the argument the book is making.
- 01In writing
Why Demos Pass and Products Fail
The distribution a demo samples, and the one production actually sends.
- 02In writing
Defining Done for an Open-Ended Output
Turning "a good answer" into criteria two reviewers would agree on.
- 03In writing
Building the First Dataset From Real Traffic
Sampling, labelling and splitting a seed set you will not contaminate.
- 04In writing
Outcome, Trajectory, and System Metrics
Three families of measurement, and what each one cannot tell you.
- 05In writing
Scoring the Path, Not Just the Answer
Path-invariant assertions for runs that legitimately take different routes.
- 06In writing
LLM Judges Need Their Own Test Sets
The judge as a system under test, against a human-annotated reference set.
- 07In writing
Calibration, Bias, and Disagreement With Humans
Chance-corrected agreement, bias probes, and a published error bar per score.
- 08In writing
Benchmarks Saturate; Your Suite Should Not
Why public scores stopped discriminating, and what to grow in their place.
- 09In writing
Wiring Evals Into CI as a Release Gate
Thresholds, regression rules, and who is allowed to override a red run.
- 10In writing
Online Evaluation and Guarded Rollouts
Shadow traffic, canaries, and the signal that triggers a rollback.
- 11In writing
Tracing With OpenTelemetry GenAI Conventions
The span hierarchy the conventions define, and where they deliberately stop.
- 12In writing
Incident Response for Systems That Are Fluently Wrong
Detecting, containing and explaining an error that reads as confident.
Evidence
What this book is arguing against.
Third-party sources behind the book's premise. Every figure in them belongs to the party that published it and is attributed to them in the text.
- AgentLens: Production-Assessed Trajectory Reviews for Coding Agent Evaluation
arXiv · 2026-07-07 · Research paper · reported
- LLM-as-a-Judge: Why Frontier Models Fail 50%+ Bias Tests
Adaline · 2026-04-08 · Industry report · reported
- AI Benchmarks 2026: Top Evaluations and Their Limits
Kili Technology · 2026-04-13 · Industry report · reported
- Datadog Agent Observability natively supports OpenTelemetry GenAI Semantic Conventions
Datadog · 2025-12-01 · Vendor engineering · reported
Continue through the library
Adjacent operating manuals.
The Agent Trust Boundary
Assume the injection succeeds. That single assumption reorganises the whole problem: security stops being a matter of instruction wording and becomes a question of identity, authority and containment. This book draws the boundary — who the agent is, what it may call, where its code runs, what it may reach, what it may remember — and places enforcement outside the model, at the gateway, the tool or the destination system. It ends with the evidence an auditor asks for, because from 2 August 2026 some teams have to produce it.
Read the full book
AI-Native Product Architecture
Most AI features fail for reasons that have nothing to do with the model. They fail because nobody decided whether the system suggests, drafts or acts, and so nobody designed the undo, the abstention, the escalation or the cost ceiling that each of those shapes implies. This book makes the interaction shape the first decision and derives the rest from it: reversibility, confidence display, routing, unit economics, latency budgets, caching that cannot serve a wrong answer faster, and the evidence that tells you to cut the feature.
Read the full book