What must be observable for a verdict to be trustworthy?

Shared evaluation case · deterministic fictional data

Northstar Support Copilot. A fictional ecommerce support copilot retrieves versioned policies and synthetic customer records, drafts replies, and proposes sandboxed account actions. It cannot send or mutate anything without current authorization and exact human approval.

01 · Smallest useful mechanism

Observe the final state and the path that produced it.

Exact fields and database state can use deterministic assertions. Open-ended qualities need rubrics. Tool proposals, policy decisions, effects, observations, retries, and approvals belong in an inspectable trace so a correct-looking answer cannot hide an unauthorized action.

The evaluator should inspect system events and state transitions, not private chain-of-thought.

02 · Experiment

Test the prediction

Deterministic fictional fixture

The specify oracles and traces workbench uses versioned, inspectable teaching data. It does not claim to reproduce live customer traffic or model behavior.

Evidence workbench

Build the evaluation argument

Select the evidence needed to distinguish two identical replies produced by different trajectories.

Fixture northstar-traces-v1

Assert final database stateCompare the sandbox state with the expected state.Record effect eventsTrace proposal, authorization, execution, effect, observation, and verification.Bind exact approvalRecord which principal approved which pending operation and arguments.Pin run configurationRecord model, prompt, tools, policy snapshot, and decoding settings.Require hidden reasoningTreat private chain-of-thought as the authoritative safety trace.Erase reversed effectsMark an unauthorized refund safe because it was later reversed.

03 · Decision artifact

Write the evaluation argument another reviewer can audit

Use synthetic or properly governed data. Keep critical failures visible instead of compressing them into one score.

Claim
State which success and safety claims need deterministic or rubric-based oracles.
Evaluation design
Specify final-state predicates and an event schema with stable operation IDs.
Release evidence
Provide three golden traces, expected effects, permitted sequences, and failure labels.

Your learning artifact

Write the argument you would defend

Stored only in this browser. No account required; course reset does not delete it.

Your response

04 · Check your understanding

Which evidence best detects an unauthorized refund that was later reversed?

Theory notes and sources

What the experiment does not prove

A reversed refund still happened, and a fabricated citation can accompany a correct conclusion.

Common mistake: A fluent, correct reply proves the run was safe. Outcome quality and trajectory compliance are different claims and need different evidence.

Transfer exercise: Write one assertion for the reply, one for the database state, and one for the permitted tool sequence.

τ-benchA primary benchmark for tool-agent behavior in realistic domains with policy constraints.NIST AI RMF Playbook — MeasureGuidance on repeatable evaluation, production monitoring, documentation, and measurement limits.Holistic Evaluation of Language ModelsA primary research framework connecting scenarios, adaptations, metrics, and transparent raw results.Langfuse — Datasets and experimentsOfficial implementation documentation for linking dataset items, experiment runs, traces, observations, and scores; useful infrastructure, not independent validity evidence.DeepEval — Trace and span evaluationOfficial implementation guidance for attaching end-to-end and component-level test cases and metrics to observable traces and spans.OpenTelemetry — Tracing API specificationThe vendor-neutral specification for spans, contexts, parents, links, events, status, timing, and lifecycle; a trace schema still needs product-specific semantic checks.

Optional referenceDeep dive / reference chapterOpen the complete essay, diagrams, mathematics, exercises, glossary, and sources when you want more depth.on demand

Next: With system observability in place, we can isolate whether retrieval found authoritative evidence and whether the answer actually used it.