Define the eval objective

Contract: references/interaction-contract.md. Calibration, templates, provenance: references/objective-card.md.

Trigger

• Nobody can name the decision the eval supports.
• A benchmark score being used to argue a claim it may not license.
• A product goal ("make the agent more helpful") that needs to become measurable.

Do

1. State the decision in one sentence. If several are plausible, ask which one the eval must support — the highest-information question available here.
2. Write one sentence each: construct, population (and its weighting), intended use and non-use.
3. Answer both questions, not just the first. Verification: do the scorers hold the spec? Validation: do these scorers reflect outcomes users actually need? Validation is the half teams skip.
4. Where validation has no evidence behind it, say so and elicit criteria from experts and real user signals rather than asserting what users want. Label unevidenced desires Assumed.
5. Write supported and unsupported claims as separate lists, then name the evidence that would change the decision. An objective no result could overturn is not an objective.

Avoid

• Do not pick metrics, thresholds, or scoring methods.
• Do not let the population drift toward whatever data is easy to collect; most bad datasets fail here, before a single item exists.
• Do not accept a benchmark name in place of a product property.
• Do not imply model-level claims when the object under test is a system — harness, tools, and context are part of what you measure.

Check

• Decision, construct, population, use, and non-use each in a sentence.
• Supported and unsupported claims listed separately.
• Both V&V questions answered, with validation's evidence source named.

Risk

• An ambiguous decision produces an eval that is locally measurable and strategically irrelevant.
• A construct defined by convenience gets optimized: the score improves, the product does not.
• Loose population statements ("our users") license claims far broader than the sample supports, and nobody notices until the number goes external.

Braintrust

Write the decision and intended claim into the experiment description so the pre-registered claim travels with its results — a claim in a doc gets reinterpreted after the fact. Record the population and its weighting in dataset metadata. Make non-use enforceable: name datasets for their population (prod-weighted-py-backend, not eval-set-2) so quoting a number out of scope is visibly wrong, and keep benchmark scaffolding in separate, clearly named datasets — once mixed with production-derived items, "measures our construct" vs. "measures the benchmark's" is unrecoverable. Validation cannot be answered offline: it needs production scoring sliced by escalation, reopen, and abandonment. If nothing scores production yet, validation is Not yet measurable, and standing up online scoring is the blocker.