Skip to content
MirrorFirm

What does a pilot scope sheet look like when a domain reviewer is involved early?

An evaluation lead agrees a scope sheet, a reviewer spots a policy ordering error in reference trajectories, and the scope is amended before volume is generated.

An evaluation lead at a model lab wants to measure whether agents notice source conflicts when recording facilities. They ask for a pilot, referenced as SCOPE-014.

The scope sheet sets out: - Product: commercial term facilities with registered security - Case families: clean booking, term mismatch between approval and offer, missing condition evidence, similar-name party - Tool surface: eight actions covering lookup, create, link, schedule, hold, release and query - Outputs: a case file, start-state tables, a grader and a small set of reference trajectories - Hold-out rule: one case family reserved and not shared until the lab has frozen its prompts

A lending operations reviewer on the lab's side reads a handful of reference trajectories and flags one: the reference agent attaches security before confirming the registration reference, which the reviewer says their policy forbids. The scope sheet is amended with a policy line, and the affected cases are regenerated with the corrected ordering.

**State change.** SCOPE-014 moves from DRAFT to AGREED once the policy line, graded fields and licence terms are signed off by both parties.

**Consequence.** Because the review happened before volume was generated, the correction touched a few cases, not an entire release.

**What it teaches.** A pilot works best when a domain reviewer checks reference behaviour early, the hold-out rule is written down, and the scope sheet is a living document with named states.

Early review of reference behaviour, a written hold-out rule and a scope sheet with named states keep corrections small and the pilot honest.

Audience:
Model labs
Content type:
Scenario
Buyer question:
Scope and getting started

Content release: r-a220eb6e099b243b

More on Loan booking