Scorecard

Good: Acceptance tied to business outcomes; difficult slices and critical errors visible

Weak: Leaderboard or cherry-picked demo; failures omitted

Evidence: Paired results, denominators, errors and slice reports

Good: Buyer-controlled holdout, justified split, calibrated graders and reproducibility

Weak: Training/test overlap; score without rubric; supplier controls all answers

Evidence: Split manifest, access policy, run config and adjudication record

Good: Rights, labels, lineage, base choice and bounded experiments connect to failures

Weak: More data or more training proposed without diagnosis

Evidence: Inventory, label guide, provenance and ablation plan

Good: Common accepted-task denominator; one-off/recurring costs and downside cases

Weak: Token saving only; missing labour, retries or minimum commitments

Evidence: Itemised quote and scenario workbook

Good: Actual serving config, load tests, fallback, rollback and observability

Weak: Notebook handover with deployment left implicit

Evidence: Architecture, traces, operational runbook

Good: Appropriate handling, access, retention, incident and change controls

Weak: Generic badges with no product/contract scope

Evidence: Data-flow review and contractual controls

Good: Artefact schedule and feasible export/redeployment path

Weak: “You own it” without transferable files or permissions

Evidence: Licence list, export package and rehearsal plan/result

Good: Named team, dependencies, milestones and change control

Weak: Unnamed resources and vague acceptance

Evidence: Work plan, responsibility matrix and reference questions

Good: Owners, cost, regression suite and triggers to retain/replace

Weak: Automatic retraining retainer without diagnosis

Evidence: Maintenance schedule and baseline refresh policy