Application Evaluation Criteria

Score any foundation-model application against four buckets — domain capability, generation quality, instruction-following, and cost/latency — rather than a single vague "quality" number.

An application that ships without an evaluation contract is worse than one never deployed: it costs to maintain and costs more to tear down, and many production AI features have unclear ROI precisely because nobody can tell whether they help. Structure that contract up front under eval-driven development as four independent criteria buckets, illustrated by a legal-contract summarizer:

  1. Domain-specific capability — does the model understand the subject matter at all (legal contracts, code, medicine)? Constrained by architecture, size, and training data; evaluated via domain benchmarks. See domain-capability evaluation.
  2. Generation capability — is the output coherent, faithful, and safe enough for the task? Classic NLG axes (fluency, coherence) matter less for strong models; factual consistency and safety dominate. See generation-capability evaluation.
  3. Instruction-following capability — does the model obey your instructions (format, length, vocabulary, banned phrases), independent of whether those instructions are well written? Essential for structured outputs and product constraints. See instruction-following evaluation.
  4. Cost and latency — is the quality worth the wait and the bill? Treat non-negotiable thresholds as hard filters before ranking survivors. See LLM model selection criteria.

Keep the buckets separate in the suite: a model can understand sentiment labels (domain) yet emit HAPPY/ANGRY instead of the required NEGATIVE/POSITIVE/NEUTRAL set (instruction-following), or produce a fluent summary (generation) that invents clauses absent from the source (factual-consistency evaluation). One aggregate score hides which failure mode to fix. Wire each bucket into offline prompt evaluation and agent grader types so harness and prompt changes regress against the criteria that actually matter for the application — not only against the easiest metric to automate.

Sources

  • AI Engineering (Chip Huyen), ch. 4