Grades

Write the rules your own build will follow: data you will not send, human checkpoints, failure disclosure.

Capabilities, limits, cost, and risk. Pass at 80%.

Versioned prompts, a schema, a passing test suite, and a README someone else could use.

Schema design, injection, and evaluating a prompt change. Pass at 80%.

Packaged, installable, tested, and documented.

Secrets, validation, and async failure modes. Pass at 80%.

With a cost dashboard and a written failure policy.

Backoff, idempotency, and cost control. Pass at 80%.

Completes a real end-to-end task in your own domain.

Tool schema design, control-flow choice, and state. Pass at 80%.

Over documents you actually own.

Chunking, embeddings, citation, and refusal. Pass at 80%.

Baseline, three interventions, and an honest verdict on each.

recall@k, reranking trade-offs, and query rewriting. Pass at 80%.

With a documented safety envelope.

Stop conditions, supervision, and agent failure modes. Pass at 80%.

With a baseline and at least one real regression caught.

Metric choice, judge bias, and regression design. Pass at 80%.

Live URL, public repository, evaluation numbers, three-minute video demo, and a written design defence.