Measure field match and abstention quality

A plausible result is not yet evidence. Evaluate with field exact-match on clear fixtures, high abstention precision on blurred fixtures, schema validity 100%. The test fixture should contain an easy positive case, an easy negative or baseline case, and the boundary case most likely to flip. Separate assertions about software contracts from claims about model quality.

Fixtures must include clear, rotated, blurred, and adversarial images. Measure field exact match, abstention precision/recall, schema validity, latency, and cost. Slice by condition.

The artifact's user-facing goal is specific: return a typed answer with confidence and evidence location, or abstain when pixels are unreadable. Its accepted input is authenticated upload bytes plus a question string, after MIME and size validation. Those statements are intentionally narrower than “build an AI system.” Narrow scope lets us inspect every input and expected result, and it prevents a toy result from being presented as a production claim. System shape for this chapter: an upload boundary checks authentication, MIME signature, dimensions, and byte limits; an image worker strips metadata and resizes safely; a provider adapter sends text and image parts; a schema validator rejects malformed or overconfident output. Keep model calls behind adapters, keep authorization and validation in deterministic code, and carry stable IDs and versions through every response. That separation lets you decide whether a bad result came from input handling, retrieval, inference, validation, or deployment. This page's job is the evaluation step: a plausible result is not yet evidence. Setup baseline for the chapter (run once per machine, not secrets in git):

python -m venv .venv && source .venv/bin/activate
pip install fastapi uvicorn pillow pydantic httpx
export MAX_IMAGE_BYTES=5000000

If hardware or a hosted provider differs, preserve the interface and expected behavior. Do not present provider syntax as universal—when a vendor adapter is unavoidable, keep it behind a thin boundary and test with a fake first. The deliverable is not “it ran once”; it is a reproducible artifact another developer can inspect, including expected output and one deliberate failure related to overconfident misread of a blurred total. Operationally, write down the owner of this stage, the command you ran, the observed output, and the next page's dependency on that output. If you cannot point to a file, fixture, metric, or config key, the stage is not done. Prefer small, reviewable increments: one contract, one path, one metric, one failure, one gate. When tradeoffs appear—latency versus quality, hit rate versus false hits, local privacy versus cloud quality—record both numbers instead of moving the threshold until the report looks green. The chapter ships only when evidence for field exact-match on clear fixtures, high abstention precision on blurred fixtures, schema validity 100% and a rehearsed recovery path exist beside FastAPI /analyze endpoint with normalization, schema validation, and model pin.

Save this as lesson.py and run python3 lesson.py. Prefer the standard library or the pinned packages from the setup block so the example stays reproducible.

fixtures=[{"name":"clear","expect":"$42.10"},{"name":"blur","expect":"abstain"}]
pred={"clear":"$42.10","blur":"abstain"}
print(sum(pred[f["name"]]==f["expect"] for f in fixtures)/len(fixtures))

Expected output: 1.0 fixture accuracy on the tiny set. Exact floating-point formatting may vary slightly, but the asserted behavior must not. Read the output as evidence about this stage, not merely proof that the interpreter started.

Print per-case expected versus observed values. Aggregates that look green can still hide a severe slice failure. At the evaluation stage, save the smallest failing fixture beside the expected result. Change one cause at a time and rerun the exact command printed above; that makes the repair reviewable and keeps this chapter's progressive artifact reproducible.

Require the stated thresholds for field exact-match on clear fixtures, high abstention precision on blurred fixtures, schema validity 100%. Do not delete hard cases to green the report. For this evaluation page, preserve the fixture and result as evidence for the next page. Label observations separately from conclusions: a passing assertion establishes the behavior it names, while broader usefulness requires the chapter's full evaluation set and stated operating limits. Primary metrics for the chapter remain field exact-match on clear fixtures, high abstention precision on blurred fixtures, schema validity 100%.