Guided Lab

Project: an auditable suggestion evaluator

Download Orientation Toolkit. The proposals.json examples are original authored policies, not captured AI responses. This offline path requires no model, API key or subscription.

1. Read the task and policy contracts in the README. Calculate expected outputs for ordinary, empty, threshold and duplicate cases before opening the reference implementation.
2. Implement parse_policy, predict and evaluate in ai_starter.py. Interpret the three allowed fields; never use eval, shell execution or generated Python code to interpret proposals.
3. Run python3 check.py ai starter and python3 cli.py ai starter proposals.json. Explain every failure reported, including the difference between a malformed policy and a valid-but-wrong policy.
4. Add a case not used to choose the candidate. Preserve its expected result before evaluation. Record the exact fixture set; do not generalize its score to overall model accuracy.
5. Complete one tutor/review cycle using authored examples or an assistant you already have access to. With a real tool, supply only non-sensitive minimal context, record tool/version/date and keep its actual response. Otherwise label the cycle simulated. No live account is necessary to pass this orientation gate.
6. Analyze the lesson's injected-document example and state which action is unauthorized. The schema rejection check is not a test of real model injection resistance.
7. Close the help source and complete the independent transfer. Reuse the existing AI-use log for assistance and privacy decisions.

Deliver evaluator code, expectations, actual report, counterexample, assistance record and unaided explanation. Full agent frameworks, model training and live API integration are deferred. Passing this lab does not establish the effectiveness of a particular AI product.