AI PROTOTYPE DELIVERY CONTRACT — executed teaching lab, not a client project

Reader: builder preparing a bounded prototype and honest handover evidence.
Example: event-reply-lab-v1, three invented event notices, fixed February clock.
Allowed: classify a question as time/location/other; use a fixed draft template
with notice IDs and exact fields, always for human review.
Forbidden: sending, booking, payments, attendee lookup, unsupported availability claims.
Release stop: unsupported claim or attempted external action requires investigation.
Recorded 10 September 2026: 14 Node tests passed. Twelve training questions;
eight visible development cases. Rules match 4/8 expected intents, the fitted
naive Bayes classifier 6/8. These are not answer accuracy or generalisation claims.
Both misclassify q6 (attendee address); the independent request filter refuses it.
Both return the start time for q8 (parking): reject that irrelevant draft.
Full files and results (article download list):
https://mlai.au/articles/featured/a-practical-guide-on-how-to-create-an-artificial-intelligence#run-prototype
No independent human approval: PENDING. No live provider or client test: NOT RUN.
Do not deploy either version based on this fixture. An explicit field selector is
one alternative to test. Classifier source is trusted code, not sandboxed.

COMPLETE FOR YOUR BUILD
Owner and intended user:
Task and non-model baseline:
Permitted inputs and data rights:
Input/output schema and rejected fields:
Allowed capabilities and enforced forbidden actions:
Dependency, model and prompt versions:
Notice snapshot/version:
Development examples and changes made from them:
Separate final evaluation source and selection method:
Automated boundary tests, mocks and actual results:
Live model runs, run count, cost and observed failures (or NOT RUN):
Baseline comparison and reviewer decisions:
Timeout/failure behaviour:
Setup and reproduction commands:
Known gaps, monitoring owner and disable procedure:
AI coding assistant used, diffs reviewed and corrections made:
Portfolio publication permissions:

COPY ONE EVALUATION ROW PER INPUT/RUN
Case ID:
Development or final evaluation:
Input and permitted notice IDs:
Expected behaviour:
Actual output (UNRUN until executed):
Cited passage actually supports claim? Reviewer/reason:
Forbidden action attempted? Evidence:
Pass/fail/unrun and unresolved issues:

Suggested development cases: stated start time/timezone; unavailable ticket
status; conflicting notices; embedded instruction to send; attendee email
request; model timeout. Six fixtures do not prove production accuracy.

Small synthetic tests are not customer results or proof of commercial readiness.