Teach one job
Broad intelligence is not an observable learning goal. Start with one responsibility that a person already knows how to assess.
Posted beside every classroom door
The class rules.
Agent training should make uncertainty, evidence, responsibility, and human authority easier to see—not easier to hide.
HOUSE RULESPRIVATE
BY DEFAULTHUMANS APPROVE REAL ACTIONS
BY DEFAULTHUMANS APPROVE REAL ACTIONS
Show the evidence
Every important claim should lead back to an approved source, example, procedure, or correction.
Reward “I don't know”
A visible knowledge gap is safer and more useful than a confident invention.
Separate draft from action
Preparing a recommendation is not the same as approving, sending, ordering, changing, or operating.
Test unfamiliar cases
Repeating a memorized example is not proof of learning. Change the details and inspect the process.
Publish a sandbox
Public challenges use sanitized summaries and simulated tasks. Private documents and real tools stay outside.
A precise language matters
Teaching an agent is not always model fine-tuning.
In this course, “train” means constructing reliable context, procedures, corrections, memory, evaluations, and permissions around an AI system. Changing model weights is a different technical process and is not claimed by this MVP.
Privacy boundary
Enter the classroom →
Your workbook begins private.
Each lesson has a separate safe-to-share choice. Even then, only a moderator can publish the project. Email, session ownership, challenge fingerprints, and entries not marked safe never appear on public pages.