Build for understanding

Start with the learner's next move

Keating's teaching contract begins with diagnosis when learner state is unknown. The learn prompt asks one short diagnostic question, waits for an answer, builds intuition before formal structure, repairs a likely misconception, and then asks for retrieval or reconstruction. It ends with transfer or a practical consequence.

The design goal is independent understanding. A learner agreeing with a fluent explanation is not the same thing as reconstructing the idea, applying it, or remembering it later.

Make the plan adaptable

The local lesson builder composes orientation, intuition, formal core, misconception repair, worked examples, guided practice, and transfer. Policy settings influence the structure, including diagram use, question emphasis, and practice count. Domain-specific additions connect the structure to the subject: code includes live execution, while law and history call for relevant citations or sources.

Unknown topics can use a fallback topic definition inferred from domain keywords. This keeps artifact generation available, but a plausible scaffold is not evidence that its examples or domain classification are appropriate.

Use memory to choose the next question

The teaching instructions consult prior learner coverage before repeating orientation. Existing coverage should lead toward retrieval or misconception repair. Goals, learner memory, feedback, review signals, and artifacts have explicit storage locations so they can inform later work and remain inspectable.

Keep the distinction between recorded evidence and inference. A saved quiz answer is an observed submission. A conversation signal or feedback rating can guide teaching, but cannot establish durable mastery by itself.

Render an interaction; save an artifact when useful

The adaptive teaching skill favors inline OpenUI for learner-facing questions, study plans, concept maps, and other teaching surfaces. Questions should give the learner a meaningful action and wait for its completion. Persisted plan and map commands are for cases where a filesystem artifact is useful, rather than a prerequisite for rendering every card.

Verification is a separate responsibility. The skill asks the teacher to consult or generate a verification checklist before factual instruction, and to keep unconfirmed claims qualified. A generated checklist must still be worked through.

Keep improvement evidence independent

teaching-bench runs fresh training episodes with the active tutor revision. auto-improve proposes one skill and compares the candidate against the incumbent on paired validation and a sealed release holdout. Both gates must pass before the exact revision can activate. A failed or rejected proposal stays inspectable.

The fixed gate requires independent case families, no family regression, no failed critical criterion, and the documented improvement and statistical thresholds. The holdout is consumed before execution; retries, renamed cases, or the force flag cannot renew it. The force flag only bypasses cooldown.

Measure people separately from models

The learning-check workflow records fixed prechecks, immediate postchecks, delayed recall, and transfer for the currently supported assessment topics. Missing stages stay unknown; they are not converted to zero. Assistance status is recorded, and assisted or unknown-assistance submissions are excluded from unaided outcome summaries.

Synthetic tutor episodes and model judgments test bounded teaching behavior. Historical benchmarks describe recorded evidence. Neither establishes a causal improvement in human learning. Even observed pre/post differences need careful interpretation because the fixed banks, self-reported assistance, and lack of randomized groups limit conclusions.

keating teaching-bench
keating learning-check start fractions --learner learner-1
keating learning-check list
keating learning-check show <check-id>