Every story sits in one of four schools, inside one subject, and declares one level. Start with the school that matches the question you are asking, then choose how much reasoning you want to do.
4 schools14 subjects13 published storieswhat is mapped but not yet written →
Four schools divide the curriculum by the question they answer. Each holds three or four subjects.
Levels describe the reasoning a story asks for, not how obscure its vocabulary is.
Subjects sit inside a school. Content areas are the exact capabilities inside a subject.
Introduces one principal mechanism and asks the reader to recognize or directly apply it.
5 storiesThe regression suite recorded sixty real rider-days from production. Two years later it approved the wrong fare rule, then reported the correct fix as a regression.
The model passed every launch check. Then one policy sentence changed, and its confident answer stayed frozen in time.
Five cells were healthy and held every job. The shared layer in front of them could not find out which cell to ask.
PackProof inspected every package through six cameras. Its weekly training table still invented physical evidence for packages no technician had examined.
The schema allowed only real, open warehouse zones. One of those valid choices still could not hold the pallet it was assigned.
Requires tracing interacting mechanisms and comparing material production tradeoffs.
4 storiesThe backfill matched every offset, key and timestamp. One value it read from outside the log had changed its mind about a day that had already been settled.
A fraud-review slowdown became a fleet-wide retry amplifier—and the AI orchestration layer multiplied the load before anyone saw the feedback loop.
The candidate ran safely in shadow mode. The incumbent still decided which field evidence could exist—and therefore which model looked better.
The router chose the correct reasoning model. Adaptive inference then spent the verification window the critical request needed.
Requires defending a cross-boundary production decision under ambiguity, changing evidence, or organizational constraints.
4 storiesFuture traffic returned to the old model release. One approved deal had already crossed the boundary in an old summary, then propagated through booking, collateral, notices, evaluation, and learning data.
Barge-in stopped the audio, the model, and the turn, and asked the tool to stop. The payment had already been accepted, and the assistant said it had been held.
The model proposed the right recovery. The human approved it. Then one lost cloud response split what the workflow knew from what production had already done.
Every service enforced its own narrow permission. Together, those permissions released one employee’s exact records to another person.