Change only the context
Made for the earlier edition of this course, which ran model-written code through a course launcher. The current course does not require one.
9:27 · subtitles: English, 中文 ·
download
An agent acts only on what reaches it, so the context is an intervention you can test. Trace the request path first; then compare contexts on the same task by changing the context only, with the question, the model, a fresh call, one call each and the checker fixed. Predict before running, keep predictions apart from what was observed, and claim no more than one pair shows. The same holds in the middle of a session: build the next request four ways from one long history, run the same check on each, and check a summary against the transcript as well as by the next check.
Not passed is not failed
Made for the earlier edition of this course, which ran model-written code through a course launcher. The current course does not require one.
8:51 · subtitles: English, 中文 ·
download
A requirement has to be traced from an observed event, not invented in the request. Ask about a past event; keep observation, inference and what you cannot conclude apart; write the problem statement by hand and check each clause against the observation; write an acceptance criterion that labels its design choices; then judge a generated candidate against that criterion. The trace runs S4-A → P1 → AC1 → AC1-mixed → predicate repair, and a passing test cannot turn a design choice into user evidence. The same criterion is then handed to an agent in stages (research, propose, plan, implement, test), with a person's decision at every stop: a candidate is read against its bound before it is tested, a failed check means the later checks were not run, and a round that stops after Plan, as the course's worked round on ci-dashboard does, is a legitimate result.