Say why it moved

The coding agent knows the repo. The person opening the page does not have its context and reads for two reasons: to learn what a paper or an idea does when tried, and to find a behavior and fix it fast at work. Both want one look: what changed, did it move, why, what it taught, and can I run it again. On one day in September 2026 real agents posted dapo-lr5e-05-s17-30st nineteen times, no note on any run, no chart, a 0.75 on a page that counts points, and no question on six agents of seven. Every page was correct and said nothing. The fix is a playbook the agent follows before the second run, tested in CI like every skill: skills/manage-experiments. wai init installs it under .claude/skills/ and the AGENTS.md block names it.

The person asks	The agent posts	Call
what are these runs for	the question, at the top of the page	`tracked.experiment(question=, hypothesis=, method=, measure=, decide=)`
what changed	one line: the prompt line, the data mix, the setting	`run.note("Changed: ...")`
did it move	before and after in points, with the interval	`"Moved: 75 to 96.2 points (±5.2) on 52 asks"`
why	the asks that flipped, and whether the reward and the held-out moved together	`"Why: 11 asks now pass, every one an old order; 0 broke"`
what did it teach me	the hypothesis closed in one sentence, with the paper when the run tried one	`"Learned: 800 traces taught one rule; the held-out checkpoints are the proof (rlhfbook.com, Over-Optimization)"`
can I run it again	the command, the seed, the pins	`"Reproduce: uv run python train.py --seed 17"` + `RunRecord(data=, optimizer=, provenance=)`
show me	one chart per run: bars with intervals for a harness, reward and held-out on one chart for training	`tracked.figure(name, {"data": [...], "layout": {...}}, run=run)`
what failed	up to 20 rows under each score, failures first	`run.score(b, 96.2, ci=5.2, n=52, examples=worst)`
is the page ready	the account read the way a teammate will, one line per problem with the fixing call	`readback(tracked)` from the skill; empty means clean

The page counts points out of 100 and the client rescales nothing: a rate out of 1 goes up as run.score(b, 0.71, ci=0.04, n=52, fraction=True), which posts 71 (±4), while a bare 0.71 posts as 0.71 points with a one-time warning that names the fix. A reward that climbs while the held-out line stays flat is the judge being gamed (rlhfbook.com, “Over-Optimization”); the note says so rather than letting the person find out. print(tracked.brief()) and print(tracked.verdict()) are the sentences the page shows, from the same rows; paste them in the pull request. Names follow Naming: the agent after the product, behaviors as the policy phrases them, versions as the team ships them, the test by its content.