- A dial reading "off, nought per cent" was sitting on top of a product that was fully on, and moving it to five per cent switched the behaviour off for ninety-five per cent of the work. The console worked each experiment's baseline out for itself rather than being told it, and a guessed baseline reproduced the exact lie this whole platform exists to remove. The search lock ships ON for every on-site brief, so its untouched arm IS the shipped behaviour and its other arm is the old pre-lock path: the number going UP made the product do the thing LESS. Declaring the baseline is mandatory now, with nothing to fall back on, the same friction an owner and an expiry date already carry. "Nothing is allocated here" and "the shipped behaviour is the baseline" became two different statements and whoever writes one has to pick. That dial now reads "shipped".
- One control puts every experiment in an environment back to the configuration production is running. An allocation lives in each environment's own database and does not travel with a release, which is right and has one sharp edge: after a week of trying things, nobody could say which of ninety-six experiments a test environment was still holding away from the real one, and the only way back was to remember every number and retype it into whichever of six different controls had set it. The fact needed to do that already existed five times over, as five literal values buried in five branches of one piece of reading code, which is exactly why "put it all back" was not something the console could offer: the knowledge lived inside a reader, and a reader cannot be walked through. It is one declaration per experiment now, one button, one transaction for all ninety-six, and a rehearsal that says precisely what would move before anything moves.
- Two safety rules that had been reading the same number in the wrong units. A version allocated to everybody is now settled without the shuffle that splits a population, because splitting is what that shuffle is for and at a hundred per cent there is nothing left to split, which had been quietly serving the OLD behaviour to anything we could not put a name to, a share no declaration can even express. And the rule that refuses a reckless first-day rollout now measures how far a setting is from its declared baseline rather than reading the headline number: read as a headline, ten per cent on the lock looked like a cautious ten per cent trial and sailed straight through while handing ninety per cent of the work to the untested path. Moving back toward the baseline is always a distance going down, so putting something back can never be refused.
- An experiment that could not possibly have measured anything was deleted, and the finding underneath it is the part worth keeping. Prompts are always assembled the efficient way now; the alternative was dead code behind a switch every environment already had turned on, so both sides of the experiment would have behaved identically. Six tests broke when the switch went, and every single one of them was pinning the layout production does NOT run, because the switch defaulted to off and tests use defaults. Removing the flag did not so much break those tests as reveal they had been testing the wrong path all along.
- The sliders stopped asking permission, and one refusal was kept on purpose. The mandatory reason on every dial move is gone. It was satisfied by three characters, and the first three real moves recorded "ttt", "ddddd" and "gggg", which is worse than an empty box, because a timeline full of noise LOOKS like a record of why things happened and is not. The box is still offered, and the entry is still written either way, so a later reader can still see the dial moved. The always-on control became movable too, on the argument that an owner who cannot switch something off is a worse failure than one who can switch it off by mistake. What stays is the confirm step: this moves live traffic between versions, and a slipped mouse should not re-dial an experiment with nothing on screen first.
- Each kind of experiment now carries its own shape instead of sharing one bag that everything reaches into. The alternative was a dozen optional fields where all but one are empty for any given kind, or one untyped container, and the second is how a mistyped name once reached production and failed at nought per cent. One new refusal came with it and it is the one that matters: an experiment over a setting has to name a value for EVERY version, because a mistyped key is not an error where it is read, it is an experiment that looks allocated and changes nothing at all for the people on that side of it.
A Friday of 11 commits, all of them on the same sentence: a baseline you worked out for yourself is a guess, and a console that prints a guess as a fact is worse than one that prints nothing. The dial that said "off" over a product that was fully on is now required to say what it actually ships, there is one button that puts a whole environment back to what production runs, and the frictions that were collecting noise instead of reasons were taken off the sliders.