- The experiment platform got its console, and the first thing that console revealed is that seven of its dials were pictures of switches nobody could throw. A dial exists because a line of code ran when a program started, which quietly makes the list of them a function of what each program happened to load: the machine that RUNS the work loads it, the console that only reads ABOUT it does not. So the board drew a card, drew a slider, and the slider reached nothing, seven times over. One morning of development logs carried twenty-eight silent refusals and nothing anywhere else said a word. The board is now one page over six different kinds of change at once, a prompt section, a branch in the code, a setting, a product cohort, each one described in a vocabulary the panel can draw without needing to know what it is looking at, and each one listable, searchable, inspectable and stoppable from that one place.
- The one experiment whose entire job is to prove the measurement is honest had recorded nothing at all, for its whole life. Zero rows in every ledger, and not one arm across 139,883 recorded model calls. It was not a typo. Writing down who saw which version was opt-in, there were four different hand-wired ways to opt in, and the experiment with nobody to remember it was the one that mattered most. Recording is structural now: one place writes it, every path reaches that place, and the hand-wired call sites were deleted rather than documented, because a call you can forget had to stop existing rather than be explained better. That always-on control is now the canary for the measurement itself, and deciding which version somebody gets was pinned as a pure calculation that may not read a clock, a coin, a setting or the database, so last Tuesday's answer can still be reproduced next year.
- The console also stopped trying to decide what counts as success. It used to ask an operator to pick a primary measure from a fixed dropdown of what the platform happens to count today, which is a second, weaker copy of something the reporting side already does properly, and two places to name a measure is two places to disagree about it. The console now answers what is running, who is on it, and whether it is hurting anything, and stops there.
- Mira inside Claude got a front door with a lock on it again, and two different lanes behind it. Yesterday's experiment, a key written into the address and a self-minted session, was reverted whole: signing in is back, and now open to any verified Google account rather than only a Fiverr one, which had been excluding precisely the audience a client surface exists for. Behind that door there are now two lanes, deliberately not the same product. A guest lane is a complete product on its own: the conversation, the brief, live matches, a full search, the PDF. The Fiverr lane signs in through the same door the website uses, brings the account and the history the person already has, and is the only lane allowed to message real talent, because "the hire continues in your inbox" is not a sentence you can say to somebody who has no inbox.
- The employee console moved out of the house. Admin traffic and client traffic had been sharing everything: the same service, the same machines, the same pool of database connections, the same database user. Two days earlier four admin page loads on one production machine each held a connection for a full minute past their own timeout and took 40% of that machine's database capacity away, while every client request behind them queued for one. Employees now have their own service, their own pool, their own budget and their own database user with its own hard ceiling. On the same day the People list stopped working out "most recently active" on every request: warm that took a fifth of a second, cold it took twelve to sixteen, and production had timed out on it three times in two days, every one of them the first unfiltered load after a quiet spell.
- A request that gives up now actually stops working. The old timeout cancelled the WAITING, not the work. A client saw the failure at ten seconds and the job behind it ran to completion sixty-one seconds later, still holding its database connection and still burning processor on the main database, while the client who saw the failure was free to press again. A timeout was the beginning of the cost rather than the end of it.
- A hunt the client walked away from stops holding talents after forty-eight hours, and the talents stopped being told there is a clock at all. In the version of the waiting experience with no deadline the client ends the hunt themselves, so the case a talent minds most, that they answered, they priced the work, and then nobody ever came back, tripped nothing whatsoever: the proposal stayed open, no one told them, and no sweep would ever have found that hunt. Forty-eight hours after the last thing the client did, every open conversation is now closed out with the note matching what that particular talent actually did. The window slides, so a client who comes back at hour forty buys another full window from that moment. And the paragraph telling each talent how long the client was collecting for was removed whole rather than switched off: our side neither knows the hunt has a deadline nor mentions one.
- The headhunting screen, rebuilt around what the tester round actually said. "Experts" reads "talents", the final step is "Matches", and every band of faces is re-centred under the summary row's five equal columns, which the prototype's layout had been quietly pulling half a column out of true. Nothing appears at zero any more: no capsule until the first proposal is in hand, no not-a-fit row until the first drop. The counts under each step now say how many people are standing there right now rather than how many have ever passed through, so somebody moving from shortlist to contacted lowers one number and raises the next, and the summary row can no longer disagree with the faces drawn beneath it. The green finish button only turns green when the hunt is over AND somebody is actually in hand.
- The rest is a long tail of things that were quietly wrong, and several had been wrong for months. The live strip of matches now opens itself the first time a hunt finishes, because the single moment its folded default costs the most is the moment real, messageable people first come back. A talent who speaks three languages has all three counted rather than silently clipped by a two-line limit, so a card stopped reading as a dangling comma with a language missing. Two real sends to talent were withheld because a brief rendered at 8.61 megabytes against an eight megabyte ceiling that had stopped meaning anything in July, when attachments moved off the request that carries them. Mira stopped telling a client "the human specialist has your brief and will take it from here" when nobody had been handed anything, the desk's own internal bookkeeping having reached her as a nameless system line with no rule saying what it meant. Every new Fiverr button now names itself in the reporting from the day it ships, instead of being folded into "other" by a list kept in a repository those buttons are not released with. And the production infrastructure plan comes back clean for the first time in weeks: five of its proposed changes were permanent noise, and applying any of them rolled four new machines serving no traffic purely to strip a label the next release puts straight back.
A Thursday of 66 commits, the largest day this site has recorded, and nearly all of it is one idea: a switch is only real where somebody can see it and move it, and a version somebody was served is only real if it was written down. Seven dials drew sliders that reached nothing; the one control experiment that exists to prove the measurement works had never written a single row in its life; a page that times out now actually stops working; a hunt a client abandoned stops holding talents; and the employees, the clients and the two ways into Mira each got their own door rather than sharing one.