- A client watched the search narrow over eight rounds down to six guitarists in New York, pressed the button that says let Mira take it from here, and got a different shortlist with a seller in Nigeria on it. Nothing anywhere reported a fault. One live run, five separate silent failures, all of them found by reading it line by line. The button did not continue anything: it opened a brand new search from cold, so every round the client had just watched converge was thrown away at the exact moment they committed to it. It now resumes from what the rail already decided, and every step of that resume fails soft back to the old behaviour rather than to an error. The judge steering the search had no way to say nothing needs to change, because the answer form demanded all sixteen filter fields every single round and leaving one out was read as deleting it, so unchanged and cleared were the same thing to write and it wrote whichever was shorter. Finishing was read as a statement about what to search for rather than about whether to keep searching, so this run finished with an empty form and deleted its own searches, the role, the category and the client's stated country in one answer. With the filters gone, the fallback took the first five words of the brief and ran a real marketplace search for the literal string “About the client: Dor is”, graded the twenty-one people that came back, shipped them, and put the same string on the client's own search-Fiverr-yourself link; an honest empty list beats a confident wrong one, so that fallback now uses the role, or the previous round's searches, or nothing at all. And the hard requirement, already based in New York, never reached the search in the first place, because one side of the system holds a country's full name while every reader of it compares two-letter codes, the same units mismatch that emptied a Mexico brief of all nineteen finalists a week ago. All five are fixed, and each one is now counted rather than merely fixed, so the same silence cannot come back unnoticed.
- The live rail keeps the people it showed you, and fills the last places instead of starting over. The rail beside the conversation refreshes as the brief changes, and pressing the button used to discard its result and run the full ninety-second hunt again to answer a question that was already answered, differently. Now the rail holds the shortlist and the button keeps it, all eighteen places. Two things had to be true first. The rail got a second engine that invents nothing: one marketplace search against what the client actually committed to, plus what Fiverr's own catalogue resolved from their brief, with no model anywhere in it, so the same brief returns the same people every time and nothing on the screen was made up. And a strict search that comes back with three cards is making a claim about the whole marketplace that our filter set has no right to make, so on the button the search is re-run with one stated constraint given back at a time, cheapest concession first, until the list fills. Every card recovered that way carries what was dropped to reach it and says so on its face, under a divider the page draws where the strict answers end. The client's own search link now also carries the category the hunt settled on, because a keyword alone lets the marketplace pick its own, and it picks a different one: the same three words landed on a category with four and a half thousand results where ours has under three thousand, sharing two of the top ten.
- A place the work physically happens in is part of the craft, not a preference to be traded away. An errand brief asked for somebody to buy items in a Swedish pharmacy and post them to the States. The search worked out on its own that this needed people in Sweden, held that for three rounds, and dropped it on the fourth, not for want of evidence: the word Sweden appears sixty-five times in what it was reading, including Mira's own line about needing a Stockholm-based shopper and four named Swedish chains. It dropped it because we told it to. The list of concessions it is allowed to make ends with give back one thing the client actually stated, and the round summary was arguing for exactly that. Thirteen Swedish sellers had been found and ten graded every round; one made the final list, behind couriers in Italy, Japan, Thailand and Turkey, each carrying based in sweden in its own record at no cost to its rank. Every stage now asks the same single question, so they cannot disagree: what would this person physically DO on day one, stand somewhere or open a laptop. Where the work happens somewhere, that somewhere is a capability like a tool or a language, and a talent outside it grades off rather than nearly right. It is also taken off the concession list entirely. When that leaves the pool thin, a short list is now the honest answer: three people who can actually do the errand beat eighteen who cannot, and the client finds that out on first contact.
- Every judging turn in every live hunt was coming back an error, and the test covering it was green. The judge's answer form is built in two different places, one for the opening turn and one for the hunt itself, and a field added the day before reached the required list in both and the properties list in only one, which the model provider rejects outright. So the hunt could neither narrow nor finish. It shipped green because the test read the first form while the hunt sends the second; the test now walks both and asserts the property that is actually enforced. Four more of the same family landed with it. A search that stops at seven people when the client is shown eighteen has not finished, it has stopped, so finishing early is refused while rounds remain, and the reason is handed back to the judge as its own next turn rather than as a silent override it cannot learn from. The service that classifies a brief into a marketplace category was being handed the budget line, the engagement block and the entire client conversation, which it rejects outright over three thousand characters, and because that enhancement is deliberately allowed to fail quietly it failed one hundred and twenty two times against twelve successes across an hour and a half of replay, with thirty-three of thirty-five hunts searching with no category vocabulary at all and nothing surfacing anywhere. A compliance line at the end of a brief was read as work to shop for, so an outreach role spent three of its nine searches looking for GDPR specialists and got privacy lawyers and a penetration tester, two of those three returning the same four people. And the judge now sees, per search, how many people that search was first to surface and how many of them graded well, because a total can never name the search at fault and it was re-issuing dead ones round after round.
- Signing in changed which A/B group you were in, and at the live setting of zero percent that meant every pinned tester dropped to the control group the moment they signed in. Every time. The group is worked out by hashing the account id, and a guest and the account they sign into are two different rows with two different ids, so signing in was re-rolling the dice. At a live dial the damage is roughly one in five signing-in guests changing group mid-session and losing or gaining a feature for no reason they could see. At zero, where the only way into the experiment is a tester pin and the pin is keyed to the guest, it was total. The claim that moves a guest's work onto their new account now carries the group across with it, written inside the same transaction as everything else, and only where it would actually differ. That shipped last night and opened the same flip in the other direction, which is this morning's first commit: sign out, browse anonymously for a minute, sign back in, and a throwaway guest session was overwriting the settled group of an account with a long history behind it. The carry now applies to a first-time account only, decided before the move rather than after, while the question can still be answered. Where both sides are established and disagree, the account wins: it is the durable identity, and the one every past project and every recorded step is filed under.
- The console learns to read one A/B group, and the page that answers that question could never fill itself. It served the readout is not ready yet, and could never stop, because the recurring tick that builds it existed on a laptop and had never been created in the cloud, so the only thing that would have written it never ran. Retrying was never going to help. The page now builds its own readout the first time somebody opens it, and the recurring refresh joins its four siblings. It also became an admin area of its own, the fifteenth, rather than borrowing a neighbour's: sharing was right for who may read it and wrong for how anyone finds it, so it had no sidebar entry and was reachable only by typing its address. And it was rendering as an unreadable strip about a third of a screen wide, clipped off the left edge, because two stylesheets had claimed the same class name and every page's styles are bundled into one file, so the name is global whichever file declares it and the later rule simply won. That is the second such collision in two days, and it is not something a careful look at the page would ever find, since a per-page preview loads only its own styles and looks perfect, so it gets a build gate instead. On top of that, the launch board, the agent health page and the people-and-projects explorer can all now be read for one group, filtered at the source rather than bolted onto each query, since a filter applied in thirteen of fourteen places is a wrong number under a filtered heading with no way to see which one was missed. The two figures that measure the whole platform rather than one group say so on the page, instead of being read as that group's own.
- Mira's outreach stops interrupting both sides, and the console stops printing a second opinion of its own ranking. Every question put to a client parks that talent's proposal until an answer comes back, and while they wait the other candidates are finishing theirs, so a talent can lose a place on the shortlist to a question that did not really need asking. The question is now put first to the person it delays, in fixed wording rather than the agent's own: waiting for a response may affect your chances of making their shortlist, so is this question essential right now, or can it wait until you have connected with the client. Only an essential answer goes through. It is measured rather than enforced, deliberately, because binning a question he has just promised to ask leaves a talent waiting on an answer that never comes, and this repo has already paid for that twice. Separately, a wrapped-up conversation is finally allowed to end: freelancers treat an inbox as something every message is owed a reply in, so a finished thread kept running on ok and thanks in both directions, and there is now a way to simply say nothing when nothing is outstanding and the last message asks for nothing. Any real question, request, new fact or client answer puts him straight back to replying. And in the console, the score printed beside each pick was a second formula over the same shortlist, disagreeing with the actual ordering on forty percent of one hundred and sixty seven sampled searches, one list showing the thirteenth pick outscoring the top one by eleven points. The column now renders the ordering itself, worked out at the moment it is read so every past search re-renders correctly, and sorted by it the scores cannot come out anything but non-increasing.
- And the day the client actually sees: a signed-in client who looked signed out, a card pinned over the chat, and a wait that finally shows the hunt. “i did sign in, but! i didnt really signed in” had three separate causes, all real. A profile row is created on any read at all, so an account that had merely opened the workspace once months ago owned an empty one, and the guard meant to avoid overwriting real work read that placeholder as real and left the whole About You the guest had just built behind on the old row. Fiverr's sign-in often sends neither an email nor a display name, so there was nothing to greet them by, and we already store what Fiverr itself calls them and had never used it. And the menu could not tell signed in but unnamed from never signed in, since both rendered a grey circle with a letter in it. A question card the buyer had already replied past stayed pinned over the chat. The brief's progress tongue painted straight down the middle of the why-Mira-picked card, undimmed, because a dialog was racing app chrome on document order instead of outranking it. An agent failing mid-turn rendered as a red line inside Mira's own thinking, with raw backend error text in it, which then outlived the turn in the browser. The guest waiting page got the treatment the concierge one got three days ago, since from outside they are one page. And the waiting dashboard now shows the hunt during the opening minutes, where it used to be a headline over a blank page while searches were running and being narrated, including whether Mira had to widen her filters to fill the list, which is the difference between these are the best people and these are who was left.
A Sunday of 37 commits, opening the week with two long branches landing at once, and the through-line is work already done being thrown away and asked again. A client narrowed a search over eight rounds to six guitarists in New York, pressed the button that hands it to Mira, and got a different list with a Nigerian seller on it: the button started a fresh hunt from cold, the judge had no way to write nothing changed, finishing deleted the filters, and the fallback then searched the marketplace for the first five words of the brief. The rail now keeps its eighteen and fills the last places by giving one stated constraint back at a time, on a second engine with no model in it. A place the work physically happens in stopped being a preference to trade away, after a Swedish pharmacy errand shipped couriers in four other countries. Every judging turn in every hunt was erroring under a green test. Signing in was re-rolling which A/B group you were in, which at the live setting of zero meant every pinned tester silently dropped to control. And the client-facing half of the day: a signed-in client who looked signed out, a card pinned over the chat, a progress tab painting through a dialog, and a wait that finally shows the hunt instead of a headline over a blank page.