- Two rules we taught the search in the morning were taken back out by mid-afternoon, and both left their numbers behind. The first said that the language a piece of finished work is IN is part of the craft rather than a preference about the person: asked for two hours of narration in American English, the search had returned eighteen people all graded strong, led by a Hebrew voiceover at rank one, a Spanish one at two and a German one at ten, on a brief whose very first sentence said American. That is a real failure and it reproduces on demand. But replayed across thirty-six briefs four separate times, the rule moved almost nothing: language match in the top three was flat, off-language picks came out nominally worse, and twenty-eight of forty-one measurements landed inside the run-to-run noise. The second said that a price far below the stated floor on a brief that names an AMOUNT is evidence about scope, since a ten dollar voice tier is a one-minute clip and not two hours. It bought exactly the internal correlation it was designed to buy, and the client never saw a penny of it: picks below the client's floor stayed at fifty-six percent before and after, at the same median of a third of the floor, while the judge held constraints more often and returned thinner lists, one brief dropping from eighteen picks to five. So both came out, with the evidence written into the revert rather than into somebody's memory, and with the real defect named where it actually lives: about a quarter of priced picks show a cheap tier from a seller who also publishes an in-band one, and the check meant to catch that dropped precisely zero sellers across all thirty-seven briefs. That is package selection, not a prompt.
- The standing preferences a talent could set, the whole feature, was switched off behind a single constant one day after it shipped, and the reason is that its two halves cannot ship apart. Not deleted. Every line of it stays exactly where it is, the tests prove both states rather than whichever way the constant currently points, and the twelve live scenarios are skipped with a reason naming the switch rather than removed, because they are the only thing pinning when the subject gets raised and when it does not. Off means Mira has never heard of it: the tools disappear from the list she is handed, the policy block and the current-preferences heading come out of her prompt, the notes about a pending or already-refused ask stop being appended, and three reads per talent turn stop happening. The search stops loading floors, rules and availability. The point of gating both ends together is what happens if you gate only one: with the talent-facing half mute and enforcement still live, somebody whose boundary is already approved is dropped from every client search with no way to see it, change it or lift it, and no way to even raise the subject, because the one who would have discussed it with them no longer knows the feature exists. Invisible filtering with no recourse is worse than either end alone. Turning it back on is one line in one file.
- Hours before that, the last two pieces of it landed: when somebody can start became a DATE, and the closing message became the one our writer actually wrote. Not before September is now stored as a date and screened against the client's own delivery deadline, the same question the marketplace vacation window is already screened against, asked of a different source. Somebody free BEFORE the deadline is kept; that asymmetry is the whole rule. It is deliberately not written as a rule sentence, because rule text goes to the grader as evidence about the KIND of work a person will not take, under an instruction that ends by saying to treat anything unclear as a conflict, so a sentence about timing in there does not go quietly inert, it plausibly grades somebody off every brief there is, and the drop is indistinguishable from a capability mismatch from every angle. Capture refuses what it cannot trust: a date already in the past screens nobody, a vague answer is a reason to ask which date exactly, and a date more than two years out is a mis-resolved relative date far more often than a plan. On the copy, the closing line now says the memory update is not instant and shows the same snapshot of stored preferences they would get if they simply asked what is set, so the two can never describe one stored row differently, and only the facet that MOVED is bolded, since the point of listing the others beside it is that the changed one stands out. Mira restates the constraint she heard, all of them when several arrive at once, and asks once whether there is anything else; she still never restates the resulting state, because the appended block is the single statement of what will be saved and a second version in her own prose is how somebody approves one thing and reads another.
- A client asked us to delete their account and we failed nineteen times in an hour, then gave up, leaving their content sitting in the row we had promised to empty. The erasure blanks a great many columns to a constant empty string, and one of those columns sits under a uniqueness rule. The moment an account held two still-open suggestions under one source project, blanking both made them identical, the database refused the second, and because the whole erasure is one transaction that refusal took the entire deletion down rather than one table. Nineteen starts, nineteen failures, no completions, then the retries ran out and the job was abandoned. The row is archived and still holds the client's writing, which is a retention of personal data and not a delete. That column now gets a tombstone derived from the row's own identifier instead of a shared blank, the same trick two other columns already use against the same class of rule, with the same guard that keeps a second run harmless. No constraint was relaxed to make it work. Test first, which fails on the old code, and the account that was abandoned still has to be re-run by hand.
- The console's main tab had stopped loading altogether, and the fix moved the filtering to where the data is. It selected every project in the system and then ran fourteen correlated lookups against each row with no index behind its ordering: five point seven seconds inside the database, nine point four end to end, against a ten second ceiling. It tipped over twenty-four times in a fortnight, and the timeout reaches the browser as the error QA reported. The people list was the same shape, thirty-seven slow requests and three hard failures in three days. Both are paged now, two hundred at a time, where a cheap pass picks the page's rows first and only those get the expensive per-row work. Paging drags three things across with it and each is a correctness fix rather than a performance one. Filters have to narrow on the server across the whole table, because a browser-side filter over a paged list would answer nothing matches from the first two hundred rows while the match sat on page nine. Sorting is split in half and the table says out loud which half it is in, since ordering by a column the browser derives would have to compute it for every row before anything could be cut. And paging is one control, draining what is loaded before fetching more, so it reads as infinite scroll and is strictly paged underneath. Three further defects fell out of the same work: a guard keeping deleted accounts out of the people list had drifted into an interpolated list where the scanner that enforces it could no longer see it, a page whose last row had no value for the column being sorted on silently ended the list and stranded everything after it, and the new preference queue was returning a server error on every single load because it handed one kind of identifier to a field expecting another.
- A four thousand dollar store redesign was ruled small and sent to plain search, and the seam that exists to stop that only ever looked at half the question. The router said low at ninety-six percent confidence while its own reading of the budget correctly said the money was well above the floor: the low came from the timeline, twelve days it judged unworkable for the scope. The guard behind it only ever vetoed a low that CLAIMED the budget sat under the floor, so this one walked straight through and a real buyer was hard-routed away from the specialist desk. The floor is the only thing that rules a client out of that desk, so routing is now an either-or: a high verdict sends them there, and the budget clearing the floor sends them there on its own, no verdict required. A low stands only when the budget is confidently under. The prompt is deliberately untouched, so the full judgement still appears in the trace and on the dashboard with its own reasoning kept beside the decision, and how often the model and the rule disagree is now a number somebody can watch.
- Five bugs off a manager's iPhone screenshots, and an audit behind them that found the layer they all came from. The keyboard rising covered the whole conversation, because the phone ignores the modern request to resize the page and instead pans it, scrolling the header off and leaving an anchor spacer where the thread should be; the app now lays itself out inside the visible part of the screen and remembers it was at the tail across the shrink. Four inputs sat a pixel or two under the size at which the phone stops zooming on focus, and the zoom persists after you look away. A checklist inside a drag handle was finger-dead with no way to tell, and now scrolls as a sibling with a measured fade at its edge. A brief still being written rendered its frame over a zero-height nothing, a white void with an orphaned close button floating in it, and now says Mira is writing. The waiting page sized itself against a header height it did not actually have, so the column overflowed its own scroll area and armed pull-to-refresh. A chip ran off the right edge because its row let each slot grow to the width of unwrapped text, so its own limit bound against a box that had already overflowed. Underneath all of that: automatic text direction for Hebrew readers on the composer, both message bodies and the chips; a counted scroll lock on the six overlays the dialog library does not cover; hover-only controls made visible where there is no hover; and real hit areas where there were none.
- The board finally learns which environment carries a fix, and it learns it from the ship rather than from a guess. The QA column said a fix was ready to check but never which environment had it, and because this team routinely ships cherry-picked branches, merged never meant deployed: ninety-two percent of those flips happened before the fix reached the staging environment, and most bounced tickets were somebody checking the wrong place. Now the deploy pipeline is the only thing allowed to write that truth. After a successful ship it works out from the history which cards are FULLY contained in what was actually deployed, allowing for a cherry-pick having a different identity but the same content, labels those and only those, and posts one evidence note on newly promoted cards. A partial ship gets nothing. Behind it sits a three-hourly recompute that owns correctness rather than speed: it rebuilds every card's state from scratch out of history, tags, run records and the board itself, asserts only the difference, strips stale labels, handles demotions, and is dispatched directly after a rollback. Its dead-man posture is inverted from the fast path on purpose, so a missing token or a schema change reds the run instead of quietly doing nothing. Everything is in shadow until a repository flag says otherwise, no failure can red a deploy, and a separate check reports commits that name no card at all as a warning and never a block. Three follow-ups the same day: the fast path gained a targeted per-push sync so a merged card gets its development label inside the deploy run instead of waiting up to three hours, a dry-run gate learned that a dispatch input arrives as the word true rather than the value, which would have inverted the shadow gate in both directions, and a scan that hit its own cap now says it only read the newest twenty rather than reporting all clear over commits it never opened.
- And a table description that existed only because somebody typed it by hand, one apply away from being deleted. The first plan after importing the reporting table came back with a single change and it was a removal: the description set at creation time was about to be stripped, because the declared resource never mentioned one. Nobody would have noticed until a reader opened the table and found the control arm undocumented. Declared now, and widened while it was open, because a reader arriving with no other context needs the two things easiest to get wrong about that table: it is the control arm and must never be joined to our own data, and its money is gross because the source cannot see cancellations.
A Wednesday of 16 commits, and the through-line is measurement deciding what stays. Two search rules taught in the morning came back out by mid-afternoon with their numbers written into the reverts: the language a deliverable is in moved almost nothing across thirty-six replayed briefs even though the failure it was reaching for reproduces on demand, and a price rule bought a grader-internal correlation the client never saw while thinning the lists. The whole standing-preference feature went quiet behind one constant a day after shipping, deleted from nothing, because with the talent-facing half mute and enforcement still live somebody would be filtered out of every search with no way to see it or raise it. Hours earlier its last two pieces had landed: when somebody can start is now a date screened against the client's deadline rather than a rule sentence that could quietly grade them off everything, and the closing message shows the same snapshot they would get by asking. An account deletion failed nineteen times in an hour and was abandoned with the client's content still in the row. The console's main tab, which had stopped loading, is paged with its filtering moved to the server. A four thousand dollar redesign was ruled small on its timeline and hard-routed to plain search, so clearing the money floor now sends a client to the specialist desk on its own. And the board learns which environment carries a fix from the ship itself rather than from a merge.