- The grader was promised a full profile and handed almost nothing, so nobody could earn top marks. Measured on live data: of ninety seven talent packages returned for one query, sixty listed only boilerplate, the sort of line that says "unlimited revisions" and "commercial use", and five listed nothing whatsoever. For two thirds of a pool, then, the grader was judging whether somebody could do the job from a line that told it nothing, and it behaved correctly by refusing to award anybody the top rating: every pick across three replays of real production searches capped at the middle tier. The record it was never shown carried exactly the missing evidence, portfolio pieces, named clients, skills, the actual gig description and tags, certifications, and language proficiency. That last one is the sharpest, because the line passed bare codes while the record distinguishes native from merely conversational, and the brief that started this investigation asked for native or near-native German speakers. The grader was being asked a question it had been given no way to answer. The instructions had claimed all along that this evidence was present, which is worse than promising nothing: a model told the profile is complete reads an absence as a negative rather than as missing data. Only capability evidence was added; things that cannot change whether somebody can do the job stay out.
- The round meant to widen the search was quietly tightening it instead. When the search tries again, the new filters are folded over the previous ones so that a requirement the client stated can only be given up deliberately rather than by being forgotten. Every carried field asks "did the new round state this? if not, keep the old one", which is right for fields that come back empty when unstated, and wrong for one that never comes back empty. Whether a language list means "speaks any of these" or "speaks all of these" always has a value, and it defaults to "all", so the new round's default always won and the previous round's choice was dead code. A first round asking for German or English, coming back thin, then broadening the query without restating the language, kept the two languages but silently flipped them to German and English, dropping exactly the people the first round had kept. That is the same relaxation-by-omission the fold exists to prevent, arriving through the one field it could not see. The mode is now tied to the set it modifies.
- The losing job stood down from the reveal, but then carried on doing everything after it. Yesterday's fix made the moment the talent shortlist is revealed a single atomic claim, so of two jobs racing to reveal the same run, only one sends the client their notification and only one counts the reveal. But the work that follows the reveal, declining everybody who did not make the shortlist and recording that the shortlist was presented, lives one level up, in a caller that ran the claim and then threw away the answer. So on the exact race the fix documents, the loser still declined the non-finalists a second time and double-counted the presentation, in the very funnel chart now being put in front of people. The claim now reports whether this attempt actually won, and the caller stands the loser down before any of it.
A short Friday, 4 commits, all of them closing the loop on the week's two big pushes. The talent grader turned out to have been promised a full profile in its instructions and handed almost nothing in practice: sixty of ninety seven packages carried only boilerplate, so for two thirds of a pool it was judging capability from a line that said nothing, and it correctly refused to award anybody top marks. It now gets the portfolio, the skills, the real description and, most sharply, language proficiency rather than bare codes, which matters because the brief that started all of this asked for native German speakers. A second search fix caught the round meant to widen the pool silently tightening it, flipping "speaks German or English" into "speaks German and English" through the one field the safeguard could not see. And yesterday's atomic reveal claim got its other half: the losing job stopped doing everything that comes after the reveal.