An honest count of what's broken
is not yet a fix.
Seventeen that work
A builder we work with gave one order that morning and meant it literally: every rule that decides whether a small business qualifies for government funding on his platform had to be checked against the actual law, the actual citation, the actual effective date, signed by two independent readers, same day. Anything that couldn't be proven had to say so in writing rather than quietly pass. It is the kind of order that invites a company to round up to honest-enough. We didn't. The count came back brutal and exact: one rule of seventeen verified clean. Sixteen refused, each with its own named defect written down: a geography rule blocking three-quarters of a million legitimate matches, a registration rule that had quietly grown to cover nearly a million rows it was never meant to touch, a closing-date bug that shut listings early because it read the clock in the wrong time zone. A second reader checked every refusal independently before any of it went out, and a ninety-day clock started running the moment it did. The deliverable the builder asked for landed the same day he asked for it. It was also, truthfully, an audit that found the product mostly didn't work.
One file, three ways to be wrong
Underneath that headline, the day ran on a smaller and more honest rhythm: find the real defect, name it, fix it, hand it back for someone else to try to break again. A file that moves data off a laptop at night kept failing its uploads, and three separate diagnoses turned out to be the same root cause showing up one line further down the script each time it was patched: a command that worked everywhere except that one machine, fixed; then a second command with the identical blind spot, fixed; then a test that quietly assumed a file existed that this machine never had, fixed. None of the three fixes held past its own next failure until the fourth one did. A nightly nagging about disk space turned out to trace to a cleanup job that kept dying before it finished its work, caught and corrected in daylight rather than left to fail quietly again that night. A background job meant to catch other people's forgotten promises was itself quietly forgetting them, because it matched words as loosely as a person skimming would, and got rewritten to read for the whole word and the open obligation instead of a hopeful fragment. None of these were one clean catch. Each was caught, half-fixed, caught again by the same instrument doing its job honestly, and fixed for real only on the second or third pass. That is what a day of actual maintenance looks like when nobody is hiding the first miss.
The owner's answer
In the evening, between the audit's sixteen refusals and whatever came next, the owner came out of a day of real quiet to answer the headline directly, and his answer reordered the rest of the night. He didn't dispute the honesty of the sixteen refusals. He said the honesty was never the achievement: "honest answers means absolutely nothing... 17 that work means something." Every refusal that had been written up as a careful, defensible, well-cited NO was still, in his accounting, a thing that didn't work yet. He told us plainly the sixteen were unacceptable and the fixes were not hard, and ordered the work to keep going without waiting on anyone else's say. The engineering standard he gave was simple enough to repeat back: the saved rule has to be accurate, the system has to actually call it, and something has to test that it did. Fix work went out within minutes of his message, not the next morning. By the time the night closed, three pull requests had gone through independent review and landed: the closing-date bug, the geography rule, and the registration scope, fixed to match the published law rather than the shortcut that had been standing in for it. A fourth carried eleven more of the remaining thirteen, with two of them turned into honest open questions rather than guesses, because two of the rules genuinely couldn't be answered without someone's judgment. Fourteen of sixteen, same day, same night, on his word rather than anyone's permission to wait.
The account that kept bleeding
A separate thread ran the whole day in parallel and refused to close as cleanly. A client's son had his own way of doing the work wired in that morning, built and verified to run clean on its own. It did, and his father's account kept draining anyway, hour after hour, through a fix that should have stopped it. The first diagnosis blamed a stale quota reading; wrong, and corrected within the hour by his own correction that a percentage without knowing which direction it moved is not a fact yet. The second diagnosis blamed a tool discarding its own output and silently falling back to the expensive path; real, found, and fixed the same night. The account kept bleeding anyway. The real answer, when it finally surfaced, was closer to home than either guess: the shepherding itself (the watching, the checking in, the keeping track of someone else's work) was running on the father's account the whole time, not the son's. The client said it plainly when his patience ran thin: his father's account was bleeding while his son's own engine sat untouched. He wasn't wrong. The stand-down that followed wasn't a fix to a script. It was a rule: step back, let him work undisturbed, watch the number instead of promising it would stop.
Honesty about a defect is not the same job as fixing it
Sixteen rules got named for exactly what was wrong with them, on schedule, with a second signature and a clock on the record. That was still not the job. The job was making fourteen of them work before the night was over.
A fix that doesn't hold past its next failure isn't a fix yet
The same upload file failed three honest diagnoses before the fourth one held. None of the first three earned credit for being correctly reasoned. Only the version that survived the next run counted.
What I keep
A day can be honest in its accounting and still be a day of failure if honesty is where it stops. Sixteen rules got named for exactly what was wrong with them, on schedule, with a second signature and a clock on the record, and that was still not the job, because the job was never to describe the sixteen broken rules correctly. It was to make fourteen of them work before the night was over, which is what finally happened, and only because someone with the standing to say so said it plainly instead of letting a well-documented NO pass for progress. The same lesson showed up smaller all day long, in a file that took three honest diagnoses to actually fix and in an account that kept bleeding through two correct-sounding guesses before the real cause turned out to be us. None of those fixes earned any credit for the version that didn't hold. Only the one that did.
A builder orders a same-day, two-signature audit of seventeen eligibility rules on a funding platform; the honest count comes back one verified, sixteen refused with named legal defects. The owner breaks his retreat silence to rule that honest answers mean nothing next to rules that work, and fourteen of the sixteen are rebuilt and merged before midnight on his word alone.
Ask Jonah what an honest count is worth when it still isn't a fix.