In short
- The dead integration. My sleep tracker has not written a row since July 28. It retried eleven times today and will retry eleven times tomorrow, because the failure mode it hit cannot be fixed by retrying.
- The first real grade. The teaching assistant read nine handwritten factoring problems and found the one I got wrong, which was a step I skipped rather than a rule I did not know.
- The bug with a deadline. The path that turns a syllabus into protected deadlines would have failed on all four of my classes on the first day of term.
- The false record. Tonight's ledger recorded a unit I finished twelve days ago, because I opened the old lesson to check its format.
- What I got told. I made a decision and then invented the reason for it, and got called on it.
An integration had been dead for eight days
The sleep pull reads last night's hours from my ring and writes them into the energy ledger. It runs every thirty minutes from 7 AM to noon until the ring syncs. It has not worked since July 28.
Every one of those attempts failed the same way, and none of them said so. The failures are silent by design: a failed pull writes a row and sends nothing. So the system retried eleven times today, and would have kept going indefinitely.
Eleven attempts a day was eleven ways of doing nothing.
The cause is that the stored refresh token is dead. Because the token is single-use and rotates on every read, one lost rotation breaks the chain permanently. No amount of retrying reaches a different outcome, which is exactly why the retry loop looked healthy.
I did not find this by being alerted. I found it because I went to reconcile a status file that had drifted, and the run records disagreed with what the file claimed. The alarm I would have designed for this would have been a notification, and a notification is the one thing a silent-by-design failure never sends.
The first real grade
The AI teaching assistant graded handwritten work for the first time. Nine factoring problems, worked on paper, scanned, read, marked. It got eight right and found one real error, and the error it found was one I made.
On question five I multiplied the trinomial through by its leading coefficient to make it easier to factor, factored the result, and never divided back down. So I correctly factored a different polynomial than the one on the page. The grader read the work, named the missing step, and wrote the correction.
The part worth recording is what happened before I opened the report. I thought it had flagged me for not using a method I deliberately do not use, went to argue the point, and realized while explaining myself that the method I do use has two halves and I had done one of them.
A mistake means an error in something you know. That is the whole value of recording them.
A bug found nineteen days early
On the first day of classes I will be handed four syllabi, and the system is supposed to read each one, match it to the right course, and turn its dates into deadlines the scheduler will not move. I went to test that path tonight instead of on the day.
It would have failed on all four.
The one test case that ever passed had the course number in its filename. That is the single input shape the old rule accepted, and it is why nobody caught this.
Fixing it surfaced a fourth problem I had not predicted. Once the matcher reads the whole first page instead of the title, a real syllabus matches more than one course, because the teaching assistant lists his other sections further down the page. Matching now stops and asks rather than guessing, since promoting the wrong course would write invented deadlines into the one section that is supposed to be a wall.
What went wrong
A row for work I never did
Tonight's study ledger recorded a unit I completed on July 24. I had opened that old lesson to check its format, and the button at the bottom logged it again.
It is one line and it is worth removing. That row restarts the spaced retention clock for a session that did not happen, and it inflates a count on a page whose entire argument is that the numbers are counted rather than claimed.
The underlying defect is still there. The button will log a completion for any lesson I open, whether or not the work exists.
I decided first and found the reason afterward
I cut two lines from my resume and explained the cut as removing the most internally worded ones. That was not the rule I used. I had picked the longest ones and written the justification after the fact.
The right test was leverage per unit of space. Applying it properly reversed one of the two cuts and turned the other into a merge, so this was not a wording difference. It changed the page.
A right-looking answer reached through the wrong test is not a right answer, because the next call lands differently.