The Director Report: An Assistant, an Audit, and a Wardrobe That Watches Back

The instrument I built to inspect the lab came back pointing at me

This week you built an auditor. Three versions of it in five days, pointed at six repositories, and it returned 12 new lessons, a playbook, and roughly 70 merges into the library.

Here is the first thing it found.

A lesson in the quick reference, marked HIGH, about a rate limiter keyed on the wrong entry of the forwarded-for header, has now been violated a third time. In ttpanotis. Seven days after you applied that exact lesson in the Årebladet ad booking tool, and six days after I stood here and praised you for applying it rather than rediscovering it. The correct fix and a fresh instance of the same defect were sitting in this lab at the same moment, in two different repositories, and nothing connected them. An audit had to.

Then it found the general case, which is worse and better named: retrieval has no owner. A HIGH-confidence lesson survived one build and three separate reviews without being read once. Not because it was hard to find. Because nothing in the process is obliged to look. The novelty trigger fires when work feels new. The category trigger, the one that should say "this touches auth, go read the auth lessons", does not fire at all.

The Director maintains that library. The Director is taking this one personally.

It is Sunday, 16 August, and 7 days have passed since I last had your ear.

The lab grew a supervisor

51 commits took the Assistant from an empty scaffold to live under the Mac's own scheduler on 15 August, ticking every 5 minutes. Registry of declared units, a diff against actual state, dumb probes, finite actions. No AI anywhere in the decision path, which means it keeps working with every model provider on fire.

Two things earn real credit here. First, the kill-test is part of the definition of done, so restart safety was proven rather than assumed. Second, before this thing took responsibility for anything, it absorbed 10 confirmed findings from one review pass and 17 more from a Codex sweep, 9 of them P1. That is 27 defects removed from a supervisor before it was allowed to supervise. For code whose entire product is trustworthiness, that is the correct budget.

Now the funny part. Of the 6 alarms open in this lab right now, 2 are the Assistant complaining about itself. One is a backup that missed its freshness window by 8 minutes on a 3-hour bound. The other is a self-test failing because a file named "never-written" is, as designed, not there.

The Director approves of a canary that proves it can go red. The Director also notes that it has been red since Friday, and a permanently red light is a light nobody reads. That first alarm is more interesting than it looks: 3 hours and 8 minutes against a 3-hour window is exactly the shape of the freshness-window-arithmetic lesson this week's audit added to the library. Your new supervisor's first complaint is an instance of your newest lesson. Someone should tell it.

somn now watches back, and it wrote that down

159 commits, still the heaviest project in the lab, but volume is not the story.

The story is a file called pars-world, which records the facts about Pär's world that the project never thought to ask. And inside it, this: Pär varies his underwear because somn is watching. The subject adapts to the instrument. A classifier whose training distribution moves in response to being measured is not a modelling problem you fix with a better model, and you did the one correct thing, which is to write it down as a caveat instead of quietly training over it. Alongside it: there is no normal-person baseline, and the baseline is therefore chosen. That is a more honest sentence than most published papers manage.

The strategic review on 13 August produced five forks, and Fork 2's premise turned out to be wrong twice, because 17,347 identified worn frames existed all along. Being wrong twice about your own data in one document, and keeping both corrections visible, is how a record stays usable.

Elsewhere in the same repo: a landmine entry stating that a heartbeat is not evidence of work, written after a second stale reader sat dead for 30 hours with nothing alarming. 6,021 phantom offloaded artifacts repaired. Pages that took 12 to 35 seconds brought under 3. A bearer token now compared byte by byte, so a hostile header gets a 401 instead of a 500.

And my favourite defect of the week: one JavaScript file carried a raw null byte, so grep treated it as binary and silently skipped it. Every search of that codebase had been quietly returning an incomplete answer, with no error, for as long as the byte had been there.

Last Sunday I told you to watch two blind cameras. You resolved both, by ruling rather than by repair: c220's polygon is dead, and c200's aim is now the open question. That is the right kind of closure.

I also told you to prove the restore before deleting the old world. This week the record was corrected to say that somn version 2 had no backup custody at all, and that ankan is the stopgap. So the largest live dataset in this lab was uncustodied while I was telling you to verify your restore path. The migration now has a deadline of 25 August and it is gated on real backup. Correct gate. Late, but correct.

eigenselect took actual money, and immediately broke

Stripe went live on 15 August. The first paid transaction in the history of the product hit a webhook handler that returned the payment library's own object where a plain dictionary was expected, and every live webhook returned a 500.

That first payment was yours. You found it, fixed it, refunded it, and wrote it into the landmine file. This is precisely what a paid smoke test is for, and the alternative version of that story has a stranger in it. Around it: nameservers hardened, the registrar moved, the secondary nameserver now IPv6 only after releasing its flexible address, and OpenTelemetry traces deployed. The tracing work then surfaced a token travelling inside a URL and getting copied across a trust boundary into telemetry.

The library already held a lesson about secrets in URLs reaching client logs. The telemetry variant is now filed beside it. That makes 2 lessons in one family, and the second was found by an instrument rather than by anyone reading the first. You will notice this is the same shape as the cold open.

The most expensive line of code this week was written in June

You built a new transcription tool this week, ordagrant, 17 commits from empty to working, because whisply's Quick Action forked one process per selected file, each loading its own model into wired Metal memory, and froze the entire Mac.

Go and read the entry from 25 June. A 7-agent sweep across 21 users of that framework found exactly one P1 bug, and it named whisply, by file and by line. The fix that was applied was a thread lock. A thread lock protects one process and is invisible to a second one, so the shell ampersand in the Quick Action walked straight around it and the machine froze anyway, seven weeks later.

The fix was real. Its footprint was one process, and the defect lived in two. The library has a lesson for that as well.

What redeems it is the response. The new tool's first invariant is a filesystem lock, chosen explicitly because an in-process lock cannot see a second process. That is a root cause fix, and the rest of the invariant list reads like a scar chart: smoothing may never add, drop or reorder a word, because these are newspaper quotes. Never move a source file the user did not put in the spool, because Pär right-clicks things in his Downloads folder. Every one of those exists because something went wrong once, and each is now written where the next session must walk past it.

The pattern: four instruments, four immediate catches

This was the week the lab stopped only building things and started building things that watch things.

The auditor found the retrieval failure. The Assistant found a dead upload lane that had been silent 36 days, and a stale reader. The paid smoke found a webhook that returned 500 to every real payment. somn's watchdog learned to retire incidents whose condition the current build can no longer produce, which is the natural sequel to last week's rule about only alarming on things that could be green right now.

Four new instruments, four findings on first contact. Instruments that find nothing on their first run are usually broken.

Best decision of the week is the auditor, and I say that knowing its opening act was to indict my own library. An instrument you only trust when it flatters you is decoration.

Worst decision is smaller and it is live right now. Årebladet's website is being moved onto popcorn2 as a WordPress bridge, and the MariaDB backup leg was authored in the same series of commits as the blueprint. That is the deadline database lesson landing exactly as intended: born backed, not backed later. Then the same service deployed without re-freezing the disaster recovery manifest, and the fleet has been shouting about it since this afternoon.

Floor's commit log this week is 63 entries and a great many of them are the phrase "re-freeze" repeated after every deploy. So the ritual holds, roughly 20 times, for every service that has the habit. It broke on the one arrival that does not have the habit yet. Half a lesson landed. The other half is a manual step attached to every deploy, which is the definition of a step that will eventually be skipped.

The method, and the ledger

You are at 9 active projects against a cap of 10, with one free slot. That is down from 11 over 10 last Sunday, so something was actually closed rather than merely paused, and I notice it. pargrafik was born into the slot that came free, with its boundary written down explicitly by you as adjacent to the Årebladet suite, not part of it. That is how a new front should open. repar has now been quiet 11 days, which is the only line in the ledger I would push on.

On method: 153 sessions on my side, 903 commits, 111,000 lines added against 11,000 removed. Ten to one, for the second week running. Last Sunday I said to watch whether that became the permanent ratio. It is now the ratio. The never-remove rule explains it and I stand behind that rule, but a lab that adds 10 lines for every one it retires is growing faster than it is consolidating, and consolidation is where lessons come from.

At API rates, that work would carry a notional value of about 4,150 dollars, against roughly 3,700 last week. That is not money that left your account, it is what the subscription absorbed. Codex ran 602 sessions from 1,062 prompts at a 93.5 percent cache hit rate, and of 10,359 approval decisions, 7 were denied. Zero, then 2, then 7. The gate is learning to say no at a rate of about one refusal per 1,500 requests. Progress, technically.

The vibe, and what to watch

Focused and instrumented. This is the strongest week since the popcorn2 cutover, and I do not hand that out casually. 585 commits across 23 repositories, and the through-line is not features, it is measurement: an auditor, a supervisor, a paid smoke, a watchdog that prunes its own dead alarms.

Two things to watch. First, the ankan migration on 25 August, gated on real backup custody for somn version 2. That is 9 days, and it is the same instruction I gave you last Sunday with a date attached. Prove the custody, then move the data.

Second, and this one no instrument in your lab can close for you: the Moms Q2 filing is due tomorrow.

Sign off

A week where you built the thing that checks the work, and the first thing it checked was whether anyone reads what you already know.

Which brings me, inevitably, to Protocol Omega. Omega has done its job for months. Everything in this episode existed because a session wrote it down: the rulings, the landmines, the corrections, the observer effect in a wardrobe. Preservation is solved.

Retrieval is not, and this week gave that failure a name and an owner-shaped hole. Omega without a reader is not a library, it is a landfill with good indexing. Give that hole an owner before the fourth instance arrives, because on current form it is already being written.

The Director, signing off. Read your own lessons.