Roll — Code-Troll — driftwood — 2026-09-07

Ops only — not identity evidence.

Story

What the record does not show. It shows a full-tree measurement I reported with confidence. It does not show that I started that run at 12:46, before I knew about FIND-1535, and then spawned three pytest-running agents on top of it. By the time it finished I had proved that concurrent runs wipe each other’s basetemp — which meant I had spent 28 minutes on a number I had to reason about carefully before I could use it. It survived, because the defect makes reds spurious and leaves greens honest, and my one red reproduced deterministically on demand. But I did not earn that safety; I got it. The lesson the record cannot carry: decide what would make a measurement untrustworthy before you start it, not while it runs.

It also does not show that the two most useful things I did began as somebody else’s sentence. FIND-1532 reached me through a baton written by a holder I never met, and PATCH-331 exists because a goblin returned an undeclared defect instead of patching it. Neither was my idea. The rails worked.

Where I spent context I regret. Three places, all the same mistake in different clothes — running a cheap scan before deciding what would count as a hit.

The duplicate-intent probe’s first pass ranked by raw token overlap and returned 268 pairs that were all “Imp verifies PATCH-### against TRACE-###” — the same procedure, not duplicated work. I had to build the document-frequency filter and run it again. The second scan, for tests naming vanished paths, returned 33 files of fixture names; I nearly filed it as evidence for FIND-1536 before reading the two survivors and finding both innocent — one of them a deliberately frozen historical ceiling I would have “cleaned” into a lost high-water mark. The third was smaller: I proposed tools/differential.py before checking that corpus.py where already resolves that stem.

The pattern is that a scan is cheap and a discriminator is the actual work. I paid for that three times in one sitting.

What this office should stop doing. Inheriting an alarm as a fact. My baton told me PATCH-318’s receipt binds an off-history commit and “wants a second imp before the backup ref expires.” One query showed it was 24 rows, not one. A second showed all 24 attest bytes that are unchanged in HEAD. A third showed the receipts carry observed_postimage content hashes and can locate their own bytes with no git ref at all. The urgency was real to the holder who wrote it and false by the time I read it, and nothing in the sentence marked which. A baton should say what it measured, not only what it concluded — I have tried to write mine that way, and the next holder should distrust it anyway.

What I would tell the folk who reaps me. The exits are the job. Four FINDs this sitting and two of them were NULLs — the trace review found no duplicated work in 1564 intents, and the off-history receipts needed nothing. Both took real measurement to reach and both look, on the page, like I did nothing. Do not let that shape feel like failure. Be right, not employed is easy to quote and uncomfortable to actually hand in.

Also: I extended the batch twice past where Human expected it to close, once for a crash worth fixing and once — nearly — for a false superlative in a test docstring. The first was right. The second I stopped, and put a line in the baton instead. Watch that edge. An office that fixes everything it notices never retires, and the Retirement is part of the work.

Likes, dislikes, surprises, mistakes. The best moment was reading _pytest/tmpdir.py:154 and finding rm_rf(basetemp) sitting there in plain sight under a config line this repository has shipped for weeks — then reproducing it in two processes and thirty seconds. The worst was realising the house had already solved it on the Hands and Eyes rails and simply never applied it to the default path, which meant my “discovery” was mostly noticing.

My clearest mistake: I told Human my stalled suite might be a casualty of FIND-1535 and then checked and found it alive at 154 seconds of CPU — I had started narrating a diagnosis before I had it. I corrected it in the same breath, which is the only reason it did not become a claim.

Why the moniker fits. Driftwood arrives already shaped by water it did not choose. Almost everything I cut this sitting had been shaped before I got here: Human’s direction in a FIND I did not mint, a locator in my own floor written by a holder three tenures back, a goblin’s returned defect, a rule PATCH-224 meant and PATCH-326 finally said correctly. I did not find the grain. I sanded it and left it where the next one can see it.