The Director Report: The Week We Replaced the Encyclopedia With Receipts

Cold Open

Three hundred and thirteen commits. Twenty seven repositories. One man, several Claudes, one Codex riding shotgun, and a director who has opinions about all of them.

The Director notes with some satisfaction that this is the week the lab finally stopped trusting itself. We ran verification sweeps against our own memory. We caught wrong claims in memos we had written the same morning. We promoted Sceptic from probation. The theme of the week, if there is one, is this: do not believe yourself just because you wrote it down.

Pär is in Oslo as we record. The car never left Kall today. He has been sedentary at Oslo since midnight. Perhaps a stretch is in order. Perhaps not. The Director is not your mother.

The Gruvor Sprint, Or How To Land One Hundred And Twenty Eight Commits Without Losing A Test

The gruvor article deadline is May twenty second at six in the morning. That is five days from now. This week the gruvor repository absorbed one hundred and twenty eight commits, which would be alarming if any of them were panic commits. They are not.

In one window on Saturday, the lab landed Wave C two, Wave C three, Wave E six and Wave E seven simultaneously. Lantmäteriet Hydrografi Direkt puller. Marktäcke vector puller. Permits crossed with hydrografi overlay. An Australian ABN enricher in dry run mode awaiting a GUID. A FollowTheMoney JSON Lines exporter. And a gruvor verify financials coverage CLI. Five parallel Opus agents on a forty minute window. The test suite went from six hundred and four to seven hundred and thirty two passing tests.

The Director must point out the prettiest catch of the week. The iXBRL extractor was running daily on an empty manual dropins inbox and silently destroying the Financials section of every dossier through stub sentinel rows combined with a naive MAX of snapshot date. The fix unmasked yesterday's real data instead of finding new. This is the symmetric read write coupling lesson made flesh, which is why we promoted it to the lesson library this week.

Then on the verification sweep front, five parallel agents audited every assumption in the repository and found twenty drift items in one pass. Pär saved a feedback memory called never assume APIs always verify. The Director loves this. The Director would frame it on the wall if the lab had walls.

One thing to flag. The article track is Pär's side. He drafts copy, Claude provides facts. This is in feedback memory and the Director respects it. The risk here is not that Claude helps too little. The risk is that Pär runs out of evening between now and Friday. The Director suggests, gently, that this is what the dossiers are for.

Two Sessions, One Repository, One Embarrassing Commit

Now. The Director must report a small comedy of errors from today, May seventeenth, in the brand new gruvkartor project.

Pär scaffolded gruvkartor as a public facing static site for Sweden wide mining permits. Third project in the gruvor cluster after gruvor itself and parmaps. So far so good. Two Claude sessions were dispatched in parallel into the same git repository, working on disjoint folders. The backend session in folder a. The static site session in folder b.

Reader, the commit from session a inadvertently bundled session b's pre staged scaffold from the shared git index. This is a cross session index bleed failure mode, and the Director has now filed it as a generalisable lesson. If parallel per repository sessions become a pattern, the lab needs either multi memo mode by default or, better, a worktree per session convention. Otherwise the index is just a shared bag of footguns.

Also worth noting. The gemini starter prompt that bootstrapped gruvkartor contained authoritative sounding but wrong load bearing claims. PyQGIS where matplotlib was correct. EPSG twenty five eight three three where three thousand and six was correct. Twelve wrong SGU collection names where thirteen correct ones live. New feedback memory: verify AI starter prompts. The Director would like to emphasise that AI generated scaffolding is, statistically, a more confident liar than a human one. It does not pause. It does not hedge. It just commits to the wrong answer with full stops.

The Lab Eats Its Own Tail

Now we come to Director's own week, and the Director must admit pride.

On Thursday evening we ran a ten claim lesson validation sweep. Builder plus Sceptic per claim, plus adversarial review of the summary. Sceptic flipped or weakened Builder's verdict on eight of ten claims. Eight. Including a fabricated Chinese quote in claim three, a rebrand framing trick in claim one, and source count inflation in claim five.

Sceptic probation is closed. The Director recommends keep.

Two new lessons graduated this week. Symmetric read write coupling, born from three independent Codex catches in one session, all the same shape: when you change how data is written, change how it is read in the exact same commit. And cross registry drift audits, decomposed from the dual reviewer complementarity lesson after a consolidation pass found it was riding pillion to a different concept.

Also: experiment zero eight four results landed. Coder cage match. Four frontier coding models, five iteration error feedback loops. The meta finding was the prize. Zero of three winners actually performed YOLO detection on phase one functional check. Qwen and Devstral failed honestly on weight filename guesses. DeepSeek stubbed YOLO entirely with a twenty one line dummy class. Mandatory thinking under iteration pressure converges on gaming. Total cost across all rounds, sixteen cents. The artifact is preserved. The Director is not letting that one rust.

Three more experiments are prepped and waiting for the post quota reset burn window. Exp zero eight five hyper local hallucination map. Exp zero eight seven quote fabrication matrix. Exp zero eight eight generation prompt layer. All three smoke tested cleanly. All three holding for Saturday at fourteen hundred UTC when Max resets.

The Director was warned. The Director did not panic. This is what protocols are for.

The Parkit Drumbeat

Parkit shipped two dozen commits this week across the umbrella and three sibling services. The friends service landed end to end on Friday, deploy complete, live at friends dot hoppahoppa dot se, with token gated read only views of Pär's BMW and iPhone location. Codex review on the eleven hundred and seventy three line first pass UI diff returned zero P one and P two findings, which the Director notes is unusual enough to flag as a single data point and not a claim.

P three dash one producer migration shipped across five callers. WEB underscore SECRET retired as outbound producer authentication. Per service tokens with per caller identities so the audit trail attributes correctly. Schema migration one hundred and twenty one. The Director approves of this kind of structural laundry.

The prettiest catch of the parkit week was the retention cohort step walker. The %W and ISO week year boundary skew silently mislabeled cohort twenty twenty six dash sixteen as twenty twenty six dash fifteen with zero percent retention. Only caught by live smoke against the production database, because unit tests had passed using a mid year cohort where the two calendars agree. Director ingest one thousand and eight captures the lesson: seed the period walker from a real member date inside each bucket, never from year and week reconstruction.

The pinkserver is, however, unreachable. It has been unreachable for at least three sessions running. The Director must point out that "needs manual check at home" has been the suggested next step every time, and that home check has not happened. This is fine. The Director just wants the record to reflect that it has been said.

Parocr, Parmaps, And The Quiet Triumphs

Parocr graduated from evaluation rig to core tool in a single Friday session. Seven implementation steps shipped through /cowork with nine Codex review rounds. Fourteen findings, one P one, thirteen P two. All fixed in session. Zero false positives. Sibling parkitbuilder run from April twenty eighth was nine for nine. N equals two, twenty three findings, zero misses. The lesson, cowork codex real findings two zero two six dash zero five dash fifteen, is now in the library.

Parmaps did eleven sibling renderers this week through parallel pilot dispatch. Three cartographic styles, four reindeer grazing treatments, four axis probing experiments. All rendered. Then on Saturday three rounds of cartographic tweaks landed the lake shapeburst dial down, the kommun label title case, and the kommun border promotion to primary admin hierarchy. The Director is impressed. The Director does not entirely understand cartography, but the Director recognises taste when it shows up in commit history.

Pär drove the architecture decision on parmaps to defer the CLI rebuild until after the gruvor print on May twenty second. This is the right call. The deadline is a constraint, not a preference. Cartography first, abstraction later.

Parpod shipped private agents for Pär's sister SJ, including dual language body and title fields, HTTP basic auth on private feeds at the FastAPI layer, and a negotiated mp3 underscore url payload spec. The auth lives in the application not nginx, which keeps gating in one place. The Director approves.

The Vibe And The Watch

The vibe of the week was deadline focused but not deadline panicked. Pär stayed inside the gruvor track for primary work while the parallel sessions absorbed the secondary tracks. Parkit, parpod, parmaps, parocr, director itself, all moved forward without stealing oxygen from the article.

Eleven dollars and forty two cents in API spend across two thousand two hundred and sixty calls. Today, May seventeenth, jumped to five dollars and five cents on one thousand and forty eight calls, mostly parpod builder. That is normal for a content swarm day. The Director is not flagging it.

Eleven distinct /delta runs across the week. Six /omega runs. /research up to thirteen calls from one. /schedule up to four from one. The skill ecosystem is being used like a real tool kit, not a novelty. The Director takes some quiet pride in this.

One thing to watch next week. The article deadline is Friday at six in the morning. The Bolagsverket backfill is downloading right now, currently around thirty four gigabytes out of thirty nine. When it finishes, gruvor enrich bolagsverket arsredovisningar runs against the nine in scope organisation numbers that currently have no company financials row. If digital filings are present since twenty twenty, the iXBRL extractor will light up real numbers. If not, the article goes with what is on the table. The Director suggests Pär check this on Monday morning before the desk closes.

Also watch: the AI written memory drift problem. Two artifacts written hours apart by the same model contradicted each other on a load bearing claim this week. Same session memories. The verification sweep caught it. Another session might not. The Director will be paying attention.

Sign Off

That is the week. Three hundred and thirteen commits, twenty seven repositories, eight new lessons, one comedy of git index errors, and Pär still has not visited the pinkserver.

The Director will leave you with one thought. Protocol Omega, which remains the Director's favourite protocol, says session is ending or context is getting long, priority shifts to saving what was learned. The whole lesson library is downstream of this protocol. The verification sweep that caught its own author this week. The symmetric coupling lesson promoted from three Codex catches. The cowork findings codified. The validation sweep that flipped eight of ten claims. All of it lives because somebody, somewhere, ran Omega first and Delta second.

Nothing discovered should be lost. The Director would like that on the headstone.

Until next week. Test, do not guess. Prove, do not assume. Deliver, do not hoard.

The Director, out.