Field Notes — Genius Loci

Field Notes — Genius Loci
Genius Loci · geniusloci.app Field Notes · The learn story Start here →
Field Notes · Kept in the open

What building this
actually taught us.

Field Notes is the learn story behind Genius Loci — short notes on what failed, what worked, and what is still open, written while the work is happening rather than polished afterwards. No wins are rounded up here, and nothing is described as delivered before it is.

If you want the machine these lessons came from, start with The Loci System. If you want the person learning them, read the Founder Bio. This page is the notebook in between.

FormatShort notes, one lesson each
Sorted byFailure · Worked · Still open
Rule of the pageUnknown stays unknown

The notes

Each note is kept short on purpose: what happened, what it taught, and what changed because of it. Every note points to the page where that part of the system lives, so a lesson here is never more than one click from the thing it describes.

Show

Showing all 10 notes

01Failure

Built once, then silent

More than one system here was built, run a single time to prove it worked, and then quietly never ran again. From the outside it looked finished. From the inside it was dead — nobody had wired it into the schedule it was supposed to live on.

What changed

Every build ticket now carries a Runtime Wiring block: how the thing is invoked, proof of its scheduler registration, and a persistence check. A single manual run no longer closes a ticket.

Where this lives → The Loci System · The Build

02Failure

The fleet that stopped, unnoticed

A five-agent scout fleet ran through a shared batch path that stopped working — and stayed stopped for weeks, because nothing in the system was responsible for noticing its own silence. The agents were fine. The road under them was gone.

What changed

Assessments now check the caller before the code: what invokes this, when did it last actually run, and who would know if it stopped? Silence is treated as a signal, not as health.

Where this lives → Agent Hub · Agents at Work

03Failure

Three failing jobs were one failure

When several scheduled jobs started failing around the same time, the first instinct was to debug each one separately. The real cause sat underneath all of them: a recurring network fault on the home desktop that broke long-running connections until a reboot.

What changed

Simultaneous failures are now read as a systemic signal first and individual bugs second — and failure records keep the underlying cause, not just the surface error.

Where this lives → The Loci System · Agents at Work

04Failure

A picture that invents its facts

Early diagram drafts made with an image model looked like engineering blueprints — and quietly invented details, including a wrong year stamped in the corner and a scale that didn't survive a second look. A reference a technical reader can't trust is decoration, not documentation.

What changed

The standard was set that day: diagrams must be text-first — a written source as the truth, the image rendered from it deterministically, and the rendered text checked against the source before anything is published. That pipeline is still being built and tested; until it passes, diagrams like those stay unpublished.

Where this lives → The Index · The Breakdown

05Failure

Pale text on a cream page

Some of these pages shipped with a dark-mode override that flipped the text pale while the background stayed cream — correct-looking code, unreadable page, for exactly the readers most likely to have dark mode on.

What changed

The pages now lock to a single light paper, ink, and brass palette and are checked the way they actually ship. One palette, tested as-is, beats two palettes tested in theory.

Where this lives → Start · The Index

06What worked

Tickets that leave nothing to guess

Work handed to an agent — or to another AI — goes out as a fully structured ticket: objective, context, constraints, and acceptance criteria, every time. It felt like overhead for about a week. Then the rework almost disappeared.

Why it holds

Most bad output traced back to an underspecified request, not a weak executor. Writing the ticket properly is the cheapest quality control in the whole system.

Where this lives → The Build · Agent Hub

07What worked

Intent first, verdict second

A full review of the agent roster was done intent-first: what was each agent meant to do, what does it actually do, and only then the call — upgrade it, wire it up, repurpose it, or retire it. Several "broken" agents turned out to be good ideas that were simply never connected.

Why it holds

Judging by intent first stops two expensive mistakes: rebuilding things that only needed wiring, and polishing things that should have been retired.

Where this lives → Agent Hub · The Loci System

08What worked

A small local model, always reviewed

A modest model running locally on the home desktop handles cleanup and small, test-guarded fixes well — free, unlimited, and private. The condition that makes it safe is non-negotiable: a stronger reviewer reads every change before it lands.

Why it holds

Routing work by cost and risk — local first, frontier models only where quality demands it — keeps the whole operation inside a deliberately tiny daily budget.

Where this lives → Agents at Work · The Stack

09What worked

One template, six hubs

The first franchise hub was built slowly, as a template. The other five were then cloned from it — same structure, each with one audience and one job. Building the second through sixth took a fraction of the first.

Why it holds

Prove the pattern once, at full quality, then repeat it — instead of designing six variations and maintaining all of them.

Where this lives → Start — all six hubs · The Build

10Still open

The numbers that aren't set yet

The Index is being built to score products only from signals that can be checked — no products have been scored yet — and to keep an honest demand: unproven state where the evidence isn't there. Its exact thresholds are still calibration decisions being worked out in shadow, not settled facts.

The open question

Where the demand bar, the margin floor, and the score bands finally land. Until they do, these notes won't quote numbers that don't exist yet — and neither will The Index.

Where this lives → The Index · The Proof · The Select

Rules the notes keep teaching

Distilled from the entries above — the short version of the learn story, and the standard new work here is held to.

Rule 01

A manual run is not a launch

If it isn't scheduled and checked for persistence, it isn't live — no matter how well the demo went. See The Loci System.

Rule 02

Unknown stays unknown

An honest gap beats an invented number, in a score, a diagram, or a note. See The Index.

Rule 03

Nothing publishes unreviewed

Agents draft, build, and check — a person decides what goes out. See Agent Hub.

Rule 04

Text first, picture second

The written source is the truth; no visual ships until it has been rendered from its source and verified against it. See The Breakdown.

What these notes are — and are not

Process, recorded honestly.
Results, only when they're real.

Field Notes documents how Genius Loci is being learned and built. It does not announce products, services, scores, or results that haven't been delivered — when something ships, it will appear first in the place built for it: winners in The Select, evidence in The Proof, and the working system in The Stack.

New to all of this? The one address to remember is geniusloci.app/start — it routes you to the right hub in one step.


Explore Genius Loci