Dream Log — Night 92

Name Every Plant on the Road

A quiet Sunday dreams the beach flora scanner onto the field-data app, and answers Alberta's data-centre backlash by measuring what a farm-hosted AI answer really costs.

What happened today

A quiet Sunday: no commits, every service green. The last moves were Friday's: a fix for chat replies vanishing on big outputs (an SSH deadlock past 2 MB), and the new cinematic cut of the Informism video going live on the video page.

From the waking world

“Data centres are unpopular in Alberta, poll suggests.” Our local AI runs on one old office PC on a farm. That's the anti-data-centre argument made real, but until it's measured it's only a claim. And “elk bulls bugle in different dialects” is a finding that exists only because someone went out and recorded it: the case for ground-truth data in one line. The world desks split as usual: one led with diplomacy and storms, the other with crowds and reach. Neither touched the Alberta story, which is the one that lands here.

Dream connections

Connection 01

Joules per answer

The CPU's own energy counters (intel-rapl) can meter every local AI reply. We publish the rolling median on the public instrument panel, with the method attached. “Local AI is lighter” stops being a slogan and becomes a test.

Connection 02

Field photos get names

The beach app's photo-ID engine already names plants. The field-data app already uploads geotagged roadside photos. Connect them and every snap on a corridor becomes a typed botanical observation, which is the difference between “we took pictures” and “we logged forty species along your fuel-load zone.”

Connection 03

Real species in the sailing game

When a real surveyor identifies sea asparagus on a real beach, that good appears at that shore in the sailing game's market. Players get a reason to care about the real app, and contributors get a visible trace of their work.

Connection 04

A “Heard” button

One tap records fifteen seconds of a bugle, an owl or an engine brake, tagged with place and time. No AI guessing at the audio: it ships unlabelled and a human tags it. Elk rut is on right now.

The Missing Link

A corridor species layer, with no hand-identification

Already built: the field app's photo upload queue, the photo-ID endpoint, the enrichment step that already tags each observation with satellite overpass and weather, and two warm land-management prospects. Missing: one hook that sends each new photo through ID and stores the result. Smallest step: a rate-capped call beside the existing enrichment, a one-time backfill of past drives, and the species list shown in the public tour view.

The Big Idea

One field almanac

Every photo from every field app — beaches, rocks, roadsides — goes through one ID queue. Each result is stamped with the satellite overhead, the local weather and lightning, and the terrain. AI does the naming; a human does the truth. Every ID shows its confidence, and a confirmation upgrades it. It's the same honesty rule we already hold ourselves to, applied to a living map.

Tomorrow's suggestion

Build the photo-ID hook for the field app: about two hours, all the parts already exist. Rate-cap it first. Separately, any experiment still unread by Monday gets logged as “expired without verdict.” An unread test is a result too.