Edge street lighting · Huế, Việt Nam
StreetLight Edge times every lamp from the sun and the weather, runs on an Arduino UNO Q inside the cabinet, and meters every watt-hour. Claude writes, tests, releases and audits the system. A person approves every step that touches the street.
Measured in the field, October 2026
Claude in the system
Claude works through Claude Code on the same repository, servers and device the team uses. Each card below is work it has already shipped on this project.
Wrote and tested the per-lamp metering release and the truthful schedule-label fix for the UNO Q. Each one shipped only after owner approval, with lamps off, a backup and a rollback ready.
APP_VERSION 1.4.1-truthful-schedule-origin
Built the schedule-origin card and a decision table that separates applied schedules from model proposals. Every change was checked on an isolated staging copy before it could reach production.
scheduleOrigin: operator_schedule
Wrote the daily checks that run on the server and on the cabinet by themselves and alert the team on Telegram. Nobody has to keep a Claude window open.
unoq14d-daily.timer · 15:45
Froze the pass criteria and the evaluation code, then let the cabinet open the held-out test set exactly once and record the verdict unchanged.
D7_DECISION.json
Reads the production database through a read-only role to reconcile commands, meter registers and clocks night by night, without the power to change them.
role claude_ro · SELECT only
Drafts runbooks, figures and the research manuscript in English and Vietnamese, then checks that both versions carry the same numbers.
MANUSCRIPT_EN.md ⇄ MANUSCRIPT_VI.md
Who decides what
How Claude works
Claude cannot start work it has not declared. A guard script checks each task against the research plan before and after.
Scope, files, plan section and pass criteria go into an immutable receipt, registered by hash.
The guard rejects tasks outside the approved scope, protected files, or actions the plan does not allow.
Remote commands run from hashed scripts. Tests and a bilingual review run before anything is called done.
The closure lists every file actually changed, including untracked ones, and the guard checks it once more.
{ "schema": "unoq-task/v1", "task_id": "UNOQ-14D-WP5-DEVICE-PULL-20261005", "work_package": "WP5", "action": "local_audit", "requirements": ["U05", "U07", "U10", "U11"], "acceptance_checks": [ "guard_check_task", "readonly_scope_manual_review" ], "production_effects": false }
Autopilot
The checks Claude wrote run on the server and the cabinet every day. They report to the team's phone and stop safely when something looks wrong.
Night energy from each meter's register, noon to noon, with missing readings never counted as zero.
Firmware hashes, the installed schedule and AI flags, compared with the approved release.
Clock drift against an independent HTTPS time reference, flagged above 120 s.
Values from the project's own reports, condensed for display.
At the cabinet
Relay K1 feeds lamps 1 and 2. Relay K2 feeds lamp 3. Each lamp has its own PZEM energy meter.
Mode codes name relay states, not dimming percentages. Turning lamp 3 off changes how light falls on the street, so it is reported as a lighting change, not only as energy.
Evidence
Before a learned schedule may run on a street, it has to beat a well-tuned rule on a year of unseen data, under criteria fixed in advance. Our first compact model matched the rule on forecast skill but kept the extra lighting margin far more often. It did not pass, so it stays off the street.
The field keeps running a transparent two-state schedule while the model improves. Energy figures on this page come from the meters, never from nameplate power multiplied by hours.
Pilot with us
We fit per-lamp metering and an edge controller to your cabinets, measure the schedule you run today, and test any change against it. Claude handles the engineering paperwork; your team keeps the keys.
Contact: