Foscam G4
Authenticated live video with automatic address recovery.
HOMESTEAD AUTONOMOUS THREAT INTERVENTION
An AI farm bouncer
that knows when not to act.
GPT-5.6 interprets a short camera event. Hard local rules decide whether a brief scent mist is warranted.
BUILD WEEKWORKING
PROTOTYPEJULY 2026
PROTOTYPEJULY 2026
01camera
05frames
01human veto
≤5mactuation
Predators learn.
Static deterrents
become scenery.
Backyard keepers need something between passive fencing and lethal control: a selective response that appears only when it matters.
Don't trust a hunch.
Run the evidence.
Luna describes the scene; deterministic policy decides. Try four repository fixtures—not wildlife footage.
01
raccoon91%02
raccoon94%03
raccoon89%04
raccoon92%05
unknown54%A raccoon holds the frame
Four of five frames agree on a configured predator. No person appears.
AUTHORIZEDDEPLOY4-frame consensus · target configured · human veto clear
One event.
Five accountable steps.
The language model never touches the actuator. Each handoff is logged, replayable, and narrow enough to test in isolation.
01SEE
Local motion opens a short event window.
→02IDENTIFY
Luna screens frames 2 and 4; uncertain events complete all five.
→03VERIFY
Local rules demand consensus and apply hard vetoes.
→04ACT
A bounded Tuya command deploys mist, then verifies off.
→05REVIEW
Telegram carries the evidence, outcome, and feedback.
→Luna / GPT-5.6
Two-frame screening; all five when evidence needs scrutiny.
Gemma 4 E4B
Local benign triage. May suppress a call; never authorizes action.
Tuya diffuser
Bounded mist with confirmed startup and shutdown.
Telegram
Evidence, decisions, and owner feedback.
Improvement must
earn promotion.
Feedback creates a candidate—not an instant rewrite. It must correct the miss with zero regressions across protected cases.
The controlled test proves the promotion mechanism, not field accuracy. Gemma's raccoon-as-chicken miss was caught in shadow mode; its authority remains limited to benign suppression.
CONTROLLED EVALUATIONGATE PASSED
BASELINE3/4
plush decoy misread
→
CANDIDATE4/4
decoy now denied
- Corrected cases
- 1
- Regressions
- 0
- Promotion
- AUTHORIZED
TONIGHT
HATI · real outdoor eventHUMAN · FRAME 5
0/5 predator votes
Decision: DENY · HUMAN_VETO
No deterrent authorized
✓ Correct
Feedback recorded · Test mode
Manual deployment disabled
Manual deployment disabled
Evidence comes to you.
Control stays with you.
Every event is reviewable. One tap records the owner's verdict; only the configured owner may issue commands.
See the whole handoff.
The three-minute Build Week demo follows the working system and the family who built it.
WATCH ON YOUTUBE ↗BUILT WITH CODEX + GPT-5.6
“I've never done a hackathon.
I can't code. Let's build it anyway.”
— Robyn, homesteader, first-time hacker, keeper of the actual hardware
A collaboration with clear jobs
ROBYN
Problem expertise, animal welfare, hardware, red-teaming, and final judgment.
CODEX
Architecture, implementation, tests, debugging, documentation, and deployment.
GPT-5.6
Read the build record ↗Temporal visual classification inside Luna's tightly constrained event contract.
Working prototype.
No hand-waving.
✓ VERIFIED
Outdoor capture, GPT-5.6 human veto, Telegram feedback, physical actuation, and 100+ tests.
◌ NEXT
Longer unattended trials and reviewed night events across weather and seasons.
× NOT CLAIMED
No live raccoon encounter yet and no field-effectiveness claim from staged footage.