Field devices
Approved fixed or mobile devices report factual events, health and telemetry. A device may observe; it does not decide that a person is guilty or authorise force.
Technology brief · evidence labelled
A local nervous system for farm security.
Scarecrow is not one robot. It is a vendor-neutral edge gateway that receives signals from approved devices, creates a common incident record, keeps humans in control and remains useful when internet access is unreliable.
Reference architecture
Five layers, one accountable workflow.
The architecture is deliberately modular. Cameras, alarms, environmental sensors, quadrupeds and future aerial systems can change without rewriting the evidence and operator layers.
Device adapters
Each vendor integration is isolated behind a narrow adapter. It translates vendor data into the Scarecrow event contract and blocks unsupported commands.
Farm gateway
Local services authenticate events, remove duplicates, correlate signals, create incidents, store evidence and track device health.
Operator console
An authorised human reviews context, marks an event safe, uncertain or threatening, requests a permitted deterrent, or calls a human response partner.
Response network
Farmers, control rooms, farm watches, emergency services and lawful security providers follow agreed procedures and retain their own authority.
Evidence layer
Time-stamped events, device state, operator actions and integrity checks support review, maintenance, investigations and pilot measurement.
Event lifecycle
From signal to reviewed outcome.
The system separates machine observation from human judgement. Confidence scores can prioritise attention, but cannot replace verification.
Observe
A sensor or mobile platform reports motion, thermal activity, tampering, a camera event or device failure.
Normalise
The adapter adds stable identity, time, zone, severity and factual details using a versioned event format.
Correlate
The gateway rejects invalid or repeated messages and opens a review incident for material events.
Decide
A person checks video, schedule and local context, then records a classification and note.
Act and learn
Permitted requests are queued for humans or approved systems. The complete trail supports tuning and audit.
Safety boundary: the current gateway accepts only observation, review and request workflows. It contains no firearm control, no autonomous-force route and no live robot command adapter.
What exists today
The software foundation is real; robot performance is not yet proven.
Authenticated event intake, duplicate protection and local SQLite storage.
Human-review incidents, operator decisions and an audit trail.
Explicit rejection of unknown and autonomous-force actions.
A browser simulator using synthetic events and no real farm data.
Private farm configuration, device heartbeat, battery, signal and degraded-state monitoring.
Local notification queue, audit-chain verification and exportable evidence bundles with checksums.
Live device integration until a platform, protocol and emergency-stop method are selected.
Detection accuracy, patrol reliability, latency, weather tolerance and commercial savings require field evidence.
Try the public simulation. It demonstrates the decision flow, not a deployed security product.
Open the demoPublic evidence interpreter
AI is the explanation layer—not the source of truth.
A small on-server language model can combine and explain reviewed Project Scarecrow records. Retrieval and validation surround the model so it cannot quietly turn an idea into an “implemented” claim.
Hard separation: the public interpreter has no robot-control route, no shell, no private-vault access, no farm data and no internet tools. The website remains complete when the model is stopped.
Only reviewed public records enter the prompt.
The server ranks a versioned public evidence index and supplies only the closest records, each with a fixed source ID and implementation status.
Answers must survive validation.
The model returns structured data. Unknown citations, missing grounding, unsafe requests and malformed output are rejected in favour of deterministic evidence text.
Production load wins.
One inference may run at a time. Input, context, output, queue, memory, request rate and runtime are capped; high server load automatically uses the curated fallback.
Vendor-neutral integration
Hardware earns access through an adapter and a safety case.
The first real integration should be a non-actuating camera, alarm or simple sensor. Mobile robotics comes only after emergency-stop, safe-state and terrain constraints are understood.
| Candidate | What we need from it | Required proof | Status |
|---|---|---|---|
| Fixed alarm or sensor | Authenticated event and heartbeat | Reliable identity, replay protection, loss-of-link alert | Best first integration |
| Network camera | Event metadata and review imagery | Privacy zones, time sync, access controls, lawful retention | Protocol/vendor required |
| Quadruped platform | Pose, health, camera and supervised route execution | Emergency stop, speed limits, no-go zones, charging and recovery | Hardware decision required |
| Aerial platform | Optional observation under approved operations | SACAA pathway, licensed operation, geofencing and weather limits | Outside initial pilot |
| Control-room system | Alert handoff and acknowledgement | Named responsibilities, secure interface, full audit and PSiRA review | Partner requirements needed |
Degraded-condition requirements
Cloud optional. Safe failure mandatory.
The lab gateway proves local intake, local records, device-health state and a local notification queue. The remaining behaviours below are explicit target requirements, not claims of field operation.
| Condition | Current proof | Target requirement |
|---|---|---|
| Internet lost | Local intake and records are implemented and automated-test passed. | External delivery retry is planned; no outbound worker exists yet. |
| Device silent | Heartbeat age produces a visible degraded or offline state in the lab gateway. | Real device loss-of-link behaviour remains blocked on a selected device. |
| Location uncertain | No mobile navigation is connected. | A future mobile unit must slow or stop in a predefined safe area. |
| Sensor conflict | Human review states exist; real multi-sensor correlation is not implemented. | Conflicting inputs must remain uncertain rather than escalating automatically. |
| Power risk | No physical backup-power test has been run. | Orderly shutdown and backup power require hardware and a recorded test. |
| Update failure | No signed-update system exists. | Signed packages, staged rollout and verified rollback are planned before field use. |
Engineering contribution
What we can build before buying a robot.
Lab control plane implemented
Configuration, health, local notification queue, integrity-checked evidence export, protected operator state and ten automated tests. Evidence qualifier: automated-test passed on 15 September 2026.
Sensor bridge
Connect one real non-actuating device and measure reliability, latency, tamper and offline behaviour.
Robot adapter harness
Define a simulator-backed contract for health, pose, route request, acknowledgement, expiry and emergency stop.
Controlled field pilot
Deploy only after site survey, safety review, operator training and agreed acceptance thresholds.