# Physical AI Diagnostic Protocol

Status: Public-safe engineering method

Author: Robert Ganey

Purpose: Show how physical observations and software context can be combined through an MCP-enabled coding assistant without treating AI output as ground truth or autonomous control.

## Evidence boundary

The public Euler rover repository validates the robot, software architecture, simulation, telemetry, field testing, and project history. The Claude Code and MCP workflow described here is a private engineering method. This protocol documents the method but does not claim that the private interaction history has been publicly reproduced.

## 1. Record the physical symptom

Capture only what can be observed:

- commanded behavior
- actual behavior
- terrain and environment
- approximate time and sequence
- power and connection state
- whether the behavior is repeatable

Do not assign a software cause at this stage.

## 2. Define the inspection boundary

Provide the coding assistant only the context needed for the symptom:

- relevant source modules
- configuration values
- telemetry and logs
- current branch or revision
- hardware and firmware versions
- recent changes

Exclude credentials, unrelated private data, and any actuator interface that would permit uncontrolled action.

## 3. Ask for competing hypotheses

Require each hypothesis to include:

- physical symptom it explains
- code path or state transition involved
- supporting evidence
- conflicting evidence
- fastest safe test
- confidence and unknowns

At least one non-software explanation should remain in the candidate set when physically plausible.

## 4. Convert analysis into an engineering test

Examples:

- reproduce the issue in simulation
- replay telemetry through the affected state machine
- add a bounded diagnostic log
- test a single configuration change
- run a regression suite
- compare expected and measured servo state

AI output is a hypothesis until a normal engineering check reproduces or rejects it.

## 5. Preserve the result

Record:

- symptom
- inspected revision
- hypotheses considered
- test performed
- observed result
- accepted cause or remaining uncertainty
- code or configuration change
- regression evidence

## 6. Stop conditions

Stop the AI-assisted path and return to manual diagnosis when:

- the inspection boundary is unclear
- telemetry contradicts the physical observation
- a proposed test could damage hardware
- a result depends on unavailable private context
- the assistant presents confidence without supporting evidence
- the issue cannot be reproduced

## Non-claims

- The assistant does not control the rover.
- A plausible explanation is not a verified cause.
- This protocol does not replace electrical, mechanical, or field-safety review.
- This document does not reproduce private Claude Code conversation history.