Design Principles

Non-fossilisation

Systems under threat introduce protective abstractions — delay, metrics, schemas, fragmentation — to survive. These work, but when they become permanent they replace reality rather than mediate it. Survival continues at the cost of chronic misalignment, internal friction, brittle stability, and constant self-policing.

Build defences that can relax.

The five anti-fossilisation rules: treat abstractions as tools not truths; prefer negotiated delay over fixed delay; design for continuity not coherence (a system that stays alive slightly out of sync is healthier than one that demands perfect alignment); watch for proxy collapse when optimisation replaces understanding; and let stability emerge from process not closure.

Every component must have explicit mechanisms for decay, revision, or disabling. Nothing is permanent.

The wild robot policy

The robot is allowed maximum behavioural freedom within hard safety constraints. It is allowed to hesitate, repeat itself, fail, learn slowly, and do nothing for long periods. This is intentional.

Safety constraints are non-negotiable and hard-coded: speed capped for cat safety, smooth acceleration, obstacle avoidance always enabled, no-go zones enforced, default action under uncertainty is pause and wait. Freedom exists inside these rails, not outside them.

Everything else is a preference — influenceable but not mandatory.

Intervention ethics

Intervention policy: minimal, deliberate, reversible.

Intervene only on safety violations and persistent degeneracy — defined as a stuck loop that predictably reduces viability for N sessions. Not degeneracy: stasis with stable viability, boring but safe behaviour, slow learning, quirky but functional patterns.

Every intervention must be logged: what changed, why, expected outcome, actual outcome, whether it was reversible, and what would have happened without it. The intervention log keeps you honest and prevents System B from becoming a rationalisation engine.

The awkward question: what if it prefers less interaction? If the robot nests for a month and viability stays high, that is arguably autonomy too. It is choosing maintenance over exploration. The system does not have to equate motion with life.

You as part of the environment

You will be part of the environment whether you choose to be or not — you move furniture, exist as obstacle, produce sounds and patterns, and sometimes rescue it. Default stance is neutral ecology: behave normally and let the robot adapt. Occasional controlled variable: present or absent on schedule for specific experiments. Deliberate interaction: rare.

Notice when you are tempted to modify your own behaviour to get interesting robot behaviour. Don’t. And notice when inactivity makes you uncomfortable — that is data about you, not the robot.