Evidence before verdicts.
A conclusion is only as trustworthy as the evidence beneath it. Preserve what was observed, where it came from, and what remains unknown.
A working philosophyv0.1 / evolving
Principles
Eight propositions for building autonomous systems whose consequential decisions can deserve trust.
Authority must have a boundary.
Autonomy is not unlimited permission. An agent should know which decisions it owns, which require escalation, and which are mechanically prohibited.
Uncertainty is information.
A system that hides uncertainty is not more decisive; it is less honest. Material ambiguity belongs in the decision, not in a footnote.
Enforce invariants; do not merely prompt for them.
Prompts express intent. Tests, permissions, policies, and runtime boundaries protect the conditions that must always hold.
Constrain safety, not strategy.
Give agents freedom to find better approaches inside a small set of clear, durable boundaries.
Outcomes are the final evaluator.
A plausible decision is not necessarily a good decision. Reconcile recommendations with what production later revealed.
Disclose context progressively.
The right context at the right moment beats a giant instruction manual. Navigation and retrieval are part of the harness.
Memory should accumulate judgment.
Useful memory is not a transcript. It captures precedents, exceptions, outcomes, and the reasons a future decision should differ.
These principles are deliberately unfinished. They should change when production evidence, stronger arguments, or better systems prove them incomplete.