Probability Paths & Blind Spots

How isolated agents traverse different regions of probability space, revealing errors that self-review cannot find, and how a debate and a person turn their different answers into one decision. Each diagram plays once; step through it at your own pace.

Published: January 15, 2026 · Updated: September 30, 2026

The single-agent problem. An agent that reviews its own work walks the same path again, so its blind spots survive, and it defends its choices.

Knowledge isolation. Agents exchange documents, not context, so one agent's framing does not carry into the other's review.

Different probability paths. The same task in separate contexts sends agents through different regions of the space; each reaches things the other does not.

The math. Two 90% agents with independent blind spots both miss only 1% of the time. Independence is the assumption: different models help it hold.

Synthesis. A plan's open questions are argued between agents on different models into recorded decisions, and a third agent checks the whole plan for contradictions.

Human arbiter. What a debate cannot settle (priorities, taste, risk) goes to a person, who also signs the plan off.

The Key Insight

Isolation ensures agents take different paths through probability space.
Different paths = Different blind spots = Better combined coverage.
Debate turns the paths into one decision; a person owns the ones debate cannot settle.