I
O-1 Phenomenon
The Governance Paradox
Why heavy governance can make AI look safer while reasoning worse.
AI Adoption Science
Janus Labs studies how governance affects AI reasoning. We publish AI Adoption Science research, maintain the Janus Protocol, and build evaluation infrastructure for teams that care about reliable agent behavior.
Many AI systems are wrapped in prompt overhead, reporting requirements, and control layers on the assumption that more governance means better outcomes. In practice, that overhead can compete with the reasoning it is supposed to protect.
We call this the Governance Paradox: as governance load rises, reasoning quality can fall. It is an observed phenomenon, not a finished theory, and it sits at the center of the Janus Labs research program.
Four published articles tracing one question — how do you add governance to a reasoning system without degrading the reasoning? Each is tagged by how far its evidence has earned its keep.
Why heavy governance can make AI look safer while reasoning worse.
Why AI models tell you what you want to hear — and why that is dangerous.
Why AI governance should tighten only when performance slips.
Tiered memory and hierarchical summarization that carry trust across sessions without carrying the noise.
Separate roles for generation and critique. The Builder advances the task. The Watcher checks for drift, repetition, and failure without crowding the main reasoning path.
A design principle: governance should stay quiet when the system is healthy and become visible when deviation appears. The goal is higher verifiability with lower overhead.
A lightweight escalation heuristic: N=1 pass, N≥2 warn, N≥3 halt. Useful on its own, and stronger when paired with semantic and confidence signals.
Measure what your governance layers cost in context, latency, and operator effort. Find out whether extra process is improving verifiability or just adding friction.
Move claims from Observed to Validated. The taxonomy provides a shared vocabulary for epistemic status. The gaps in the matrix are the research agenda.
Ask vendors what kind of evidence sits behind their safety claims. The taxonomy gives buyers a way to distinguish observation, hypothesis, and validated results.
One measured piece at a time, with sources attached. No cadence promises, no marketing list — just the next article when it’s ready to defend.
Subscribe on SubstackAlex is a product and technology leader in Sydney and the creator of the Janus Protocol. Janus Labs is his independent venture for making AI agent reliability measurable — one falsifiable claim at a time.