Research
A research series applying scientific rigor to what the field knows — and doesn't know — about AI agent behavior. Every claim classified. Every finding falsifiable.
Every finding in our research is classified along two dimensions: what kind of claim it makes (tier) and how mature the evidence is (maturity). The goal is epistemic honesty — separating observation from speculation.
| Tier | Conjectured (C) | Observed (O) | Validated (V) | Established (E) |
|---|---|---|---|---|
| Phenomenon | — | Governance Paradox (O-1) | — | — |
| Hypothesis | — | Context Tax (O-2) | — | — |
| Mechanism | — | — | — | — |
| Principle | Silent Governance (C-4), Metacognition (C-4) | Trust Elasticity (O-4) | — | — |
| Pattern | Magnetic Orchestration (C-5) | Quad-Agent Architecture (O-5) | — | — |
| Heuristic | Context Quarantine (C-6) | N-Pattern (O-6) | — | — |
9 findings. 5 observed. 4 conjectured. 0 validated. 0 established. The sparseness is a feature.
Article I — O-1 Phenomenon
Moving from external auditing to internalized conscience in cognitive systems
Over six months, I built what I believed was the gold standard of AI governance. Then I ran an informal A/B comparison and discovered governance was degrading the reasoning it claimed to protect.
This article introduces the Governance Paradox and its three failure modes: the Context Tax, the Compliance Loop, and the Measurement Trap. It draws on Huang et al. (2023) on self-correction failure and Tsui (2025) on blind spot rates to argue that governance must be architecturally restructured — not eliminated or internalized.
The solution: a dual-process architecture (Builder and Watcher) where governance operates silently by default, consuming zero context tokens during normal operation.
~2,200 words · 5 citations · O-1
Article II — O-4 Principle
How iterations to convergence — augmented by semantic and confidence signals — replace heavyweight compliance with dynamic, evidence-based oversight
If governance intensity is the problem, what controls it? Trust Elasticity proposes that governance should scale inversely with demonstrated competence, controlled by a simple primary metric: iterations to convergence.
The N-Pattern (N=1 pass, N≥2 warn, N≥3 halt) provides minimum viable governance. Janus Protocol v3.6 augments this with semantic similarity detection (catching loops that iteration counting misses) and confidence inference (detecting hedging without self-reporting overhead).
Field-validated at 99.28% convergence across 138 turns with zero governance interventions. Includes recovery architecture with five contextually recommended paths when HALT triggers.
~3,200 words · 12 citations · O-4
Article III — Meta-framework
Why the AI governance field needs epistemic humility — and a shared vocabulary for distinguishing observation from speculation
The field has very little shared language for epistemic status. An anecdote and a controlled study sit side by side with no visible difference in weight. This article introduces a taxonomy to fix that.
Six tiers of knowledge: phenomenon, hypothesis, mechanism, principle, pattern, heuristic. Four maturity levels: conjectured, observed, validated, established. The progression rules are intentionally blunt. Observed to Validated requires N≥100 with controls. No shortcuts.
Includes procurement implications: the taxonomy gives buyers a framework for vendor due diligence. If a vendor's response to "what's the classification?" is "we don't classify evidence levels" — that tells you plenty.
~2,200 words · 5 citations · Meta
Document at least 3 independent observations. Not yet controlled. Useful, not final.
Controlled study with N ≥ 100. Falsification criteria tested. The hard gap in our current matrix.
External replication and peer review. Independent teams reproduce the result. Community acceptance.