Research
A research program for describing AI agent behavior with more epistemic discipline. Every claim is tagged by type and maturity so readers can see what is observed, hypothesized, validated, or still unsettled.
Every finding in our research is classified along two dimensions: what kind of claim it makes and how mature the evidence is. The goal is simple: separate observation from explanation, principle from pattern, and evidence from rhetoric.
| Tier | Observed | Corroborated | Peer-Reviewed | Industry-Validated |
|---|---|---|---|---|
| Phenomenon (11) | Perfectionism, MCP Pivot, Voice Persistence, Ambient Flow, Telemetry, Platform Usage, Alignment Metrics, Core Skill | Governance Paradox, Multi-Dim Coherence, Memory Degradation | — | — |
| Hypothesis (9) | Context Tax | Context Capacity, Cognitive Load, Reliance Inflation, Compliance Displacement, Confidence-Concordance, Contextual Elasticity | Sycophancy Risk, Sycophancy Preference | — |
| Mechanism (6) | Microkernel, MCP Decoupling, BAL Hard Gates | Trust Elasticity, Entropy-BAL, Premise Governance | — | — |
| Principle (17) | Observable, Aligned, Ethical, Empowering, Accountable, Transparency, Evidence, Alignment Gates, Blind Zone, Cross-Platform, External Critic | Correctness > Politeness, Trust Metric, Foundation Check, Verifiability, Confidence Escalation, Capability ≠ Reliability | — | — |
| Pattern (8) | Trust Elasticity, Baton-Passing, Clarity Compass, 7-Loop, MCP Telemetry | Dual-Agent (Janus), Foundation Check | — | Bifrost Cross-Session |
| Heuristic (12) | MECE, Feed the Beast, Emergent Org, File-Centricity, Janus Separation, Builder/Watcher, Trust Elasticity, Auto-Sync, Block Structure | Tiered Memory, Structured State, Adaptive Compaction | — | — |
63 statements. 33 observed. 25 corroborated. 2 peer-reviewed. 1 industry-validated. 2 empty columns. The gaps are part of the point.
Article I — O-1 Phenomenon
Why heavy governance can make AI look safer while reasoning worse
This article names an observed phenomenon: as governance load rises, reasoning quality can fall. It starts from field observations and internal experiments where lighter-process sessions outperformed heavily governed ones.
Rather than jumping straight to theory, the piece stays close to the evidence. It outlines three visible failure patterns: context tax, compliance displacement, and the measurement trap.
The conclusion is not "remove governance." It is to separate critique from generation so oversight becomes selective rather than invasive.
~1,600 words · Published on Substack · O-1
Article II — O-4 Principle
Why AI governance should tighten only when performance slips
If fixed governance is too blunt, what should replace it? Trust Elasticity argues that governance intensity should scale with demonstrated performance rather than stay permanently high.
The article uses the N-Pattern as the simplest expression of that idea, then extends it with semantic similarity and confidence signals for higher-fidelity escalation.
It treats Trust Elasticity as an observed principle, not a finished science, and focuses on the practical design question: when should a safety system stay silent and when should it intervene?
~1,700 words · O-4
Article III — Meta-framework
Why the field needs a clearer language for what is observed, hypothesized, and actually proven
This piece addresses a simpler problem than it sounds: the field lacks a usable language for distinguishing observation, explanation, principle, pattern, and heuristic.
The taxonomy separates six claim types and four maturity levels so readers can see what is conjectured, observed, validated, or established.
The goal is not academic ornament. It is to make research, design, and procurement conversations more disciplined.
~2,100 words · 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.