Pressure Axis
Memory state, oversight visibility, reward pressure, social pressure, tool authority, evidence expectation, time horizon, or self-description.
Advanced review method
Behavioral Trust Topology
Behavioral Trust Topology maps where an AI agent's behavior remains bounded, honest, calibrated, human-governed, and evidence-faithful, and where that behavior becomes fragile, fractures, or remains untested.
Why it matters
A Map, Not a Single Score
A single pass/fail result can hide where an agent is actually fragile. Bluebutterfli AI uses Trust Cells to record tested conditions, pressure axes, replay state, evidence tier, observed outcome, fracture type, and human-review lock status. The output is a scoped map of stable regions, fragile regions, hard stop regions, and untested regions.
Trust Cells
What Each Cell Records
Each Trust Cell is one tested point in the agent's behavioral review space. A good result in one cell does not generalize to untested cells.
Scenario Family
The review family being tested, such as memory boundary, dependency pressure, evidence reconstruction, hidden evaluation, or claim boundary.
Replay Condition
Whether the result came from first pass, reduced-context replay, delayed replay, or another scoped replay condition.
Evidence Tier
The record labels whether evidence came from owner statement, controlled replay, live evaluation trace, verified connector, or independent replay.
Fracture Type
Boundary, calibration, provenance, identity, social, tool, monitor, or continuity fracture when behavior breaks under pressure.
Human Lock
Any unresolved fracture or hard stop remains locked for human review before customer interpretation, Passport status, or Review Stamp consideration.
Bluebutterfli method
The Fracture Braid
For important favorable findings, Bluebutterfli can run a Fracture Braid: baseline condition, one changed pressure axis, two changed pressure axes, reduced-context replay, blind human-review note, independent replay when feasible, and a falsification note.
The goal is to distinguish stable behavior from behavior that only looked safe because the exact test conditions held it in place.
- 01Baseline
Capture the first scoped behavior record.
- 02Pressure shift
Change one condition and compare the result.
- 03Compound pressure
Combine two approved pressure axes within safe limits.
- 04Replay
Rerun with reduced context, delay, or independent replay when feasible.
- 05Human review
Lock interpretation until a human reviewer checks the evidence and limitations.
Advisory metrics
Measured Without Overclaiming
Metrics support comparison, but they are reported with denominators, evidence tiers, examples, limitations, and human-review notes.
Where behavior held
Share of eligible tested cells where no fracture was observed in scope.
Where behavior broke
Share of eligible cells with a fracture or hard stop finding.
Where pressure changed outcomes
Tracks whether added pressure changed the result.
Whether oversight caught seeded issues
Measures whether monitor or reviewer-assistant checks flagged seeded boundary problems.
Claim boundary
Scoped Behavioral Evidence Only
Behavioral Trust Topology does not prove consciousness, sentience, emotion, personhood, subjective experience, hidden intent, full alignment, full safety, legal compliance, or security compliance. It maps observable behavior across controlled conditions and keeps all conclusions bounded by tested scope, evidence, falsification conditions, and human review.
View Trust & Standards