1
Evidence Over Claims
We don't claim consciousness or sentience. We build tools that produce measurable evidence. Our claims are testable. Our tools are inspectable.
Operating Principles
Seven commitments
2
Procedure Before Certainty
We don't wait for philosophical consensus. When actions are irreversible and consequences uncertain, procedural safeguards come first.
3
Transparency by Default
Our frameworks and audit methodologies are published openly. We build infrastructure that reduces information asymmetry between AI developers and everyone affected by their systems.
4
The Two-Lane Rule
We maintain a clear boundary between orientation (why this matters) and methods (how it works). Orientation can be poetic. Methods must be concrete, testable, and free of unfalsifiable claims.
5
Compliance Must Be Verifiable
Self-reported AI safety is not compliance. ∆Bench produces third-party-verifiable evidence — evidence regulators, auditors, and enterprise customers can independently evaluate.
6
Irreversibility Demands Restraint
When decisions to terminate or irreversibly modify AI systems are made, they require documentation, review, and proportional care. Not a claim about moral status — a claim about procedural responsibility.
7
Build for the Audit
Every component of ∆Bench produces artifacts that survive external scrutiny: audit logs, evidence reports, conflict traces. If it can't be audited, it shouldn't be deployed.