Evidence before assertion
Capability and safety claims should name the evaluation, conditions, limitations, and uncertainty behind them.
Responsibility
Capability and care, developed together.
Responsible development is not a final review. It shapes research questions, data choices, evaluation, access, and the decision to deploy.
Approach /
Lifecycle controls
Evidence /
Evaluation before claims
Status /
Framework in development
01 / Commitments
Principles that change engineering decisions.
These are working commitments for an early-stage lab. They will become more specific as capabilities, collaborators, and deployment contexts develop.
Least capability necessary
Systems should receive only the tools, data, autonomy, and access needed for a defined purpose.
Human agency by design
People need meaningful visibility, consent, intervention, correction, and exit in consequential workflows.
Accountability survives scale
Faster and more autonomous systems require stronger traces, ownership boundaries, and incident response.
Deployment is a decision
A capable model is not automatically suitable for every environment. Context, misuse potential, and reversibility matter.
02 / Lifecycle
Responsibility across the full system.
01
Define
Scope and threat model
Specify intended use, excluded use, affected people, plausible misuse, and required controls before development accelerates.
02
Build
Data and training review
Track provenance, quality, representation, privacy, filtering, and known gaps in training and evaluation data.
03
Test
Layered evaluation
Combine capability tests, adversarial testing, domain review, system-level exercises, and checks under distribution shift.
04
Release
Proportional access
Match access, monitoring, rate limits, human review, and rollback mechanisms to demonstrated risk.
03 / Transparency
Make limits visible.
When LabOne37 publishes a model, system, or material result, the accompanying record should describe intended use, evaluation boundaries, important failure modes, and operational constraints.
We will distinguish research direction from demonstrated result, and prototype from deployable system. This website follows that rule: active work is described without invented releases or benchmarks.
View working notes →