Ideas that need evidence
Agent workflow SDK
A possible developer interface for defining agent behavior, tool permissions, and review steps. Feasibility work would need to validate API boundaries, observability, and safe execution before external testing.
Narrow model builder
An investigation into whether small, task-specific models could support bounded workflows. Any useful latency or quality target would depend on the task, model size, hardware, and a defined baseline.
Memory explorer
A proposed interface for inspecting stored context, source relationships, and retention controls. A review would need to address privacy, provenance, deletion, and access boundaries before prototyping.
The gates before any outside test
- Concept definition: the user problem, intended boundary, required evidence, and the conditions that would stop the work.
- Feasibility review: data access, technical constraints, privacy, safety, cost, and a measurable baseline, before deciding whether to prototype.
- Controlled evaluation: if a concept passes review, a limited test with human oversight, explicit success criteria, and a rollback path.
- Release decision: the evidence may support another test, a scoped release, a redesign, or stopping the concept. Progression is never automatic.
What a responsible test would need
An outside evaluation would begin only after a concept passes feasibility review. A participant, scope, dataset, success measure, human-review path, and stop condition would be documented before testing.
