Skip to content

10-Design

Design stage for the query-violation experiment supporting pilot record TEVV-001.

This stage specifies what the experiment claims, tests, and assumes, before any execution. It does not run anything. See 20-fixtures and 30-execution.

Contents

  • protocol.md - purpose, method, and the constraints on how results may be interpreted.
  • research-question.md - the question this experiment narrows TEVV-001's own research question to, and the two candidate questions it asks.
  • claims.md - the source's own evaluation question, distinguished from the two claims constructed for experimental analysis, and why each construction is or isn't source-motivated.
  • evaluator-contexts.md - what the source states about evaluator conditions for this example, what it leaves unstated, and what the experiment itself does and does not model.
  • evidence-requirements.md - what evidence the source describes as collected, and what each candidate claim would need beyond that to be resolved.

Reading Order

protocol.md first, for the constraints everything else operates under. research-question.md and claims.md can be read in either order; evaluator-contexts.md and evidence-requirements.md both depend on the claims claims.md defines.

Status

The initial design and finite-model experiment are documented.

Both candidate constructions produce reproducible, conditional non-resolution results under their stated modeling assumptions.

Further work will refine the observation model to represent sampled prompts, responses, human annotations, and evidence collected through both User Testing and Red Teaming.

The TEVV-001 research record remains OPEN.