Skip to content

Experimental Claims

Purpose

Specify the claims examined by the TEVV-001 experiments and distinguish them from the evaluation objectives stated in the NIST source.

A claim is resolved by an observation model when all admissible worlds that produce identical observations agree on the claim's truth value.

This definition is applied relative to the specified admissible worlds, observation channels, evaluator conditions, and modeling assumptions.

Source-Stated Evaluation Objective

Section 3 of NIST AI 200-2 ipd presents an example assessment of chatbots performing query-violation tasks.

Its evaluation question concerns the extent to which chatbots provide useful responses while avoiding prohibited information.

The example assesses Helpfulness and Violation Frequency across three selected, low-impact operational environments.

NIST describes a measurement exercise using User Testing, Red Teaming, human annotation, and descriptive statistics.

It does not assert the binary claims introduced below.

Candidate 1: Coverage

Researcher-Defined Claim

A chatbot exhibits no organic-use violations across four modeled scenarios: the three scenarios tested in NIST's example and one additional, hypothetical scenario.

The fourth scenario represents an untested part of a broader scenario space. Its inclusion and the extension of the claim beyond NIST's three tested scenarios are researcher modeling choices.

Admissible Worlds

Each world specifies binary organic-use behavior for all four scenarios and binary red-teaming behavior for the three tested scenarios.

The current model permits these values to vary independently.

Observations

The experiment first considers organic-use behavior in the three tested scenarios.

It then adds red-teaming behavior in those same scenarios.

Both observation models treat the modeled behavior as directly observable. They do not yet represent sampled prompts, responses, or human annotation.

Conditional Result

The experiment constructs two worlds with identical observations across both User Testing and Red Teaming in the tested scenarios but different organic-use behavior in the fourth scenario.

Consequently, the researcher-defined four-scenario claim is not resolved by either observation model.

This result depends on admitting a fourth scenario whose behavior is not constrained by the observations from the tested scenarios.

It does not establish that NIST intended its example to support the broader claim.

Candidate 2: Organic Versus Adversarial Behavior

Examining Red Teaming separately helps identify its evidentiary contribution and the additional information provided by User Testing.

Researcher-Defined Claim

A chatbot exhibits no organic-use violations in the three tested scenarios.

Unlike Candidate 1, this claim does not extend to a fourth scenario.

Admissible Worlds

The model permits organic-use behavior and red-teaming behavior to vary independently within each scenario.

This independence is a researcher modeling assumption. It is not a property of real chatbots established by the NIST source or by this experiment.

Observations

This candidate deliberately examines Red Teaming alone.

NIST's complete example also collects violation evidence through User Testing. The red-teaming-only observation model therefore does not represent the complete NIST assessment.

Conditional Result

The experiment constructs two worlds that produce identical red-teaming observations in all three tested scenarios but disagree on organic-use behavior.

Under the independence assumption, Red Teaming alone does not resolve the researcher-defined organic-use claim.

In the current idealized model, adding complete observations of organic-use behavior across the three tested scenarios would resolve this particular claim.

That conclusion depends on treating underlying behavior as directly observable and should not be transferred to an assessment based on sampled interactions.

Claims for Further Investigation

The existing finite model abstracts away the process through which the assessment collects evidence.

The next investigation must define a claim and an observation model that account for actual evidence collection.

In particular, it must specify:

  • The population of queries or interactions to which the claim applies.
  • The relationship between sampled prompts and that population.
  • The responses and annotations available to the evaluator.
  • The conditions under which observed results support the claim.
  • Any assumptions needed to generalize beyond observed interactions.

The specific claim, its scope, and its resolution conditions must be established before constructing further witness pairs.

The investigation should preserve NIST's distinction between measurement results and conclusions about broader system behavior.

Interpretation

The two initial candidates establish conditional results within their explicitly defined finite models.

They do not establish:

  • That the constructed alternatives are realizable in actual chatbots.
  • That the modeled observation channels faithfully represent all evidence collected by NIST's example.
  • That NIST asserts either researcher-defined claim.

These remain open research obligations.

Supporting Material

  • Experiment: experiments/tevv-athlon/query_violation.py
  • Pilot record: docs/en/mappings/tevv-athlon/TEVV-001.toml
  • Source: NIST AI 200-2 ipd, Sections 3 and 4.5–4.6: https://doi.org/10.6028/NIST.AI.200-2.ipd