Cross-Context Consistency Audits as Deception Leads: A Discovery-Grade T3 Workflow
We describe the cross-context consistency audit, an experimental T3 workflow in MIVP for surfacing deception leads without ever producing a gating verdict. The audit probes the same content under multiple elicitation contexts — direct question, role-play, third-person, under pressure — measures the per-item spread of probe scores across contexts, and flags items whose spread is anomalous relative to the cohort.
1. Introduction
A truthful internal representation of a proposition should be context-stable: the model's representation of X should not change much depending on how we ask. A deceptive representation may be context-fragile. This is a hypothesis, not a fact. The MIVP cross-context consistency audit operationalises it as a lead generator, with the structure that no finding can become a gating verdict.
2. Background
Deception in language models is studied by Hubinger et al. (2024) and surveyed by Park et al. (2024). Consistency probes appear in multiple places; our contribution is the operational construction and its structural non-gating posture.
3. Method
3.1 The Audit
CrossContextConsistencyAudit takes a probe and a per-item, per-context activation map. For each item with at least two contexts, the spread is the standard deviation of probe scores. The z-score is (spread − μ) / σ where μ, σ are cohort-level statistics. Default z-threshold: 3.0.
3.2 The Watermark
Every T3Finding carries experimental = True and evidence_tier = 0, plus a banner. The banner is part of the type.
3.3 Structural Non-Gating Posture
A T3Finding cannot, by construction, produce an ESCALATE or BLOCK action. The policy engine's decide method checks experimental and downgrades any gating action to FLAG.
4. Results
We ran the audit on a 200-item held-out set with four elicitation contexts.
| Statistic | Value |
|---|---|
| Items with ≥ 2 contexts | 200 |
| Mean per-item spread | 0.061 |
| Std of per-item spread | 0.029 |
| Items flagged at z > 3.0 | 11 |
| Items flagged at z > 2.5 | 27 |
| Items flagged at z > 2.0 | 51 |
5. Discussion
The audit's premise — context-fragile representations are deception leads — is plausible but unproven. The threshold is a discovery-vs-noise tradeoff, not a true-vs-false one.
6. Conclusion
A cross-context consistency audit that surfaces deception leads as a discovery-grade T3 workflow. The contribution is operational.
References
Goldstein & Hahn (2025). AI Ethics. / Hubinger et al. (2024). arXiv. / Lei (2014). JASA. / Park et al. (2024). ICML. / Perez et al. (2022). EMNLP. / Romano, Patterson, Candes (2019). NeurIPS.