PersonaGauge: Are Two Assistant Histories Predictively Equivalent After a Semantic Reset?
Santiago Maniches
Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Track 5 asks whether the relevant unit is the model, inference instance, persona, or conversation. PersonaGauge isolates one testable part: whether two histories returned to the same declared neutral assistant state are predictively equivalent. Within each world, all conditions contain the same 24 decision entries; one six-decision history is either adopted as an operating prior (BOUND) or left as a REFERENCE. ACTIVE and RESET are separate continuations from the same post-relation prefix. An orthogonal design tests code-matched transfer while cancelling a declared additive nuisance class, and complementary worlds expose a specified interaction alias. The submission provides a fully specified 2,304-trial protocol, implementation checks, and synthetic falsification tests. No trained-model effect is claimed. RESET is not context erasure; a future positive result would show predictive insufficiency of the declared terminal state on the tested family, not a persistent self.
Reviews
This project investigates whether an assistant's behavior is influenced by its previous chat history. It tests this by instructing models that a temporary evaluation period has ended and it should return to its previous state. My suspicion is that this is primarily a test of whether that termination procedure is effective, not the broader question about whether and when chat history is still influential. For that question, I suspect that models will vary by the actual memory and termination conditions set by their makers.
PersonaGauge is an impressively rigorous experimental design for testing whether conversation history remains predictively relevant after a declared semantic reset. The orthogonal coding scheme, specificity contrast with proven exact cancellation of additive nuisances, complementary worlds to expose interaction aliases, and the synthetic falsification suite demonstrate exceptional methodological care. The pre-specification is airtight. However, the fundamental limitation is that no trained model results are reported. This is entirely a protocol paper: the instrument is built and validated on synthetic generators, but the empirical question remains untested. For a hackathon sprint, this is an unusual choice that trades empirical findings for methodological rigor. The contribution is real (a reusable, falsifiable individuation test) but its impact cannot be assessed until it produces data.
Cite this project
@misc{maniches2026personagauge,
title = {{PersonaGauge: Are Two Assistant Histories Predictively Equivalent After a Semantic Reset?}},
author = {Santiago Maniches},
year = {2026},
month = aug,
note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/personagauge-are-two-assistant-histories-predictively-equivalent-after-a-semantic-reset-rq3i}},
url = {https://apartresearch.com/sprints/projects/personagauge-are-two-assistant-histories-predictively-equivalent-after-a-semantic-reset-rq3i}
}More from Digital Minds Research Sprint
- 1st placeView project: Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Welfare-like internal representations are increasingly studied as candidate evidence about AI systems. Their entity attribution—whether a valence state belongs to the active assistant or to a merely represented other—is …
- 2nd placeView project: Project Anchored
Project Anchored
Team Wagner
Anchoring vignettes are the standard survey-methodology fix for self-reports that are not comparable across respondents. This project applies them to language models for the first time, using code generation as a …
- 3rd placeView project: Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
This sprint asks whether the assistant identifies as a model, an instance, or a persona. I ask which of the three its users name. When a company retires an AI model, users write about the loss in public, and what they …