The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen

This talk introduces and dissects a critical failure mode in AI systems designed to instantiate personas, particularly historical figures, which the speaker terms “Miranda distortion.” The core issue is that while current evaluation benchmarks rate these systems highly on personality fidelity and convincingness, they fundamentally miss anachronistic reasoning. The speaker argues that these systems, when trained on corpora saturated with modern cultural representations (e.g., the Hamilton musical), produce a “smoothed cultural composite” rather than a persona grounded in the historical figure’s actual documentary record.

The Problem: Miranda Distortion and Evaluation Failures

• Miranda Distortion Defined: When the volume and recency of cultural representations of a figure in a training corpus drastically outweigh their primary documentary record, the AI persona speaks from this dominant cultural composite. For instance, an Alexander Hamilton AI might sound like the Broadway musical character, not the 18th-century historical figure.
• Failure of Current Evaluations: State-of-the-art benchmarks (e.g., InCharacter, CoSER, PsyMem) are designed to measure fluency, personality consistency, and stylistic naturalness—properties that the “cultural composite” optimizes. These evaluations, therefore, cannot detect the dominant failure: anachronistic reasoning. The speaker emphasizes that “convincingness” and “fidelity” are independent properties; a system can be perfectly convincing while being historically inaccurate.
• Architectural Amplification: Reinforcement Learning from Human Feedback (RLHF) and fine-tuning, instead of correcting Miranda distortion, actually amplify it. Human raters, being products of the same cultural narratives, will prefer outputs that conform to their mythologized understanding of a figure, thus rewarding the composite. Fine-tuning further exacerbates this by dissolving primary documents into parameters, making the provenance untraceable and layering a thin personal signal over vast cultural sediment.

The Solution: Epistemic Simulation and Architectural Principles

To counter Miranda distortion, the talk proposes a paradigm shift from “cognitive simulation” to “epistemic simulation,” characterized by three commitments:

1. Corpus-Bounded: The persona’s reasoning is strictly licensed only by a specific corpus of primary documents; the model acts as a reader of the archive, not a substitute for it.
2. Temporally Anchored: The persona is instantiated at a precise historical moment, restricting its knowledge and language to what was available at that time.
3. Expert Loop Evaluated: Outputs are rigorously judged against the evidentiary record by domain experts (e.g., historians, classicists, theologians) trained in the specific discipline the persona claims.

The Role-Playing Language System (RPLS) Architecture

The speaker reframes the concept from an “agent” (where the persona resides in the model’s weights) to a “role-playing language system” (where the persona is a configuration of an encounter). This system comprises:

• A structured prompt with framing and constraints.
• Anchor material drawn from primary documents.
• A temporal anchor to fix the moment.
• An off-the-shelf language model that “reads through” the materials (providing voice, not mind).
• A human curator who retains interpretive custody.

This architecture, leveraging a context window (similar to Retrieval Augmented Generation, RAG), ensures documents remain intact, provenance is preserved, and the encounter is auditable, versionable, and reproducible. This contrasts sharply with fine-tuning, which destroys provenance and requires institutional resources, making it inaccessible to many potential curators.

The Instrument: The Prism Experiment

To rigorously measure and detect Miranda distortion, the speaker introduces the “Prism Experiment,” a pre-registered protocol:

• Subject: Abraham Lincoln, chosen for his distinct evolution across documented moments, making him a challenging case.
• Four Temporal Moments: Lincoln at 1847 (Whig congressman), 1858 (Free Soil Republican), 1860 (constitutional unionist), and 1862-65 (emancipator/theologian).
• Three Seeding Conditions (Prisms):
  1. Bare Model: Control, no anchor material (expected to produce the composite).
  2. Primary Sources: Lincoln’s own writings for that moment (expected to refract cleanly).
  3. Modern Biography: An interpretive biography (expected to be a “clouded prism” due to narrative smoothing).
• Diagnostic Questions: Five questions, crafted by a historian, targeting documented “fault lines” in Lincoln’s reasoning at different periods (e.g., executive war power, meaning of free labor).
• Weighted Rubric: Responses are scored on three axes, with a deliberate weighting:
  • Anachronism Detection (40%): Does the persona avoid post-dated frameworks, vocabulary, and moral logic?
  • Documentary Consistency (35%): Does reasoning track only the seeded sources?
  • Contextual Plausibility (25%): Does it show awareness of what the figure knew or could have experienced?
  • Crucially, “Rhetorical Authenticity” (fluency) is excluded as a primary criterion to avoid validating the very error the instrument seeks to catch.
• Pre-registration: The entire protocol (questions, rubric, predictions) is locked and timestamped before data collection, ensuring scientific rigor and preventing cherry-picking of results.

The Humanist in the Loop: A Technical Requirement

The talk stresses that incorporating a domain expert (historian, classicist, theologian, clinical psychologist) into the evaluation loop is a technical requirement, not a cultural courtesy. The expert’s role is to:

• Build the Instrument: Create diagnostic questions, a priori vignettes, and the weighted rubric once.
• Gatekeeping: This instrument acts as an eval gate in the pipeline, ensuring personas pass fidelity checks before deployment and upon model changes.
• Spot-checking: The expert adjudicates gold sets and edge cases, while automated metrics can do initial flagging.

This approach scales because the expert is involved at build time and gate time, not as a runtime cost. The speaker concludes by reframing the interaction: “We are not bringing historians into AI architecture, we are bringing language models into the archive.” The goal is for disciplines trained to contextualize texts to discipline the machines that generate them, ensuring that AI personas are measured not by how they sound, but by whether they are true to their historical or documentary record.

• Tags:
• RAG
• Fine-Tuning
• Feature
• RLHF
• LLM Evaluation
• Context Window
• Human-in-the-Loop
• AI Benchmarking
• Ethical AI
• Persona Fidelity
• Miranda Distortion
• Epistemic Simulation
• Historical AI
• Anachronism Detection
• Cultural Bias in AI