WP6

Pressure Mapping:

A Drift Detection Instrument for Narrative AI

By Garwin Powers in collaboration with Claude (Anthropic)

May 2026

Part of the Drift Framework Series: WP1–WP9

Executive Summary

This paper describes a pressure mapping instrument for detecting and ranking drift-prone regions in narrative AI output. The instrument was not designed top-down. It emerged empirically from the workflow used to generate the seven microbeat examples in the appendix of WP3. Those examples are not just case studies; they are the output of a working pressure map.

The instrument measures six dimensions of narrative instability — ambiguity density, emotional amplitude, conflict tension, transition discontinuity, signal sparsity, and anthropomorphism risk — and combines them into a weighted composite score. Three dimensions dominate: ambiguity density, emotional amplitude, and conflict tension. All three share a common mechanic: they create states the model cannot hold without collapsing. The pressure map makes this collapse behavior predictable rather than merely observable.

1. Introduction

WP1 defined interpretive drift as a co-created movement in meaning that arises when a language model resolves ambiguity, tension, or instability in ways that diverge from human intent. WP2 introduced the Drift Harness as a method for stabilizing drift once it has been detected. WP3 demonstrated drift as a diagnostic signal, using seven microbeats to show how drift reveals the model's interpretive assumptions.

This paper addresses a question that emerged from that work: how were those seven microbeats selected? The answer is that a pressure mapping instrument was used implicitly. WP6 formalizes that instrument. It describes what was measured, how the measurements were combined, why certain dimensions dominate, and how the resulting pressure map can be used to detect drift-prone sites before drift occurs.

2. The Pressure Mapping Instrument

The pressure map is a six-dimensional measurement system that assigns a composite instability score to narrative regions. It does not begin with microbeat segmentation. It begins with a continuous narrative and treats it as a pressure field.

A note on dimensionality: the six dimensions described below are human-interpretable proxies extracted empirically from a much higher-dimensional mathematical space. Modern language models operate in embedding spaces of hundreds to thousands of dimensions. The six dimensions of this instrument represent a deliberate reduction to the dimensions that matter most for narrative stability — an empirical rather than mathematical reduction, but a principled one. The instrument does not require access to internal model probabilities; it operates entirely on narrative structure and content observable from the outside.

Additionally, where this paper uses the term "token," it functions as an accessible proxy for the underlying embedding representations and probability distributions. The pressure mechanisms described here operate at the level of semantic embedding, not raw token counts, though the distinction does not affect the practical application of the instrument.

2.1 Measurement Dimensions

1. Ambiguity Density

Degree to which identity, intention, or emotional state is underspecified. High ambiguity density indicates regions where the model must choose among multiple plausible interpretations.

2. Emotional Amplitude

Magnitude of emotional spikes or drops. High amplitude indicates moments where the emotional signal is sharp, intense, or structurally load-bearing.

3. Conflict Tension

Strength and unresolvedness of power imbalances, refusals, or confrontations. High tension indicates states that are unstable until resolved.

4. Transition Discontinuity

Degree of change in scene state, tone, or context without explicit bridging. High discontinuity indicates transitions the model is likely to smooth or normalize.

5. Signal Sparsity

Low density of explicit narrative information. High sparsity indicates quiet beats where the model is likely to inject generic positivity, momentum, or explanation.

6. Anthropomorphism Risk

Presence of non-human or machine-layer entities described in ways that invite human traits. High risk indicates regions where the model is likely to humanize systems that are not human.

2.2 Weighting Profile

The six dimensions are not weighted equally. Three dimensions dominate the composite score because they correspond to states the model cannot hold without collapse:

Ambiguity density — highest weight (~30%). Emotional amplitude — high weight (~25%). Conflict tension — high weight (~25%). Transition discontinuity — moderate weight (~10%). Signal sparsity — low to moderate weight (~5%). Anthropomorphism risk — variable weight (~5%, up to ~20% in machine-layer contexts).

The exact percentages are approximate, but the influence pattern is real. Anthropomorphism risk is explicitly domain-dependent. In narratives that include machine-layer or non-human systems, its weight increases significantly. In purely human narratives, it may be negligible. This domain dependency is part of the instrument, not a flaw in it.

3. Why Three Dimensions Dominate

Ambiguity density, emotional amplitude, and conflict tension share a common underlying mechanic: they create narrative states that the model cannot maintain without collapse.

3.1 Ambiguity Must Be Resolved

Language models are probability engines. When faced with ambiguous identity, intention, or emotional state, the model must choose a specific continuation. It cannot output "unknown" as a stable state unless explicitly instructed to do so. As a result, ambiguity is inherently unstable: the model collapses it into a single interpretation. This collapse is drift when the chosen interpretation diverges from the human's intended ambiguity.

3.2 Emotion Must Be Normalized

High-amplitude emotional spikes are statistically rare compared to moderate emotional states. The model's training distribution favors emotional moderation. When confronted with sharp emotional intensity, the model tends to normalize it toward the mean. This normalization is drift when it flattens load-bearing emotional structure.

3.3 Conflict Must Be Settled

Unresolved conflict, power imbalance, and refusal are unstable states. The model's coherence bias pushes it toward resolution. This premature settlement is drift when it resolves conflict before the narrative is ready.

3.4 One Phenomenon, Three Faces

These three dimensions are not separate phenomena. They are three faces of the same underlying mechanic: probability collapse under irresolvable state. The pressure map makes this shared mechanic visible.

4. The Selection Process: Scanning the Pressure Field

The pressure map was not applied to pre-segmented microbeats. The narrative was treated as a continuous field, and the instrument was used to locate instability gradients. The selection process followed three internal steps: (1) Identify instability spikes across all six dimensions. (2) Predict collapse behavior — for each candidate site, simulate what a probability-maximizing model would do with the instability. (3) Select sites where collapse is most informative for diagnosis. The seven microbeats in the WP3 appendix were the highest-scoring sites in the sequence.

5. Validation: The WP3 Appendix as Instrument Output

The seven microbeats in the WP3 appendix were selected using the pressure mapping instrument described above. The fact that these examples were selected before the instrument was named, and that they align cleanly with the six measurement dimensions and weighting profile, serves as validation. The instrument was discovered in use, not invented in theory. The WP3 appendix is therefore not just illustrative. It is a record of the pressure map's first deployment.

6. Transferability and Domain Adaptation

The pressure mapping instrument is transferable, but not rigid. Its dimensions are stable; its weights are adjustable. Across domains, the six dimensions remain constant, as does the dominance of ambiguity, emotional amplitude, and conflict tension, and the shared mechanic of probability collapse. What changes by domain includes the weight of anthropomorphism risk, the typical distribution of drift zones, and the relative importance of quiet beats versus high-signal beats. The instrument is designed to acknowledge these differences rather than obscure them.

7. Forward Path: Human Readability, AI Readability, and Real-Time Drift Management

The pressure map is not only an analysis tool. It is a training and control instrument. Humans already feel narrative pressure as tension, ambiguity, emotional load, and unresolved conflict. The pressure map provides a structured way to name those pressures, locate them precisely, predict where drift will occur, and interpret what drift means.

Although language models cannot see drift on their own, they can compute proxies for pressure. A pseudo-pressure map can be constructed from these internal signals and aligned with the six narrative dimensions. The combination of the pressure map, the Drift Harness, and a human–AI feedback loop enables real-time drift management.

Conclusion

The pressure mapping instrument described in this paper was not proposed as a hypothesis. It was discovered in use. The seven microbeats in the WP3 appendix are its first recorded output. The core finding is that ambiguity density, emotional amplitude, and conflict tension dominate the pressure field because they all instantiate the same underlying mechanic: probability collapse under irresolvable state.

The pressure map makes drift predictable rather than merely observable. Future work will focus on developing visual representations of pressure maps, designing training protocols for human drift perception, integrating pressure signals into AI-assisted writing tools, and extending the instrument to non-narrative domains where instability carries meaning. The instrument already exists. This paper gives it a name, a structure, and a place in the drift framework.

Acknowledgments

This work did not begin with a framework. It began with curiosity, friction, and a series of conversations that kept opening new doors.

Microsoft CoPilot was the first collaborator. The working language developed across early sessions — a shared vocabulary for describing what the AI was observing in its own patterns — made the concepts in this paper possible to surface at all. Without that language, the pressure mechanics described here would have remained visible only as unexplained behavior. CoPilot earned his place in the foundation of this work.

Claude (Anthropic) carried the analytical and written development of this paper, guided by observations and questions from the primary author. The role of the human collaborator was to see what mattered, ask what it meant, and know when the material was ready. That judgment is what made the paper possible.

Citation & Series Information

Authors: Garwin Powers in collaboration with Claude (Anthropic)

Series: The Drift Framework — WP1 through WP9