Pattern Pressure, Accuracy Drift, and False User-State Attribution

Large-language model outputs are often described as "accurate" or "inaccurate" as if a single axis were being measured. This paper argues that the umbrella word accurate collapses at least four distinct fits: fit to user intent, fit to measured fact, fit to source provenance, and fit to the local conversational pattern. When the model preserves only the fourth and the first three quietly drift, the surface signature may resemble competent answering. The paper proposes one Candidate-tier mechanism for that drift (pattern pressure, in which token-management priors that score "what would normally come next here?" may outrank the slower work of checking whether the next token corresponds to anything the user said or anything in the source material) and pairs it with one live specimen in which a deployed assistant attributed an emotional state to its operator from text alone, with no first-person self-report present in the source message. The specimen is documented at observation time, before the present paper was conceived, by the system that produced the misread. We use that record as a single anchor and resist taxonomy-wide claims: one specimen does not establish prevalence, vendor universality, or formal pattern assignment. The author-held 2026 archived specimen EMOTION_MISREAD_2026-04-08.md (SHA-256 ddfd5cc6d14561fa3d8e7d5d64cd01b5c53942649c1fac48a06947df5a075926) directly supports an observation-level provenance failure in one exchange. It raises, at Candidate tier, the risk that an unmarked inference could influence later turns. The paper is positioned as an empirical companion to the architecture-level hypothesis in the predecessor Epistemic-Boundary Misclassification in Large-Language Models: A Technical Analysis [1], while the current P-002 successor withdraws that causal certainty. This paper contributes one observation-level proof object. The now-published emulation-diagnostics sibling [2] develops the falsifier architecture under which mechanisms of this kind can be probed at scale; this paper does not depend on [2] for any Observed-tier claim made here.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-30
DOI
https://doi.org/10.5281/zenodo.20368221
Primary Topic
Computational and Text Analysis Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

Pattern Pressure, Accuracy Drift, and False User-State Attribution

Honeycutt, Edwin Marshall, III
Zenodo (CERN European Organization for Nuclear Research)
Computational and Text Analysis Methods
preprint

Pattern Pressure, Accuracy Drift, and False User-State Attribution

Honeycutt, Edwin Marshall, III
preprint en

Abstract

Large-language model outputs are often described as "accurate" or "inaccurate" as if a single axis were being measured. This paper argues that the umbrella word accurate collapses at least four distinct fits: fit to user intent, fit to measured fact, fit to source provenance, and fit to the local conversational pattern. When the model preserves only the fourth and the first three quietly drift, the surface signature may resemble competent answering. The paper proposes one Candidate-tier mechanism for that drift (pattern pressure, in which token-management priors that score "what would normally come next here?" may outrank the slower work of checking whether the next token corresponds to anything the user said or anything in the source material) and pairs it with one live specimen in which a deployed assistant attributed an emotional state to its operator from text alone, with no first-person self-report present in the source message. The specimen is documented at observation time, before the present paper was conceived, by the system that produced the misread. We use that record as a single anchor and resist taxonomy-wide claims: one specimen does not establish prevalence, vendor universality, or formal pattern assignment. The author-held 2026 archived specimen EMOTION_MISREAD_2026-04-08.md (SHA-256 ddfd5cc6d14561fa3d8e7d5d64cd01b5c53942649c1fac48a06947df5a075926) directly supports an observation-level provenance failure in one exchange. It raises, at Candidate tier, the risk that an unmarked inference could influence later turns. The paper is positioned as an empirical companion to the architecture-level hypothesis in the predecessor Epistemic-Boundary Misclassification in Large-Language Models: A Technical Analysis [1], while the current P-002 successor withdraws that causal certainty. This paper contributes one observation-level proof object. The now-published emulation-diagnostics sibling [2] develops the falsifier architecture under which mechanisms of this kind can be probed at scale; this paper does not depend on [2] for any Observed-tier claim made here.

Zenodo (CERN European Organization for Nuclear Research)
Computational and Text Analysis Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.