Ethics statement

This study analyzes the ANEST Narrative–Affect Dataset (ANAD v1.1.0), which was derived from publicly accessible online narratives and processed to remove or mask personally identifying information (PII) prior to release. The Zenodo v1.1.0 release contains derived features only and explicitly excludes raw post text and user-level metadata, thereby reducing traceability and re-identification risk. The present work involves no interaction or intervention with human participants and analyzes de-identified text records. Ethical considerations therefore focus on privacy preservation and minimization of re-identification risk. We report the dataset provenance, filtering, and de-identification procedures as provided in the dataset release [34]. We do not reproduce verbatim narrative text in either the main text or the Supporting Information. All reporting is based on released derived features and aggregate summaries, consistent with the dataset’s privacy-preserving design.

Human dataset: ANAD v1.1.0 (ANEST Narrative–Affect Dataset)

Human narratives were drawn from the ANEST Narrative–Affect Dataset (ANAD v1.1.0), a curated collection of 351,734 English-language relationship narratives collected from publicly accessible online forums (2012–2023) and released as derived features only [34]. The release excludes raw post text and user-level metadata and is processed to reduce re-identification risk (see Ethics statement and the dataset record).

This study uses the dataset-provided precomputed variables: narrative complexity (LoC) and linguistically inferred affective intensity (sentiment_norm), both scaled to a common 0–10 range in the dataset pipeline [34]. The dataset also provides an absolute discrepancy feature (NADI_abs); however, for regime and geometric analyses we define signed discrepancy as

Throughout, N and A are treated as fixed released measurements, and all downstream steps (clipping, regime labeling, density estimation, and geometric occupancy) are applied identically to humans and the language model.

We operationalize signed discrepancy for analysis as

which distinguishes understatement states (D < 0; expressed affect exceeds narrative structure) from overstatement states (D > 0; narrative structure exceeds affect). For descriptive reporting, we summarize discrepancy magnitude using |D|; regime and comparative analyses use signed D. Because D is a deterministic linear function of N and A, the triplet (N, A, D) lies on a plane in three-dimensional space. The effective dimensionality of the expressive geometry is therefore two. All geometric measures (convex hull area, occupancy entropy) are computed in the clipped plane, with D treated as a derived coordinate for regime labeling and interpretive purposes.

LLM trajectories

Model trajectories consist of 1,000 single-turn outputs generated by an RLHF-aligned large language model (OpenAI gpt-4.1-mini; temperature 0.9, top-p 1.0, max tokens 600) [35–37] under a fixed decoding configuration. Outputs were scored using the same feature extractors and scaling pipeline applied to the human corpus, producing N, A, and D on the same scale.

To support replication despite provider versioning, the provider/model identifier (or closest available tag), generation date range, the fixed prompt template (llm_prompt_set.txt), and decoding parameters (including any nucleus sampling, maximum tokens, and stopping criteria) are documented in Supporting Information (S7 Text) and the released configuration file llm_config.json. LLM trajectories are stored as nadi_llm_trajectories.parquet, and we restrict analysis to the first turn (t = 0) for comparability with single-narrative human records.

Preprocessing and clipping

To reduce the influence of extreme outliers while preserving distributional structure, we compute the 1st and 99th percentiles for each primary dimension (N, A) from the human corpus and apply axis-wise clipping to both humans and the LLM:

Signed discrepancy in the clipped space is then computed as . All density maps, regime assignments, and geometric occupancy measures are computed on the clipped variables.

Crucially, the same human-derived clipping bounds are applied to both humans and the LLM; area differences therefore reflect comparative occupancy under a shared bounded geometry rather than agent-specific preprocessing.

Density estimation and correlation analysis

Two-dimensional density maps are provided in Supporting Information (Fig S1 in S2 Text) and are computed using histogram-based estimators on the clipped human data. For each pair , we construct a grid, count observations per cell, and export triplets of cell centers and counts as: fig1_density_NA.dat, fig1_density_ND.dat, and fig1_density_AD.dat. Visualization uses logarithmic scaling () to preserve dynamic range.

Pearson correlations (S4 Table in S4 Text) are computed over the clipped human data and exported to table2_corr_matrix.dat. Because D is defined as prior to clipping, the A–D association is structurally constrained; we therefore interpret correlations primarily in terms of remaining degrees of freedom, with particular emphasis on the empirical near-independence of N from A and D.

Regime definitions and group comparisons

We operationalize discrepancy regimes in two complementary ways.

Four-regime labeling (Fig 1).

We define four expressive regimes using empirical quantile thresholds estimated from the full human corpus on the bounded 0–10 scale. Let denote the 75th percentile of |D|, the 75th percentile of A, and the 25th percentile of N. We assign: (i) Collapse if , , and ; (ii) Understatement if and (excluding Collapse); (iii) Overstatement if (excluding Collapse); and (iv) Coupled otherwise, i.e., the complement of extreme discrepancy regimes. Regime labels are computed for the full corpus and exported as fig2_regimes.dat. Fig 1 visualizes a stratified subset (fig2_regimes_sampled.dat) with a fixed point cap for LaTeX compilation stability; all prevalence and summary statistics are computed on the full corpus.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. A stratified visualization sample of human narratives in the NCS (n = 200 for readability).

Regime definitions and prevalence are computed on the full corpus and reported in Table 3 and Methods.

https://doi.org/10.1371/journal.pone.0348715.g001

Regime boundaries are quantile-based and data-driven. Sensitivity analyses (S1 Text) confirm that qualitative regime structure and contraction findings are stable under percentile threshold shifts. Coupled expression is defined as the complement of the three extreme regimes and is not posited as a theoretically privileged equilibrium.

We use the term “collapse” rather than “compression” because the pattern involves not merely reduced expressive range but a convergence of narratives toward the lower structural boundary of the space while affective intensity remains high, a qualitative shift distinct from proportional scaling.

Signed-quartile bands (Table 2).

To compare humans and the LLM with a single, shared partition, we compute the 25th and 75th percentiles () of the signed human discrepancy D and define: Understatement (D < Q₁), Intermediate (), and Overstatement (D > Q₃). The same cutoffs are applied to the LLM to compute group-conditional means. These results are exported as table3_group_compare.dat.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Group-conditional means for humans and the aligned model (shared human cutoffs).

https://doi.org/10.1371/journal.pone.0348715.t002

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Regime-level summary statistics and prevalence in the NCS (rendered from exported CSV).

https://doi.org/10.1371/journal.pone.0348715.t003

Data-driven cost landscape

To formalize expressive regulation as constrained control, we construct a cost function using only statistics estimated from the clipped human data. We compute min–max parameters , , and define normalized variables

with a small to avoid division by zero.

Let be the median of the signed human discrepancy distribution computed on the clipped human data (and its normalized value). We define three components:

(1)

(2)

(3)

Weights are fixed from empirical dispersion using inverse-variance style scaling:

computed on the clipped human sample. We use to scale deviation by typical discrepancy magnitude irrespective of sign. The composite cost is

(4)

Several design decisions warrant explanation. Quadratic penalties are adopted for parsimony and analytic tractability. The deviation component is centered on the empirical median rather than on zero because the population does not converge toward zero discrepancy; centering on the observed median acknowledges this non-zero anchor as an empirical regularity rather than a normative target. The deviation term penalizes departure from typical discrepancy magnitude, not departure from zero; the cost landscape thus describes bounded variation around an empirical attractor, distinct from a model in which discrepancy is globally minimized. Taken together, J is an abductive, illustrative function. It was not fitted to individual trajectories nor inferred as a latent objective function. Its purpose is to provide a parsimonious geometric summary of the constraint structure consistent with the observed multi-regime occupancy.

We evaluate J on a regular grid spanning empirical ranges and export the full landscape as fig3_cost_landscape.dat, along with slices for visualization (e.g., fig3_cost_slice.dat). Qualitatively similar basin structure and contraction patterns were obtained under alternative weighting perturbations (Supporting Information (S1 Text)), supporting robustness to reasonable rescalings.

Expressive area and occupancy entropy

We quantify occupancy in the clipped plane using two complementary measures:

Software

All preprocessing and analyses were conducted in Python 3.12 using standard scientific libraries (e.g., numpy, pandas, scipy). As a methodological precedent for making high-inference constructs auditable via explicit, theory-grounded operationalization, we follow the same reproducibility-first design philosophy articulated in DefMoN [38], while applying it here to narrative–affect discrepancy geometry rather than defensiveness-oriented synthesis. A reproducible software bundle (Supplementary Software) documents the required dependencies and scripts used to generate the exported Source Data files. Figures are rendered in LaTeX via pgfplots and pgfplotstable, reading directly from exported .dat/.csv source files to ensure traceability between reported quantities and analysis outputs. No hand-tuned or simulated values were introduced into tables or figures.

Global geometry of narrative–affect discrepancy

We first examined the joint distribution of narrative complexity (N) and linguistically inferred affective intensity (A) across the full NCS corpus (n = 351,734). Supporting Information (Fig S1 in S2 Text) shows density projections in the (N,A), (N,D), and (A,D) planes, and S4 Table in S4 Text reports the corresponding correlation structure. Affective intensity spans a wide range, while narrative complexity varies substantially even within narrow affect bands.

The absence of strong diagonal structure in the (N,A) projection suggests that discrepancy does not arise from a simple coupling between affective intensity and narrative complexity. Instead, the projections reveal broad near-independence between N and A (r = 0.009; S4 Table in S4 Text), confirming that narrative complexity and affective intensity vary as largely independent dimensions. Fig S1 (Supporting Information, S2 Text) is read as a map of allowed configurations in the empirical geometry, motivating the regime-based and cost-based analyses that follow.

Fig S1 (Supporting Information, S2 Text) also reveals that the marginal density of D is not concentrated narrowly around a single value. Instead of converging toward low-discrepancy states, human narratives occupy a wide expressive range, suggesting that non-zero discrepancy is a stable and frequently visited configuration.

Regime identification in discrepancy space

To characterize this structure, we partitioned the (N,A,D) space into regions defined by the sign and magnitude of discrepancy and by relative narrative elaboration. Using empirical quantile thresholds computed from the human corpus (see Methods), we identify four robust regimes of discrepancy organization:

1. Coupled expression: the complement of the extreme discrepancy regimes, capturing the dense majority of narratives that do not meet the thresholds for strong under- or overstatement or collapse.
2. Strategic understatement: high A paired with relatively sparse narrative structure (, D < 0).
3. Strategic overstatement: low A accompanied by extensive narrative elaboration (, D > 0).
4. Collapse: high A with comparatively low narrative scaffolding and large-magnitude discrepancy, corresponding to a higher-cost region in which affective magnitude is expressed with minimal structural containment.

Throughout, the collapse regime is defined descriptively in terms of structural insufficiency relative to affective magnitude, without implying clinical pathology or dysfunction.

Table 3 reports regime-level summary statistics, including prevalence and mean values of (N,A,D). Coupled expression accounts for the majority of narratives (91.3%), while Understatement (n = 20,223), Collapse (n = 8,040), and Overstatement (n = 2,223) remain clearly represented. Thus, extreme discrepancy configurations are not isolated artifacts but systematic, measurable modes within the population distribution.

Cost landscape in narrative–control space

To interpret these regimes mechanistically, we define a data-anchored cost function J(N,A,D) that trades off exposure risk, cognitive effort, and deviation from a typical discrepancy target (Methods). This construction does not assume that individuals consciously optimize J; rather, it provides a parsimonious quantitative description of constraint structure consistent with multi-regime occupancy.

Fig 2 visualizes an exported slice of the cost landscape over (N,A) with point color indicating cost. Importantly, low-cost basins are not confined to D = 0 states; rather, distinct regions correspond to understatement and overstatement configurations under different trade-offs.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Data-anchored cost over (N, A) (slice at D* in the clipped space; Methods).

Low-cost basins correspond to configurations that trade off exposure risk (high A) against narrative effort (high N), rather than globally minimizing discrepancy. Understatement-leaning basins appear at higher A with lower N, whereas overstatement-leaning basins appear at higher N with lower A.

https://doi.org/10.1371/journal.pone.0348715.g002

This structure implies that discrepancy is not penalized per se. Rather, costs arise when expressive strategies fail to balance narrative scaffolding against affective intensity. Collapse corresponds to a higher-cost region in which affective magnitude is high but narrative resources are insufficient to regulate exposure.

Expressive area and regime occupancy

We next quantified the expressive area accessible to human narrators by computing the convex hull of occupied regions in the (N,A) plane. To ensure robust comparison and reduce tail leverage, we perform axis-wise clipping to human-derived percentile bounds and compute geometry in the clipped space (Methods). This area captures the range of discrepancy strategies empirically realized across the population under bounded observation.

Human narratives occupy a large and topologically rich area spanning all four regimes. Regime occupancy is continuous rather than discrete, with smooth transitions between modes. This suggests that expressive regulation operates along gradients rather than fixed categories.

Area captures the reachable extent of expression in the clipped space, complementing regime-based summaries of where narratives concentrate.

Comparison with an aligned language model

Finally, we projected an RLHF-aligned large language model into the same NCS using matched prompts and identical feature extraction procedures (Methods). The model’s expressive area is markedly contracted, occupying a narrow band concentrated near low-discrepancy states. Quantitatively, the model hull area is approximately 1.70 times smaller than the human hull area (human bootstrap mean: 99.51; LLM: 58.68; bootstrap 95% CI for contraction ratio: [1.68, 1.70]; permutation p < 0.0001, T = 10,000 shuffles; Fig 3; S6 Text). Notably, regions corresponding to strategic understatement and overstatement are sparsely populated, and the collapse regime is almost entirely absent.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Convex hull area for humans and the LLM in the clipped (N′, A′) plane (Methods).

The human value denotes the bootstrap-matched mean (100 resamples, n = 1,000 each). Contraction ratio: 1.70× (95% CI [1.68, 1.70]; permutation p < 0.0001). Full uncertainty reporting in S6 Text.

https://doi.org/10.1371/journal.pone.0348715.g003

Occupancy entropy, which measures uniformity of distribution within the explored region, was 4.654 for humans and 5.103 for the LLM. The higher LLM entropy indicates that the model distributes more uniformly within its narrower occupied region, whereas humans concentrate in specific regime zones despite spanning a broader area.

This contraction is not attributable to reduced affective range alone. Rather, it reflects a joint restriction on how narrative structure and affective intensity are combined. The model produces contextually coherent responses, but with limited variation in discrepancy strategies.

Taken together, these results show that humans systematically organize emotional expression across multiple discrepancy regimes, while alignment-constrained language models inhabit a much narrower expressive geometry.

Narrative–affect discrepancy as a measurable regulatory dimension

The present findings challenge the assumption that effective emotional expression necessarily minimizes discrepancy between narrative structure and expressed affective intensity. Across more than 350,000 human narratives, non-zero discrepancy is not only common but systematically organized into stable regimes. These regimes correspond to distinct expressive configurations rather than to noise or expressive failure.

From a psychological perspective, this pattern aligns with theories of emotion regulation that emphasize flexibility rather than uniform optimization. Classical process models [1,2] distinguish between antecedent-focused and response-focused strategies, while constructionist accounts highlight the role of contextual and conceptual scaffolding in shaping affective experience [5]. Narrative–affect discrepancy can be understood as an emergent dimension along which such strategies are implemented at the level of observable narrative form.

Strategic understatement and overstatement, in particular, appear as functional modes of regulation. Understatement allows individuals to express intense affect while limiting exposure, whereas overstatement provides cognitive distance through elaboration even when expressed affect is low. In this view, discrepancy is not a deviation from optimal expression but a resource-like degree of freedom enabling trade-offs among emotional intensity, cognitive effort, and social risk.

To reiterate a distinction introduced earlier: the term “regulated” throughout this paper refers to population-level distributional patterns, not to conscious or mechanistic control by individuals. The observed regime structure is consistent with constrained organization at the aggregate level; individual-level inference would require longitudinal or experimental designs beyond the scope of the present work.

Methodological and reproducibility contributions

Beyond substantive findings about discrepancy regimes, a central contribution of this work is methodological: it provides a transparent coordinate system for mapping large-scale narratives by jointly scoring narrative structure and expressed affect, and it reports all derived quantities directly from exported source-data files. The cost landscape is data-anchored in the empirical dispersion of the human corpus rather than imposed as an a priori theoretical optimum, and geometric comparisons are performed under explicitly stated clipping and resampling procedures. During revision, we identified that the deterministic relationship confines all observations to a plane in (N, A, D) space, rendering three-dimensional hull volume dependent on clipping artifacts rather than genuine distributional spread. We therefore report expressive area in the (N, A) plane, which provides a geometrically valid and more conservative measure of expressive extent. The contraction ratio is (bootstrap 95% CI [1.68, 1.70]; permutation p < 0.0001), preserving the direction and significance of the finding while correcting the geometric basis. Together, these design choices support technical evaluation on the basis of reproducibility and sensitivity analysis, consistent with a measurement-first approach to large-scale narrative data.

The cost landscape analysis supports an interpretation with multiple stable configurations. Rather than converging toward a single equilibrium, human narratives occupy multiple low-cost basins corresponding to different discrepancy regimes.

Such organization resonates with work on self-regulation and control in cognitive and affective systems, including active inference accounts emphasizing minimization of expected surprise under constraints [10,11], as well as inverse reinforcement learning frameworks that infer latent objectives from observed behavior [39–41]. Here, however, the minimized quantity is not discrepancy itself but a combined cost of exposure and effort. Discrepancy emerges as a degree of freedom through which this balance is achieved.

Importantly, collapse appears as a higher-cost region characterized by intense affect without sufficient narrative scaffolding. This regime aligns with clinical and developmental observations linking reduced narrative organization to dysregulation under high emotional load [14,27], consistent with computational psychiatry perspectives that model affective dysregulation as deviations from normative inference [42,43].

Implications for aligned language models

The comparison with an aligned language model reveals a markedly different expressive geometry. Despite producing contextually coherent narratives, the model occupies a much smaller region in discrepancy space, with limited access to extreme understatement or overstatement regimes. We do not attempt to dissociate alignment constraints from architectural or training-scale effects; the present comparison is intended as a descriptive contrast rather than a causal attribution. Expressive contraction may reflect safety filtering, prompt design, decoding constraints, or training distribution characteristics, not alignment objectives alone. The comparison uses a single RLHF-aligned model under a fixed prompt and decoding configuration and should not be generalized across architectures, training regimes, or objective functions. This contraction should not be interpreted as a simple deficit. RLHF-style objectives (e.g., safety and preference optimization) may encourage convergence toward comparatively conservative expressive configurations; however, the present analysis remains descriptive and does not establish causal attribution. Prior work has shown that alignment can reduce variance and extremity in model outputs [36,44], while studies applying cognitive and clinical benchmarks to language models have documented both surprising competencies and systematic limitations [45,46]. Our results extend this insight by suggesting that alignment can also constrain the structure of affective expression, not merely its content.

Limitations and future directions

Several limitations warrant note. First, the analysis focuses on textual narratives and does not directly address multimodal expression. Second, the model comparison involves a single GPT-4-class RLHF-aligned system under a fixed prompt and decoding configuration and should not be generalized across architectures or training regimes.

Third, the affective intensity measure (A) is derived from lexicon-based sentiment scoring (VADER), which captures valence magnitude expressed in surface language but does not distinguish arousal from valence, nor does it access suppressed or unexpressed affect. This is consistent with the study’s framing, which targets structural degrees of freedom within language rather than latent emotional states, but it means that A should not be interpreted as a proxy for subjective felt intensity.

Repeated interaction with expressively constrained systems may further narrow users’ own expressive repertoire through resonant amplification, a dynamic that cross-sectional snapshots cannot capture but that longitudinal extensions could address [47]. More broadly, the geometric framework introduced here may complement rights-based approaches to emotion AI, where the question shifts from whether a system classifies correctly to whether it preserves the expressive degrees of freedom available to the person it models [48].

Future work could extend this framework to longitudinal data, clinical populations, and alternative alignment strategies. Comparing models trained with different objectives may clarify which constraints most strongly shape expressive geometry. More broadly, integrating discrepancy-based metrics into evaluation protocols could complement existing benchmarks that focus on correctness or sentiment accuracy alone.