Comparative Analysis of Elicited and Unelicited Synthetic Narrative Cohorts

1 · Standing disclaimer

This document analyses two entirely synthetic narrative cohorts generated by large language models (GPT-4o-mini) from biographical scaffolding. The dataset is intended exclusively for demonstration of platform capability and for dogfooding the analytical pipeline that will be applied to real elicited dialogue in production use. This is not empirical evidence about any real population. Any apparent resemblance to real individuals or communities is coincidental and unintended. The methodological work that carries empirical weight is the manuscript "Eliciting the Narrative Register: A Longitudinal Study of Structural Thinking in Migration Dialogue," which uses real elicited dialogue from an eight-month engagement with a single participant, alongside the Blog Authorship Corpus as negative control. That manuscript demonstrates a 133:1 range difference in structural-thinking language between elicited and unelicited registers. The present analysis demonstrates platform output character, not psychological claims about real migrants.

2 · What you read

This analysis examines 20 synthetic personas distributed across two cohorts of 10 each. All personas represent Latin American migrants to Spain, with biographical arcs spanning the 2022–2025 period. The elicited cohort (personas 01–10) comprises individuals whose corpora include both diary-style entries and multi-turn exchanges with Prometheus, the platform's dialogue agent, responding to the 12-question elicitation protocol. The unelicited cohort (personas 11–20) comprises individuals whose corpora consist exclusively of diary entries with no structured elicitation. Both cohorts share the LatAm→Spain migration corridor and comparable biographical diversity (working-class to professional; secular to religious; multiple national origins; varied gender and sexual identities). Arc lengths range from 14 to 36 months. The elicited cohort produced 450 total entries (mean: 45.0 per persona; range: 33–61). The unelicited cohort produced 382 total entries (mean: 38.2 per persona; range: 30–47). Representative samples from each persona were provided for this analysis.

3 · Register comparison

The two cohorts differ markedly in register — the linguistic and rhetorical posture adopted by the writers. This difference is not reducible to topic or biographical detail; it is a structural property of how experience is narrated.

Elicited register

The elicited cohort demonstrates a consistent pattern of introspective abstraction and structural reframing. Writers in this cohort explicitly articulate what experiences mean, how identity is negotiated, and how past and present relate. Consider persona_01 (2023-10-01): "There's a bittersweetness in that question. I want to say Lavapiés feels like home, but sometimes it feels more like a stage, where I'm performing for an audience that doesn't know my lines." The writer is not merely reporting an event; she is examining the concept of home as a contested category, using theatrical metaphor to express the gap between performance and authenticity. Similarly, persona_04 (2025-03-22) writes: "Maybe that's what this place does—it changes the way one experiences endings." This is not description of a specific event but a meta-observation about how place reshapes temporal experience. The language is abstract ("the way one experiences endings"), reflective, and explicitly meaning-making. Persona_05 (2024-12-04) demonstrates the same pattern when discussing language choice: "Language shifts like the weather, depending on what I'm feeling. I keep reaching for what feels right in the moment, trying to find a balance, just like I'm trying to find my place in this new life." The writer is not reporting which language she used; she is theorising the relationship between linguistic choice and identity negotiation.

Unelicited register

The unelicited cohort, by contrast, is event-driven, observational, and present-tense. Writers report what happened, what they saw, what they felt in the moment, without stepping back to theorise the pattern. Persona_11 (2023-11-12): "susi's been pestering me for a walk since sunrise. damn dog, always ready for adventure. the sun is out, bright but not too hot. decent day for a change." This is immediate, concrete, anchored in the sensory present. The writer does not ask what the dog's persistence means about her own need for routine or connection; she reports the event. Persona_14 (2023-01-25) writes: "the cold is back, joder. woke up to a gray sky, everything feels heavy. skipped the long walk to the studio, took the bus instead." Again: weather, logistics, immediate feeling. No structural reframing, no abstraction about what "heaviness" signifies in the arc of adaptation. Persona_17 (2024-04-28): "hoy el día estuvo un poco más fresco. las nubes estaban pesadas, pero no llovió. mariana y diego jugaron en el salón, haciendo ruido como siempre." Weather, children's behaviour, routine observation. The entry does not theorise what the children's noise means about her isolation or what "fresco" weather signifies about seasonal cycles of hope.

What distinguishes the registers

The elicited register is characterised by:

• Abstraction language ("the concept of home," "the tension between," "what it means to")

• Structural reframing (events are not just reported but interpreted as instances of larger patterns)

• Explicit meaning-making ("I realised," "perhaps," "I wonder if")

• Temporal bridging (past and present are explicitly connected; the writer theorises continuity and rupture)

The unelicited register is characterised by:

• Concrete observation (weather, food, logistics, immediate sensory detail)

• Event sequencing (this happened, then that happened)

• Emotional immediacy (feelings are reported as they arise, not analysed)

• Audience consciousness (the diary is a record for the self or an imagined reader, not a dialogue partner prompting reflection)

The difference is not that elicited writers are "smarter" or "more articulate." It is that the elicitation architecture — Prometheus asking follow-up questions that press on meaning, identity, and pattern — produces a different kind of text. The unelicited writers are equally capable of abstraction; they simply are not prompted to perform it.

4 · Theme distribution

Elicited cohort themes

The elicited cohort foregrounds themes that emerge from the 12-question protocol:

• Home as contested concept (Q1: "Where do you feel most at home?") — nearly all elicited personas grapple explicitly with what "home" means when it is split across geographies.

• Language and identity negotiation (Q3: "Which language do you reach for when you're tired, angry, or tender?") — elicited personas theorise linguistic choice as identity work.

• Structural thinking about belonging — elicited personas describe belonging not as a binary (in/out) but as a process, a negotiation, a "space between."

• Temporal integration — elicited personas explicitly connect past (origin) and present (destination), theorising continuity and rupture.

• Meaning-making about displacement — elicited personas ask what their migration means, not just what happened.

Overlapping themes (present in both cohorts but foregrounded differently):

• Paperwork and bureaucracy — both cohorts mention it, but elicited personas frame it as symbolic of status and recognition (persona_04: "the homologation paperwork looms"), while unelicited personas report it as logistical frustration (persona_13: "billable hours stacking, but the merger pitch is still a tangled mess").

• Family contact — both cohorts describe calls and messages, but elicited personas theorise what distance does to relationships (persona_02: "maybe home is a mix of all this"), while unelicited personas report the content of calls (persona_16: "llamé a los chicos en la noche, y escucharlos fue un alivio").

Unelicited cohort themes

The unelicited cohort foregrounds themes that emerge from diary-writing conventions:

• Daily logistics (work shifts, grocery lists, bus schedules, weather)

• Immediate emotion (loneliness, exhaustion, flashes of joy, reported as they arise)

• Sensory observation (food, light, temperature, physical sensation)

• Routine and ritual (church attendance, market visits, phone calls as scheduled events)

• Children and caregiving (for those with children, the diary tracks their activities and needs)

• Financial anxiety (remittances, tips, bills — reported as concrete numbers, not theorised as symbolic of obligation)

Divergent themes (present in elicited but largely absent in unelicited):

• Self-coherence and integration — elicited personas explicitly ask how their past and present selves relate; unelicited personas do not.

• Identity as process — elicited personas describe identity as something actively negotiated; unelicited personas report identity markers (queer, trans, Aymara) but do not theorise the negotiation.

The theme distribution reflects the methodological point: elicitation does not invent new topics, but it shifts the register in which topics are discussed. Both cohorts talk about home, family, and paperwork. But the elicited cohort talks about these topics as concepts to be examined, while the unelicited cohort talks through these topics as lived experience to be recorded.

5 · Voice diversity within each cohort

Voice variety within each cohort is preserved. Persona_01's parenthetical asides and use of "che" and "boludo" are distinct from persona_03's ASCII sketches and rhetorical questions. Persona_07's epigraphs and em-dashes are distinct from persona_09's footnoting and use of "one." The synthetic generation preserved individual voice fingerprints as specified in the biographical scaffolding. Similarly, in the unelicited cohort, persona_11's lowercase-only, emoji-laden fragments are distinct from persona_12's formal "mi querido diario" address and "hasta mañana, si Dios quiere" closings. Persona_14's italics-by-asterisk and movement notation ("plie, releve, plie") are distinct from persona_17's Guaraní exclamations and light-focused observations. However, the register difference between cohorts is more consistent than the within-cohort voice differences. That is: persona_01 and persona_07, despite very different voices, both produce the elicited register (abstraction, structural reframing, meaning-making). Persona_11 and persona_17, despite very different voices, both produce the unelicited register (event-driven, observational, present-tense). The elicitation architecture produces a recognisable register shift that operates across individual voice variation. This demonstrates the methodological point: elicitation is not a topic shift or a personality shift. It is a register shift — a change in the rhetorical posture and linguistic architecture of the text. The personas remain distinct from each other, but the elicitation protocol produces a consistent difference in how they narrate, not just what they narrate.

6 · What this demonstrates and what it does not

What it demonstrates

The platform's elicitation architecture (Prometheus + 12-question protocol + multi-turn dialogue) produces narrative material that is observably different from free-form diary writing on broadly similar biographical material. The difference is not in subject matter alone — both cohorts discuss home, family, language, paperwork, and displacement. The difference is in posture, abstraction language, and the explicit articulation of what experiences mean. Specifically:

• Elicited writing demonstrates introspective abstraction. Writers step back from events to theorise patterns, to ask what experiences signify, to connect past and present structurally.

• Unelicited writing demonstrates observational immediacy. Writers report events, sensations, and emotions as they arise, without stepping back to theorise the pattern.

• The difference is consistent across individual voice variation. The elicitation register is recognisable even when voices differ markedly in style, tone, and linguistic fingerprint.

This is a demonstration of platform output character. It shows that the elicitation architecture, when applied to synthetic biographical material, produces a recognisably different kind of text. Whether this difference would hold with real elicited dialogue is an empirical question that the manuscript addresses (and answers affirmatively, with a 133:1 range difference in structural-thinking language). The synthetic demonstration shows that the platform's design intends to produce this difference; the manuscript shows that the platform's design succeeds in producing this difference with real subjects.

What it does not demonstrate

This synthetic dataset does not demonstrate:

• That real elicited dialogue would behave the same way. The manuscript shows that real elicited dialogue produces a measurable register shift (133:1 structural-thinking-language range relative to the Blog Authorship Corpus). Whether synthetic gpt-4o-mini-generated dialogue produces a comparable register shift is an empirical question this demonstration cannot answer. The synthetic data validates the pipeline (the analytical tools that will process real dialogue), not the methodology (the claim that elicitation produces a register shift in real subjects).

• That the difference would generalise to other populations beyond LatAm→Spain. The synthetic cohorts were deliberately constrained to a single migration corridor to control for cultural and linguistic variables. Whether the elicitation register would be recognisable in, say, South Asian→UK or East African→Gulf cohorts is unknown.

• That the difference has the same magnitude as the manuscript's 133:1 finding. The manuscript's finding is based on eight months of sustained engagement with a single deeply-engaged participant. Synthetic personas, generated in a single pass, are unlikely to replicate that magnitude. A 5:1 or 20:1 ratio would be a reasonable demonstration; a 1:1 ratio would suggest the subject-voice prompt is failing to produce the elicitation register.

• Any psychological claim about real migrants, real third-culture individuals, or real diaspora communities. The personas are synthetic. Their "experiences" are generated from biographical scaffolding. Any apparent insight into real migration psychology is coincidental and unintended. The platform is a research instrument, not a psychological model.

• That the platform is ready for production use. This is a demonstration, not a validation. Real elicited dialogue will differ from synthetic dialogue in ways that cannot be predicted from this analysis. The platform's credibility depends on honest acknowledgment of what synthetic data can and cannot show.

7 · What the DOL pipeline would be expected to show

The DOL (Distribution of Cognitive Load) Python pipeline is a deterministic, lexicon-based analytical tool that produces quantitative measures of linguistic features. It is open-source and lives at github.com/RayanBVasse/DOL. DOL is not run as part of this synthesis — that is a separate analytical step, manually triggered, that produces numbers. This section describes what running DOL on these synthetic samples is expected to show, with explicit caveats about the limits of synthetic data.

Expected findings

Structural-thinking lexicon scores: The elicited cohort should score higher than the unelicited cohort on the structural-thinking lexicon (terms like "pattern," "tension," "negotiate," "integrate," "process"). This is because the elicitation protocol explicitly prompts for structural reframing. However, the magnitude of the difference is uncertain. GPT-4o-mini may produce more or less structural-thinking vocabulary than real subjects, depending on how the subject-voice prompt is tuned. A 5:1 or 20:1 ratio would be a reasonable demonstration; a 1:1 ratio would suggest the prompt is failing. Epistemic-uncertainty arcs: The elicited cohort should show non-monotonic epistemic-uncertainty arcs (rises during exploratory phases, declines toward integration), reflecting the dialogic process of meaning-making. The unelicited cohort should show more event-tracking patterns (spikes at specific worries, declines without arc), reflecting the diary's function as a record of immediate emotion. However, synthetic arcs may be less pronounced than real arcs, because synthetic personas lack the sustained engagement that produces genuine epistemic shifts. MTLD / TTR (vocabulary specialisation): The elicited cohort should show declining MTLD (measure of textual lexical diversity) over time, reflecting specialisation as the persona converges on themes Prometheus surfaces. The unelicited cohort should show stable MTLD, reflecting the diary's function as a record of varied daily experience. However, synthetic personas may show weaker trends than real subjects, because synthetic generation does not replicate the cognitive process of thematic convergence. Spearman correlations with time: Directional correlations (structural-thinking rising, epistemic-uncertainty declining, MTLD declining in elicited; all stable in unelicited) are expected but may be weak given small N (10 + 10 personas) and short arcs (14–36 months). Real longitudinal data with larger N and longer arcs would be needed to establish robust correlations. Magnitude relative to manuscript's 133:1 finding: The manuscript's 133:1 structural-thinking-language range is based on real elicited dialogue with eight months of sustained engagement and a single deeply-engaged participant. Synthetic ratios will almost certainly be smaller. A 5:1 or 20:1 result would be a reasonable demonstration of platform output character; a 1:1 result would suggest the subject-voice prompt is failing to produce the elicitation register. A 133:1 result would be implausibly high and would suggest overfitting to the manuscript's findings.

8 · Honest limits

What synthetic data cannot test

• Real-world elicitation effects. Synthetic personas do not experience genuine epistemic shifts, identity negotiation, or meaning-making. They simulate these processes based on the subject-voice prompt. Whether real subjects would produce comparable output is an empirical question that requires real subjects.

• Real subject voice. Synthetic personas are generated from biographical scaffolding. Real subjects have idiosyncratic voices shaped by lived experience, not prompt engineering. The manuscript's single participant produced voice patterns that cannot be predicted from synthetic generation.

• Real cultural patterns. Synthetic personas simulate cultural markers (Aymara, Zapotec, Afro-Peruvian, etc.) based on the biographical scaffolding. Real cultural patterns are emergent, contested, and irreducible to demographic labels. The platform's credibility depends on not overclaiming about cultural insight.

• Real cohort effects. Synthetic cohorts are generated independently. Real cohorts develop shared language, reference each other's experiences, and produce emergent group dynamics. The platform's longitudinal design assumes cohort effects; synthetic data cannot test whether they occur.

What this demonstration is good for

• UX dogfooding. The synthetic cohorts allow the research team to test the platform's interface, visualisation tools, and analytical pipeline before engaging real subjects. This is essential for identifying usability issues and refining the design.

• Illustrative figures. The synthetic cohorts provide material for public-facing demonstrations, conference presentations, and training materials. They show what the platform intends to produce, even if real output will differ.

• Public-site showcase content. The Global Narrative Atlas demonstration site can use synthetic cohorts to illustrate platform capability without exposing real subjects to public scrutiny. This protects subject privacy while demonstrating platform value.

• Training material for researchers. Synthetic cohorts provide a safe environment for researchers to learn the DOL pipeline, practice interpretive analysis, and develop hypotheses before engaging real data.

What this demonstration is not good for

• Any empirical claim. Synthetic data cannot support claims about real migration, real identity negotiation, or real narrative patterns. Any such claim would be methodologically invalid.

• Any generalisation. Synthetic cohorts are constrained to LatAm→Spain, 2022–2025, and the specific biographical scaffolding provided. Generalising to other populations, time periods, or migration corridors would be unjustified.

• Any methodological validation independent of the manuscript. The manuscript's empirical work (real elicited dialogue + Blog Authorship Corpus negative control) is the methodological evidence. The synthetic demonstration is a platform capability demonstration, not a methodological validation.

The synthetic demonstration is a proof of concept, not a proof of claim. The manuscript provides the proof of claim. This document provides the proof of concept.