What is guided reflection?

Not all narrative is equally legible.

Survey responses, casual blog posts, social-media threads, and chatbot exchanges all produce text. They do not all produce the same kind of text. The register in which a person writes — privately to themselves, publicly to an imagined audience, transactionally to a service, reflectively to a non-judgmental interlocutor — determines what computational analysis can see.

Migration researchers have long understood that the questions asked shape the narratives obtained. The life-history interview is an active elicitation method; the ethnographic encounter is an intersubjective co-construction. GNA extends this principle into the computational domain: its canonical questions are not a neutral input window, but the elicitation mechanism that produces the narrative conditions under which identity trajectories become tractable.

Three phases of structured inquiry.

A canonical question set designed to prompt the three narrative registers identified in prior identity research as carrying computational signal for cognitive and emotional transformation.

Grounding

Establishing a participant’s cultural and biographical baseline. Questions probe early life environments, migration history, languages spoken across contexts, and lived experiences of belonging and non-belonging. Surfaces the participant’s narrative framework for organising experience.

Exploration

Systematically examining experiences of cultural mismatch, identity conflict, code-switching, and the cost of moving between worlds. Elicits the reflective meaning-making in which cognitive markers become detectable.

Integration

Inviting reflection on change over time — how the participant’s relationship to their own multiplicity has shifted from childhood to adolescence to adulthood, and what was once not yet understood about one’s own way of belonging.

Sample questions

Where, exactly, is home for you right now — and has that answer changed?

When do you find yourself translating, not just words but yourself, for others?

What does it cost you to belong in the different worlds you navigate?

If you think about yourself five years ago, what did you not yet understand about your own way of belonging?

The canonical question set is not fixed. It serves as the initial template for platform-mediated dialogue. In full deployment, an AI assistant administers follow-up probes tailored to the participant’s responses, deepening the reflective register without scripted rigidity.

Forty years of evidence that prompts shape register.

The elicitation-dependence hypothesis is not a novel claim. It is the operationalisation, in longitudinal AI-mediated text, of a finding that three independent research traditions have converged on over four decades.

The Pennebaker tradition

Beginning with Pennebaker & Beall (1986), expressive-writing studies established that prompted reflective writing produces measurable elevations in cognitive-processing vocabulary — specifically causal words (“because,” “reason”) and insight words (“realise,” “understand”). Pennebaker & Francis (1996) showed this with LIWC; Klein & Boals (2001, 2010) replicated and extended the finding across populations and languages, and tied the lexical signature to mediation of meaning-making and post-traumatic growth. GNA’s structural-thinking lexicon is a more elaborated version of what Pennebaker called cognitive-mechanism vocabulary.

Discourse-elicitation in clinical linguistics

The aphasia and temporal-lobe-epilepsy literatures have systematically shown that different elicitation methods (procedural, narrative, expositional, free) produce systematically different linguistic properties — TTR, propositional density, syntactic complexity, words-per-minute. Where structure is imposed on discourse, the lexical and propositional scope narrows. GNA’s MTLD-decline finding is what discourse-elicitation research would predict in advance (Fergadiotis & Wright 2011; Bryant et al. 2016).

Narrative-interview methodology

Schütze’s narrative-interview tradition, Flick’s episodic-interviewing technique, Atkinson’s life-story protocol, and Jovchelovitch & Bauer’s formalisation all rest on the premise that how one prompts determines what register one obtains. This is the qualitative-methods anchor for the same claim.

Adjacent inner-life literatures

The platform speaks to a cluster of psychological constructs that sit adjacent to the elicitation-dependence hypothesis. McAdams and McLean (2013) define narrative identity as an internalised and evolving life story that integrates reconstructed past and imagined future into a sense of unity and purpose — the constructional target the AI-driven “Prometheus” protocol elicits. Oishi and Westgate (2022) frame psychological richness — variety, interestingness, and perspective-changing experience — as a third dimension of the good life alongside happiness and meaning, measured by the Psychologically Rich Life Questionnaire. Kashdan, Barrett and McKnight (2015) unpack emotional granularity as the capacity to make fine-grained distinctions between affective states — the construct most directly addressed by DOL’s lexical-extraction pipeline, particularly its epistemic-uncertainty lexicon.

GNA does not synthesise these constructs into a single index. It offers an elicitation-and-analysis substrate on which a study could measure any subset of them longitudinally — the canonical 12-question protocol as elicitation backbone, DOL’s deterministic lexicons as analytical layer. Construct correspondence — whether GNA’s lexicon hits track the same trajectories these instruments measure — is untested and the natural subject of replication work.

• Oishi, S., & Westgate, E. C. (2022). A psychologically rich life: Beyond happiness and meaning. Psychological Review, 129(4), 790–811.
• McAdams, D. P., & McLean, K. C. (2013). Narrative identity. Current Directions in Psychological Science, 22(3), 233–238.
• Kashdan, T. B., Barrett, L. F., & McKnight, P. E. (2015). Unpacking emotion differentiation: Transforming unpleasant experience by perceiving distinctions in negativity. Current Directions in Psychological Science, 24(1), 10–16.

GNA does not claim to have discovered that elicitation changes register. It claims to have built deterministic, open-source analytical infrastructure that operationalises this finding longitudinally at scales the original studies could not reach.

A platform designed to elicit a register the pipeline can read.

In exploratory empirical work, the GNA’s analytical engine was applied to two contrasting corpora through an identical pipeline. The contrast is striking — and worth reading with care, since the two corpora differ on several confounded dimensions.

• Genre: private reflective dialogue vs public-facing blog posts.
• Audience: AI interlocutor vs imagined readership.
• Era: 2025 vs 1999–2006.
• Selection: a deeply engaged single author vs a hundred randomly sampled bloggers.
• Domain: identity / migration reflection vs varied everyday topics.

Any of these could plausibly contribute to the observed difference. The finding is consistent with an elicitation effect; it does not establish one. A within-subject contrast — the same participant writing under both conditions — would be required for a clean causal estimate.

Source: Eliciting the Narrative Register: A Computational Platform for Longitudinal Identity Research (Vasse 2026). doi:10.5281/zenodo.20142013

Adjacent surfaces, distinct purposes.

From therapy

Therapy is clinical, oriented toward wellbeing outcomes, and conducted within a regulated practitioner relationship. Guided reflection is a research method. It does not diagnose, treat, or prescribe.

From journaling

Journaling is unstructured and audience-less. Guided reflection is structured around canonical questions, longitudinal in design, and analytically instrumented from the first prompt. The structure is the source of its computational signal.

From chatbot conversation

Consumer chatbots are optimised for engagement, helpfulness, or entertainment. Guided reflection is bounded by research protocol; the dialogue serves the analytical pipeline, not the other way round. There is no character, no relationship simulation.

From surveys and standardised scales

Surveys produce quantified states across populations. Guided reflection produces longitudinal trajectories within individuals. The two methods complement, but do not substitute for, one another.

From passive social-media analysis

Passive traces are continuous, naturalistic, and audience-directed. Their register is shaped by performative context. Guided reflection produces the introspective register that passive traces cannot.

How elicited narrative becomes a trajectory.

1. 01 · Cognitive profiling

   Two lexical markers — structural thinking (40 terms) and epistemic uncertainty (12 terms) — are scored per calendar month, normalised per 1,000 words. Lexicons are theory-guided and publicly auditable.

2. 02 · Vocabulary specialisation

   Type–Token Ratio and the length-invariant Measure of Textual Lexical Diversity (MTLD) are computed monthly. Declining values are consistent with vocabulary narrowing through domain specialisation; alternative interpretations include disengagement, formulaic response patterns, lexical fatigue, or language attrition. The construct is most interpretable when combined with thematic-domain analysis (which is also computed) and with the participant’s qualitative context.

3. 03 · Topic modelling

   Messages are vectorised (TF-IDF, 60,000 features), reduced via truncated SVD, and clustered into fine-grained topics that are aggregated into interpretable macro-domains through manual review. This approach is chosen for transparency, determinism, and replicability rather than for state-of-the-art performance; contemporary alternatives (BERTopic, dynamic topic models, contextual embeddings) are under evaluation for future versions and would be implemented as opt-in pipeline variants alongside the current method, not replacements.

4. 04 · Statistical validation

   Temporal trends are tested via Spearman rank correlation against permutation null distributions (10,000 shuffles). Holdout validation confirms model stability across time windows.

5. 05 · Bounded synthesis

   An LLM receives only the extracted scores — never raw text — and translates the trajectory into readable interpretive prose under a schema-constrained prompt that prohibits speculation beyond the data. Generated reports are audited before delivery by two mechanisms: (1) an Assistant Register lexicon (71 terms covering therapeutic, motivational, chatty, and universalising drift) is run on every report and flags drift for regeneration with tightened prompts; (2) a sample of reports is reviewed monthly by a member of the Technical Council against the underlying structured data, to verify that no inferential claim has been introduced beyond the data the LLM received.

6. 06 · Reflective return

   The participant or researcher receives the trajectory as a longitudinal artefact, to be read against further dialogue, qualitative work, or independent evidence.

Deterministic extraction: Lexicon-based extraction is deterministic — identical inputs yield identical outputs from the analytical pipeline. The interpretive layer is bounded and grounded in extracted scores, never in raw text. Reproducibility in the wider scientific sense — that independent researchers obtain comparable results on independent data, and that the constructs measured correspond to what they purport to measure — is the subject of ongoing work and is not yet established.

Elicitation as a methodological condition.

The conventional reading of computational text analysis assumes the analytical pipeline is the active instrument and the text is a passive substrate. GNA reverses this. The elicitation architecture is the active instrument; the analytical pipeline reads what the architecture has produced.

This reframes the platform as a research instrument, not a passive analytical tool. A migration researcher deploying GNA to study, for example, the identity trajectories of third-culture individuals or second-generation migrants is not simply collecting pre-existing text. They are engaging participants in a structured reflective process that itself constitutes a form of identity work. The platform’s bounded interpretive output — the trajectory reports — are a record of that process.

We use “computational hermeneutics” in the sense of Kommers et al. (2026) — a methodological lens through which computational text analysis is treated as an active participant in the production of meaning, rather than as a neutral observer. We do not claim substantive engagement with the philosophical hermeneutic tradition (Gadamer, Ricoeur, Schleiermacher); the term is borrowed and instrumental, naming a working stance, not a philosophical commitment.

The questions a platform asks — and the implied audience to which a participant responds — are part of the methodological condition of computational analysis. Acknowledging this is not a limitation to be managed but a design parameter to be optimised.

— Eliciting the Narrative Register (Vasse 2026)

What we do not yet know.

• The positive demonstration is a self-study (N = 1, the platform’s developer, across eight months). Trajectories are statistically structured within the series but cannot be generalised without replication across independent participants. The infrastructure is offered as a platform for that replication, not a closed claim.
• The lexicons are currently English-only. Cross-linguistic validity, particularly for migration research where code-switching and multilingual narrative are the norm, remains untested. The architecture can accommodate additional language-specific lexicons.
• The structural-thinking and epistemic-uncertainty lexicons are theory-guided probes whose construct validity against independent cognitive measures is unknown. They are reliable under specified conditions; they are not psychometrically validated instruments.
• The structural-thinking and epistemic-uncertainty lexicons fire on vocabulary, not on cognitive register directly. Out-of-domain testing showed the structural-thinking lexicon firing most strongly on posts whose topic happened to involve structural vocabulary, not on posts reflectively structural in the cognitive sense. The lexicons are topic-and-register-sensitive markers, not direct measures of cognitive structuration.
• The blog negative-control corpus dates from 1999–2006. Different results might obtain with contemporary social-media data, chat exports, or AI-mediated logs. The historical character of the control is methodologically useful precisely because it predates the LLM era.
• The Blog Authorship Corpus is one negative control. Future work should add intermediate controls — diary corpora, free-form personal writing, life-history interview transcripts — to triangulate the register effect against multiple unprompted baselines.