Editorial transparency

How we build and review the corpus

Synthetic Personality Review is not a list of search results. It is a closed editorial corpus of 370 unique studies, selected for relevance and reviewed from primary sources. All 370 records include full-text reading, bilingual synthesis, and an approved fidelity check.

01 · Question and scope

What belongs in the field

We include studies in which personality, a psychological disposition, identity, or a conditioned persona is the object of measurement, intervention, or validation in a language model. We also include work evaluating the fidelity, risks, or uses of agents representing individuals or groups, together with methodological contributions directly applicable to psychometric evaluation of LLMs.

Inclusion criteria

Measures or tests personality or psychological dispositions in an LLM.
Induces, controls, trains, represents, or maintains traits, identity, or persona.
Evaluates fidelity, validity, safety, bias, or use of people or groups represented through LLMs.
Makes a direct methodological contribution to psychometric evaluation of language models.

Non-inclusion criteria

Persona is only a prompting label and is not analyzed as a construct or intervention.
Psychometrics only scores an unrelated task or an instrument intended exclusively for humans.
Personalization, memory, recommendation, or multi-agent work lacks a substantive question about personality, fidelity, or human simulation.
A stronger or more direct version of the same work is already represented.

Exclusion does not mean that a study is low quality. It means that the study falls outside this version's editorial question, is redundant, or is represented more directly by another work.

02 · Discovery and decision

How we decide what to include

The update begins with reproducible searches in arXiv and OpenAlex, combines queries about LLMs, personality, personas, role-play, and psychometrics, normalizes titles, and reconciles versions. The discovery cutoff was July 14, 2026; editorial closure and quality replacements were completed on July 18, 2026.

Discovery. Potential results are collected and exact title duplicates are removed.
Bibliographic identity. Title, authors, year, identifiers, and relationships between preprints and publications are checked.
Relevance screening. Title and abstract determine whether a source should be acquired and read; they are never used to manufacture the final summary.
Prioritization. When the set exceeds the limit, direct relevance and methodological contribution take precedence over tangential mentions.
Source access. An approved record requires verifiable full text. A source that cannot be acquired legally does not occupy an approved place.
Full-text eligibility. Reading can confirm inclusion, reveal a tangential scope, or identify a duplicate version.

The maximum of 370 is a version limit, not a quota. It structures priority, but it does not make a tangential study relevant or justify lowering the editorial threshold.

03 · Published funnel

What percentage was not included

The correct denominator is the set of genuinely new candidates. Of 539 unique results,42 were already in the starting collection, leaving 497 new candidates requiring a decision.

539

Unique results

Found in the literature update.

497

New candidates

After separating 42 studies already present.

129

Initial selection

26.0% of new candidates.

151

Represented at closure

30.4% after replacements and controls.

Not included · initial screening

74.0%

368 of 497 new candidates.

Not included · final corpus

69.6%

346 of 497 new candidates.

The difference comes from 22 initially deprioritized studies that were later included to replace closed sources, duplicates, or records that did not pass the final audit.

04 · Editorial production

From candidate to approved review

Acquisition

The best available primary source is acquired and a verifiable snapshot is preserved.

Full reading

Question, method, results, tables, figures, limitations, and relevant supplements are reviewed.

Extraction

Question, design, models, instruments, data, sample, findings, and limitations are separated.

Original synthesis

Spanish and English versions are written from the reading, without mechanically expanding or trimming the abstract.

Cross-check

Central claims are checked again against section, page, table, or figure locators.

Approval

Only a complete, traceable, bilingual, full-text record is published as validated.

05 · Evidence structure

What every record contains

The public page is a readable view of a broader editorial record. Abstract and summary are not the same: the abstract belongs to the authors; the summary is a Synthetic Personality Review synthesis checked against the full text.

Identity

Title, authors, year, language, identifiers, publication type, URL, and relationships between versions.

Source

Consulted version, retrieval date, access, SHA-256 hash, and actual reading scope.

Analysis

Question, method, models, instruments, datasets, sample, findings, limitations, and what the study does not establish.

Control

Evidence locators, confidence, responsible reviewer, status, review date, and approval date.

Bilingual content

Original abstract when available and independent editorial summaries in Spanish and English.

Classification

One primary track and keywords, without duplicating a study to inflate category counts.

06 · Editorial instrument

Our evaluation sheet

Every study is reviewed through the same structured sheet. It is not an automated score or a rubric that produces a grade. It requires the reviewer to document what was read, what the evidence supports, and which limitations prevent broader conclusions. These are the actual fields a review must complete before approval.

Editorial sheet · schema v1

Critical evaluation of one study

Possible states: pending, in review, needs correction, reviewed, or approved.

Block	What the sheet records	Approval condition
Identity and scope	`article_id`, status, reading scope, and relationship to other versions.	One relevant canonical work with no editorial alias.
Consulted source	URL, version, access, retrieval date, and SHA-256 hash.	Verifiable full text whose hash matches the source manifest.
Question and design	Research question, method, sample, models, instruments, metrics, and datasets.	Method described from the source without filling missing information.
Editorial synthesis	Original Spanish summary and an editorially equivalent English version.	Both summaries complete, at least 150 words, and distinct from the authors' abstract.
Critical reading	Findings, limitations, and claims the study does not establish.	All three sections contain evidence and delimit the scope of the conclusions.
Evidence	Claim and locator pairs pointing to a section, page, table, figure, appendix, or artifact.	At least one verifiable locator for the central claims.
Editorial control	Confidence, reviewer, review date, approver, and approval date.	Every responsible person, date, and fidelity control is complete.

The final decision is not based on adding points. A critical gap, such as unavailable full text, unlocatable evidence, or failure to separate results from limitations, blocks approval even when the rest of the sheet is complete.

07 · Sources and states

What each source supports

Available source	Permitted use	Approved summary?
Full journal article or proceedings paper	Reading, extraction, synthesis, and approval.	Yes, after cross-checking.
Complete author manuscript	Full reading; the consulted version is disclosed.	Yes, with preprint status disclosed.
arXiv, OpenReview, or another academic repository	Full reading and version traceability.	Yes, if the document is complete.
Official abstract only	Metadata and a possible pending record.	No.
Search snippet, third-party summary, or unsourced text	Not accepted as editorial evidence.	No.

08 · Thematic organization

Results, figures, samples, models, and conclusions are never invented.
Correlations are not converted into causality and claims are not generalized beyond the design.
Missing information is not filled with general knowledge, and deterministic rules never select abstract sentences.
Results, author interpretation, limitations, and editorial judgment remain distinct.
The English version preserves structure, numerals, and evidence locators one to one, and is linked by SHA-256 to the approved Spanish analysis.
Automated controls check schema, coverage, bilingual content, duplicates, and risk signals; they do not replace reading.
If a source disappears, becomes closed, points to another work, or duplicates a version, it is corrected or replaced and the removed record remains traceable.

A reference corpus must be able to correct errors without hiding them. The publication therefore preserves source manifests, content hashes, editorial states, and a separate removal registry.

The build is blocked if a translation loses a numeral, changes item count or order, alters evidence locators, or becomes stale relative to the Spanish content it translates.