BlossomPsy is an MBTI assessment prototype that replaces part of a questionnaire with open conversation and, when its estimator reports low confidence, image pairs. Two Doubao agents ask questions in Chinese. A shared RoBERTa-base feeds four binary heads and one 16-type classifier; modified UCB chooses which dimension to probe and PID tunes two confidence-transformation parameters during synthetic training. Thirteen MBTI-M items are converted into visual pairs by three agents plus human supervision, and nine are retained. This is an adaptive-interface concept, not a validated psychometric instrument or clinical tool.
The current arXiv v2 comparison pools 12 Chinese university students with nine Qwen/Doubao-simulated participants. All 21 complete BlossomPsy and Chinese MBTI-M. Seven match on four dimensions, nine on three, three on two, and two on one; thus 16/21 match at least three. Results are not separated for the 12 humans and nine LLMs. The latter are 43% of the sample and receive prompts instructing them to adopt an MBTI type; their consistency can stress-test the flow but does not add human psychometric evidence. MBTI-M is a reference, not personality ground truth.
The dimensional table is not arithmetically reproducible. It reports accuracy/F1/kappa of 0.76/0.79/0.63 for E/I, 0.67/0.72/0.50 for S/N, 0.95/0.92/0.90 for T/F, and 0.76/0.78/0.62 for J/P. The paper defines accuracy as correct/(correct+incorrect). In binary classification, accuracy 0.76 permits a maximum kappa of about 0.546, so 0.63 and 0.62 are impossible; accuracy 0.67 permits at most about 0.405, not 0.50. Exhaustive enumeration likewise finds no 21-case 2x2 tables reproducing those metric triplets. Only T/F is feasible. If rounded accuracies correspond to 16/21, 14/21, 20/21, and 16/21, Wilson 95% intervals are approximately [0.549,0.894], [0.454,0.828], [0.773,0.992], and [0.549,0.894]. Without labels, confusion matrices, or code, the kappas should not be cited as reliable evidence.
There is also substantive version drift. Version 1 claimed 45 participants, 12 humans and 33 LLMs, and displayed a 16-type distribution summing to 45. One day later, v2 claims 21, the same 12 humans but only nine LLMs, and replaces that distribution with a source table. Match rates, accuracy, F1, kappa, the experience radar, and conclusions remain unchanged. Match percentages correspond exactly to 7, 9, 3, and 2 of 21, but no revision explanation or data reveals which figures use 21, 45, or other subsets.
The higher-satisfaction claim rests only on a radar chart without numeric values. It does not say whether 12 humans, all 21 mixed cases, or another sample responded; items, anchors, means, spread, individual responses, order, counterbalancing, and paired tests are missing. The chart suggests a preliminary interface preference but does not establish a statistical increase in satisfaction or engagement. Completion time, dropout, errors, accessibility, and cognitive load are also absent.
Visual validation is simulation-dominated. Each of 13 pairs receives ten human responses and 40 Qwen simulations, so 80% of 50 votes are synthetic. Pairs are adapted and re-tested until exceeding 66%, and the nine retained items receive no independent validation. Matching the text item measures choice agreement, not whether the image measures an MBTI trait. No item analysis, factor structure, differential-item functioning, or psychologist validation is reported. The text recognizes gender-related visual bias and changes figures based on user gender, but does not validate the rule, address nonbinary identities, or report fairness analysis.
The classifier trains on Personality Cafe, 8,675 users with 50 posts and self-reported labels, then fine-tunes on CPME posts filtered through purported student words and translated into English. Post-filter sizes, selection vocabulary, translator, splits, indices, seeds, and author separation are unavailable. Personality Cafe explicitly discusses MBTI and can leak type names; the paper does not say whether it removes them or segments before splitting users. CPME labels users; absent an author-level split, fragments from one user may cross train and test. Duplicate controls and training-only over/undersampling are also unspecified. The strong RoBERTa row, 0.838/0.920/0.797/0.759, lacks an unambiguous test population and metric. Another table reports 84/87/86/81% in a different unspecified “contextual setting.”
Ablations use only 16 LLM simulations, one per type, without repetitions or uncertainty. The full system shows more matches than some variants, but removing PID, photos, or mUCB also changes information and conversation paths. PID converges near alpha=15.57 and beta=5.51 in a corrupted CPME environment; it is not compared with grid or random search, and the paper concedes simple alternatives may match it. Two-scalar convergence does not demonstrate calibration, better adaptive testing, or a need for PID control.
No code, checkpoint, processed data, raw evaluations, or statistical scripts were found; the arXiv archive contains TeX and figures. Consent, ethics review, compensation, recruitment, and governance of personal conversations are not reported. The defensible contribution is an exploratory UX and orchestration design for making a test more conversational. It does not establish MBTI validity, real personality, Big Five portability, causal component superiority, or suitability for clinical, educational, employment, or other high-stakes decisions.