Inter-Rater Self · v1
A model does not know you. It samples your history and infers — a biased sample dressed as insight. Run the same read across several models and the disagreement becomes the instrument. Convergence is your stable core; the widest spread is either context-dependent or unmeasurable from text. Paste each model's JSON below.
MBTI is a tool for self-reflection — a way to learn more about your personality, not a fixed label. Your type can shift across stages of life. The framework traces back to Carl Jung's personality theory, later developed into the MBTI by Isabel Briggs Myers and her mother Katherine Briggs.
One advantage of using an AI for this is that it may already have a record of how you write and express yourself — not because it "knows" you, but because it can read patterns in your past conversations to estimate where you fall on these scales. The model doesn't know you personally. Over time, your chat history across apps like ChatGPT, Claude, Grok, or Gemini builds a sample of your communication style, biases, and preferences. This works best with real history — ideally 20+ genuine conversations on at least one app.
The more history a model has, the cleaner the read. When multiple models return similar scores, that's a signal worth taking seriously. When scores vary widely across models, treat the result as inconclusive — that's the data being thin, not a hidden depth surfacing.
This is a tool, not a substitute for a mental health professional or clinician. I'm not a professional in this field — just someone who looks for ways to gradually improve and shares what's useful. If something here doesn't land, set it aside.
Paste the prompt into 3–5 models you've actually talked to (20+ conversations each, memory enabled). Don't tell any model what another said — you want independent reads, not influenced ones. Each model returns prose plus a JSON block. Collect the JSON, drop it into the dashboard below, and look at how much the scores differ between models.
Perform a personality analysis of me from our conversation history. Don't ask questions — use what you've observed. I want accuracy, not flattery, and I'm comparing your output against other models.
Score Big Five / OCEAN 0–100 (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism), each with a specific observed pattern as evidence. Then give Enneagram core type, wing, and confidence, citing the conversational behavior that typed me — how I handle being wrong, disagreement, what engages me. Then MBTI four-letter type with per-axis confidence; flag close calls.
Note where you have thin data instead of guessing.
Then output ONLY this JSON block — no prose inside it, no markdown fences:
{"model":"", "ocean":{"o":0,"c":0,"e":0,"a":0,"n":0},
"enneagram":{"type":0,"wing":0,"confidence":0.0},
"mbti":{"type":"","axis_confidence":{"ei":0.0,"sn":0.0,"tf":0.0,"jp":0.0}}}
One JSON object per model — a bare array, or objects pasted back-to-back. Schema emitted by the prompt.
Each polygon is one model's read. Tight cluster = consensus. Fan-out = disagreement.
A single read is a snapshot — and a snapshot can be cherry-picked by someone else to build a story about you. The defense is holding your own record. Hit Save this run above to download a small dated JSON file; keep it. Re-run this in a month or three, drop every saved file back here, and watch the trend per trait. The record lives in your files, not in this page or any provider — nothing is uploaded or stored.
Read drift skeptically. A change between two runs has three possible causes you can't fully separate: you changed, the model changed, or it's run-to-run noise (the same model types you differently on different days). Distrust a single jump. Trust a sustained trend across several runs — that's the only movement worth reading as real.
Convergence — multiple models landing on the same score — is your most trustworthy data; treat it as a meaningful read. A wide spread on any single trait means the data is thin or ambiguous: that trait is inconclusive, so don't read into it. This applies to everything — OCEAN scores, Enneagram type, each MBTI letter. If models can't agree on a trait, the issue is the data, not you. The output is a repeatable, comparable baseline across systems rather than one source's verdict. Ask: does this reflect how I actually think and behave? The terms are starting points for your own research — take what works, leave the rest.
The two regimes don't agree, and that's the point. When researchers fine-tune a dedicated classifier on thousands of real chats against ground-truth personality tests, it profiles people meaningfully — the ETH Zurich study below trained RoBERTa models on 62,090 chats from 668 users and beat chance across several traits, with extraversion inference improving ~44% over baseline on relationship and personal-reflection conversations. That's the privacy threat: a party holding your full history, with labels and a trained model, can read you — and the same paper warns this enables profiling, targeted influence, and what its authors call "cognitive surrender." But that is not what this prompt does. Asking a stock chat model to type you from memory, unlabeled and untrained, is the weak regime — peer-reviewed benchmarks put inter-rater agreement near zero (Cohen's κ < 0.10) and validity at r ≈ 0.27, dropping to r ≈ 0.12 in default assistant mode. So: the capability is real enough to fear from a data-holder who trains on you, and weak enough to distrust from a chat box that doesn't. The variance you measure here is the honest version of that gap.