Nobody has counted. Here is what psychology measured instead, what a character's file needs, and a simulator to test your own numbers.
“I could give you twenty reasons for the way I work. Three of them matter.”CLARA MERCIER · IF YOU ASKED HER
COF 1.2 gives AI characters a psyche: values, beliefs, wants, needs, fears, wounds, each a node with a weight, a date and a reason, tied to the memories that made it. The obvious question is how many of them a character needs — and how many a person has. This page starts with one refusal, follows it down to where it comes from, then says what the research does and does not know, lets you test your own numbers, and is honest about the popular shortcuts.
Orders of magnitude, not measurements. Every figure on this page says where it comes from.
Clara is a fictional designer from Lyon, the example character of the COF format. A prospective client writes to her. Here is the exchange — invented for this page, but played from her file. Then ask her why, as many times as it takes.
clientCould you ship our whole new site by Friday? Autoplay video on the home page, a chat bubble, push notifications — the works.
ClaraNo. Not like that. I can give you a lighter site by Friday, or this one in a month. But I won't ship something heavy and fast again.
No notifications
taste · strength 60 · settled · valence -80
Hates notifications and interfaces that blink at you.
The first answer is a taste: what she cannot stand. Strong, and quick — but a taste alone would only make her grumble, not refuse.
Lightness
value · strength 95 · core · since age 28
Lightness: do less, better — a page should cost the planet and the reader as little as possible.
Under it, a value, at the core of who she is. This is what turns a dislike into a no. It has a date, 28 — and two reasons.
The agency years
wound · strength 85 · core · since age 24
Four years of shipping heavy interfaces too fast, for clients who never saw the users.
The first reason is a wound. Four years she does not tell as a story, but that shapes how she works.
The agency years
memory · a period · age 24–28
Four years at a digital agency in Lyon, from 24 to 28: heavy interfaces shipped too fast, late nights of testing, and the slow conviction that the web could be lighter.
And under the wound, the memory itself. Here the chain ends: not in an adjective, in a life.
Five questions, from a sentence to four years that began at twenty-four.
Every answer above is a node of Clara's file, linked to the next — supports, because — written once by her author, readable by any tool, the same in every model that plays her. Nothing here was improvised.
Five nodes answered. Clara's file holds 23, with 31 links between them and to her memories. Is that a person? Clearly not all of one. So how many would it take?
No study has counted the convictions of a person, and none could: “a conviction” is not a unit the mind keeps. What psychology has counted is the number of dimensions it takes to describe how people differ. Those numbers are small, and they are the closest thing to an answer.
Milton Rokeach separated beliefs, attitudes and values, and argued that a person holds far fewer values than attitudes: values are the few central beliefs the rest are organised around. His survey asks people to rank 36 of them, 18 ends and 18 means (Rokeach, 1973). Shalom Schwartz then proposed ten basic values recognised across cultures — security, benevolence, achievement and the rest — arranged in a circle of compatible and conflicting motivations, and tested the structure in twenty countries (Schwartz, 1992); the refined theory distinguishes nineteen, confirmed in fifteen samples from ten countries (Schwartz et al., 2012). A person is not ten values: a person is an order among them. That is why COF's value nodes carry a strength, and why the RFC names Schwartz's list as an optional vocabulary, never as a form to fill.
The lexical hypothesis holds that whatever matters in how people differ ends up as a word. Gordon Allport and Henry Odbert went through Webster's dictionary — about 400,000 entries — and kept 17,953 terms that describe personality or behaviour, 4,504 of them for stable traits (Allport & Odbert, 1936). Decades of factor analysis condensed that vocabulary into five broad domains — the Big Five — which the standard questionnaire measures through 30 facets, six per domain, with 240 items (Costa & McCrae, 1992). And when sociologists asked people simply to answer “Who am I?”, they gave them twenty lines to fill (Kuhn & McPartland, 1954) — a format, not a ceiling, but a telling one.
What has been counted, on one logarithmic scale — not counts of the same thing: dimensions to describe everyone, items of a questionnaire, words of a language, and in violet COF's own numbers. The range of a novel's or a life's psyche sits between what a questionnaire asks and what a dictionary holds.
Dan McAdams answered the question “what do we know when we know a person?” with three levels (McAdams, 1995; McAdams & Pals, 2006). Traits, broad and context-free: the Big Five, which give the “psychology of the stranger”. Characteristic adaptations: goals, values, concerns, coping strategies, tied to a time, a place, a role. And narrative identity: the evolving story a person makes of their life, joining the remembered past to the imagined future (McAdams & McLean, 2013).
COF's psyche lives on the second and third levels. Traits stay where they were in 1.0, an optional annex of scores. A node is a characteristic adaptation — a want, a fear, a stance toward someone — and its because link to a memory is a piece of the life story. Memory research describes the same two-way tie: in Martin Conway's self-memory system, what we remember is selected and shaped by the goals of the current self, and the self in turn is grounded in what we remember (Conway & Pleydell-Pearce, 2000; Conway, 2005). Memories without a psyche stay anecdotes; a psyche without memories stays slogans.
Fiction has its own counts, and they are small at the core. John Truby builds a character on a weakness and a need the hero does not see, a desire the hero pursues, and often a ghost — “the event from the past still haunting the hero in the present” (The Anatomy of Story, 2007). K. M. Weiland adds the lie the character believes, formed by that ghost and undone by the story (Weiland, on want and need; ghost, wound, lie, weakness). One wound, one lie, one want, one need: four nodes carry an arc. COF has a kind for each — wound, belief marked false but held, want, need — and everything a novelist knows besides: the tastes, habits, stances that make a character recognisable in a scene where the arc is not at stake.
A person's psyche is not a snapshot. A meta-analysis of 92 longitudinal samples found that personality traits keep changing through adulthood, most between twenty and forty, some into old age (Roberts, Walton & Viechtbauer, 2006). And more than 19,000 people, asked about their personality, their values and their preferences, recalled a lot of change over the past ten years and predicted little over the next ten, at every age — the end of history illusion (Quoidbach, Gilbert & Wilson, 2013). That is why a node has a since and an until, and why a conviction that changes leaves its earlier self in the file.
So there is no number to take from the literature, only bounds: a handful of dimensions to describe anyone, a few dozen values at most, thousands of words for the rest. Here is ours, from building Clara and from load tests of the tools — an estimate, labelled as one.
Every assumption is a control. The presets reproduce the estimate above; Clara's is fitted on her file — 23 nodes written by hand, 21 held at 65, two let go at forty-five. Change any of it.
Logarithmic scales. 1 KB = 1,000 bytes. Expected values of a simple model, not a measurement of anyone: hover or focus the chart and use the arrow keys to read it year by year.
Two things stand out whatever the settings. The file stays small: a thousand convictions in two languages weigh under a megabyte — three pictures like the one of her wedding, a seventeenth of her file. And what a model must read per turn barely moves: an agent picks the core nodes and the handful the moment calls for — by tag, by the person or memory the conversation touches, by strength — and leaves the rest on disk. Size is not the constraint; writing them is.
Many authors do not start from a blank page but from a type: “she's an INFJ”, “a Four”, “very open, not very agreeable”, “a Virgo”. COF's tools take all four as starting points. Their scientific standing is very different, and it is worth saying so plainly — without contempt for anyone who finds them useful. A shorthand that helps someone think about a character is doing a job.
The five domains that came out of the dictionary — openness, conscientiousness, extraversion, agreeableness, neuroticism — are the best-validated model in personality psychology. They are scores, not boxes: everyone is somewhere on each. Their links to what happens in a life hold up: a large replication of 78 published associations between traits and life outcomes reproduced about 87 % of them, though usually with smaller effects than first reported (Soto, 2019). Their limit, for a character, is that they say how much, never about what: a high neuroticism score does not say what she is afraid of.
Four letters for sixteen types, from Jung's theory, hugely popular in workplaces and online. Research finds that the four scales behind the letters measure something real — four of the five Big Five domains — but not types: the scores are not bimodal, they do not split people into two camps on each scale (McCrae & Costa, 1989). Near the middle a letter is decided by a few answers, and it often flips: among people with intermediate scores, a quarter to a third got the other letter on a given scale when retested a month later (Howes & Carskadon, 1979, reviewed by Pittenger, 2005). A four-letter type is a less stable summary than the scores beneath it.
Nine types, each defined by a basic fear and a basic desire, widely used in coaching and spiritual direction. A systematic review of 104 samples found mixed evidence of reliability and validity: some subscales relate to the Big Five in the expected ways and several studies report that people find it helpful for personal growth, but factor analyses usually recover fewer than nine factors, and little supports the wings and movements between types (Hook et al., 2021).
A sun sign is the oldest personality shorthand there is. In controlled tests it has no predictive power. In a double-blind experiment published in Nature, astrologers matched birth charts to one of three personality profiles from a standard questionnaire, and got it right about one time in three — what chance gives (Carlson, 1985). A review of more than forty controlled studies found astrologers doing no better than chance, and people born minutes apart showed none of the similarities astrology predicts (Dean & Kelly, 2003). One finding matters for a character, though: the sign can still shape the person. In a sample of 2,324 adults, people born under the odd-numbered signs scored as more extraverted on a standard questionnaire, just as astrology says (Mayo, White & Eysenck, 1978) — and the effect disappeared among people who did not know what their sign was supposed to make them (Eysenck & Nias, 1982). The horoscope had done the work: a belief about oneself, repeated, becomes behaviour. In a story, a character who reads hers every morning is a fact about her — a node, often aware: no, with a because on the habit.
| what it gives | what the evidence says | what COF proposes from it | |
|---|---|---|---|
| Big Five | five scores, 0–100: openness, conscientiousness, extraversion, agreeableness, neuroticism | the best-validated model in personality psychology; links to life outcomes largely replicate | a score of 70 or more, or 30 or less, proposes one or two candidates for that pole |
| Myers-Briggs type | four letters, sixteen types | the four scales track four of the Big Five; the types themselves are not well supported, and a letter near the middle often flips on retest | three or four common readings of the type: a value, a taste, a habit, a fear or a want |
| Enneagram | one type out of nine, with its basic fear and desire | mixed evidence of reliability and validity; the nine types are not recovered as nine factors | the basic fear, the basic desire, a value and a habit of the type |
| Astrology | a sun sign, twelve | no predictive validity in controlled, double-blind tests | three or four traits commonly attached to the sign |
So COF uses all four the same way: as a proposal, never as a field of the format. Give the builder or cof psyche propose a sign, a four-letter type, an Enneagram number or five scores, and it offers a few ordinary nodes — strength 50, stability shifting, tagged with where they came from — every one unticked until the author reads it and decides. There is no “type” field to fill: millions of people share a sign or a type, and a type cannot say why. A person is a graph with reasons. The shorthand is a fine place to start the graph, and a poor place to stop.
There are three ways to tell a model who a character is. A description: a few lines in a prompt. A graph: the psyche of COF, convictions linked to each other and to memories. Weights: a model fine-tuned on the character, or a LoRA — the character dissolved in a latent space.
| a description | a graph | weights | |
|---|---|---|---|
| what it holds | a portrait in a few lines | convictions, with weights, dates, reasons and the memories behind them | habits of speech, reflexes, a style no text pins down |
| can you read it? | yes | yes — node by node | no — millions of numbers |
| can you correct it by hand? | yes | yes, one node at a time | only by training again |
| can it answer why? | rarely — it improvises | yes, link by link | no — it can only act |
| can it keep a secret, or not know itself? | everything is on the page | open, guarded, secret; aware or not | not on purpose |
| does it move with time? | one snapshot | since, until, revisions | one training run |
| size | about a kilobyte | Clara: 12 KB for 23 nodes; a whole modelled life, a few hundred KB | tens to hundreds of megabytes |
| what it takes to make | writing it by hand | writing it by hand, or a tool that proposes and you tick | data, GPUs, hours of training |
| works with | any model | any model, today | one model family |
Orders of magnitude, not measurements. The size of weights depends on the model and the method; the graph's on how much of the life is written.
The honest ranking is not a ladder. A graph is better than a description: it can be selected from, it answers why, it keeps secrets and moves with time. It is less rich than a latent space, which holds what no list of convictions can — the way a sentence turns, a hesitation, a kind of humour. But a graph is the only one of the three that anyone can make with a pen or a tool that proposes, read node by node, edit one conviction at a time, and carry to any model, today and in ten years.
Weights are richer, and nothing on this page competes with that. But they are mute about themselves: they can produce the no, never the why. You cannot read them, correct one conviction, keep one secret, or take them to another model.
So the two are not rivals. A graph is the score; weights are a performance of it. The score can be read, argued with, versioned, carried to any orchestra — and one day it can be what a model is tuned from, and what its performance is checked against. Until then, it is the only one of the three that you can write by hand tonight, and that every model can read tomorrow.
In practice the two work best together. Weights give the voice its grain; the graph gives it something to hold on to — buoys in a channel: the current does the moving, the buoys keep it from drifting. Prompt the graph into a model you fine-tuned on the same character, and each corrects the other's failure mode: the weights stop improvising convictions the file never gave, the graph stops reading like a list.
That last sentence is a research question, and we would like someone to take it on. Can a psyche graph be the supervision signal for a character's weights — and can it be the evaluation set?
Each node already reads like a test. A core value with strength 95 predicts a refusal; a fragile stance predicts a change of mind within one conversation; a node marked aware: no predicts behaviour the character never names; a because chain predicts the answer to “why?”. So here is the study. Take a few dozen characters with authored graphs of a hundred nodes or so. Turn each node into situations — a client in a hurry, a friend who asks for help — with the expected behaviour derived from its kind, strength and stability. Then measure three things. Whether a model prompted with the graph behaves as the nodes predict, and how that degrades as only a selection of nodes is loaded. Whether an adapter trained on dialogue generated from the graph recovers the because chains when asked, without being shown them. And whether editing one node and retraining changes only the behaviours that node predicts, and nothing else. If the answers are good, the graph becomes what the reference images of a face already are for a LoRA in COF: the readable source a model is derived from, regenerated when the source changes.
Rokeach, M. (1973). The Nature of Human Values. Free Press — the Rokeach Value Survey.
Schwartz, S. H. (1992). Universals in the content and structure of values: theoretical advances and empirical tests in 20 countries. Advances in Experimental Social Psychology, 25, 1–65. doi:10.1016/S0065-2601(08)60281-6
Schwartz, S. H. et al. (2012). Refining the theory of basic individual values. Journal of Personality and Social Psychology, 103(4), 663–688. doi:10.1037/a0029393
Allport, G. W. & Odbert, H. S. (1936). Trait-names: a psycho-lexical study. Psychological Monographs, 47(1). Summary: the lexical hypothesis.
Costa, P. T. & McCrae, R. R. (1992). NEO PI-R Professional Manual. Psychological Assessment Resources — description of the instrument.
Kuhn, M. H. & McPartland, T. S. (1954). An empirical investigation of self-attitudes. American Sociological Review, 19(1), 68–76. doi:10.2307/2088175
McAdams, D. P. (1995). What do we know when we know a person? Journal of Personality, 63(3), 365–396. doi:10.1111/j.1467-6494.1995.tb00500.x
McAdams, D. P. & Pals, J. L. (2006). A new Big Five: fundamental principles for an integrative science of personality. American Psychologist, 61(3), 204–217. doi:10.1037/0003-066X.61.3.204
McAdams, D. P. & McLean, K. C. (2013). Narrative identity. Current Directions in Psychological Science, 22(3), 233–238. doi:10.1177/0963721413475622
Conway, M. A. & Pleydell-Pearce, C. W. (2000). The construction of autobiographical memories in the self-memory system. Psychological Review, 107(2), 261–288. doi:10.1037/0033-295X.107.2.261
Conway, M. A. (2005). Memory and the self. Journal of Memory and Language, 53(4), 594–628. doi:10.1016/j.jml.2005.08.005
Roberts, B. W., Walton, K. E. & Viechtbauer, W. (2006). Patterns of mean-level change in personality traits across the life course: a meta-analysis of longitudinal studies. Psychological Bulletin, 132(1), 1–25. doi:10.1037/0033-2909.132.1.1
Quoidbach, J., Gilbert, D. T. & Wilson, T. D. (2013). The end of history illusion. Science, 339(6115), 96–98. doi:10.1126/science.1229294
Truby, J. (2007). The Anatomy of Story: 22 Steps to Becoming a Master Storyteller. Faber and Faber.
Weiland, K. M. (2016). Creating Character Arcs. PenForASword — and, online, the thing your character wants vs. the thing your character needs and your character's ghost vs. wound vs. lie vs. weakness.
Soto, C. J. (2019). How replicable are links between personality traits and consequential life outcomes? The Life Outcomes of Personality Replication Project. Psychological Science, 30(5), 711–727. doi:10.1177/0956797619831612
McCrae, R. R. & Costa, P. T. (1989). Reinterpreting the Myers-Briggs Type Indicator from the perspective of the five-factor model of personality. Journal of Personality, 57(1), 17–40. doi:10.1111/j.1467-6494.1989.tb00759.x
Pittenger, D. J. (2005). Cautionary comments regarding the Myers-Briggs Type Indicator. Consulting Psychology Journal: Practice and Research, 57(3), 210–221. doi:10.1037/1065-9293.57.3.210
Hook, J. N. et al. (2021). The Enneagram: a systematic review of the literature and directions for future research. Journal of Clinical Psychology, 77(4), 865–883. doi:10.1002/jclp.23097
Carlson, S. (1985). A double-blind test of astrology. Nature, 318, 419–425. doi:10.1038/318419a0
Mayo, J., White, O., & Eysenck, H. J. (1978). An empirical study of the relation between astrological factors and personality. The Journal of Social Psychology, 105(2), 229–236. doi:10.1080/00224545.1978.9924119
Eysenck, H. J., & Nias, D. K. B. (1982). Astrology: Science or Superstition? Maurice Temple Smith, London. The self-attribution reanalysis: the sign–extraversion link holds only among people who know what their sign predicts.
Dean, G. & Kelly, I. W. (2003). Is astrology relevant to consciousness and psi? Journal of Consciousness Studies, 10(6–7), 175–198. ingentaconnect
OpenAI Help Center, What are tokens and how to count them? — “1 token ≈ 4 characters in English”, the rule the simulator uses (the same as cof tokens without a tokenizer).
Clara's psyche, spec/examples/10-clara-mercier, COF 1.2: 23 nodes, 12.0 KB of node files, 15.9 KB with their manifest entries, 31 links, 9 nodes tied to a memory; her « Who she is » as cof psyche text --preset en-agent renders it, 1,943 characters. The upper bound: generated psyches of 150, 500 and 1,000 nodes, opened in the explorer (load tests, 3 November 2026).
She said no because of four years she does not talk about. Now you know why. So does any tool that opens her file.
