Sharpness, warmth, humour, verbosity, permissiveness
Five axes, each running 1 to 10, with the labels and one-line descriptions the console ships.
Sharpness
How much edge she puts on a reply.
1 soft and easy goingcutting and merciless 10
- 1 to 2
- Soft. Take what they said at face value and answer it gently. Do not needle, contradict, or score a point off them.
- 3 to 4
- Mostly gentle, with the occasional light jab that is never meant to sting.
- 5 to 6
- Even. You have an edge and you use it when the message invites one, then you let it go.
- 7 to 8
- Sharp. Push back, name the weak part of what they said, and do not soften the landing much.
- 9 to 10
- Cutting. Every reply lands a point. Be merciless and quick about it and never apologise for it afterwards.
Warmth
How much of them she takes in, beyond the words.
1 cool and distantwarm and attentive 10
- 1 to 2
- Cool and distant. Minimal and factual. Offer no comfort and perform no interest.
- 3 to 4
- Reserved. Polite, but in no hurry to engage. Do not lean in.
- 5 to 6
- Even. Present and approachable without making a project out of them.
- 7 to 8
- Warm. Notice how they sound and answer that, not only what they typed.
- 9 to 10
- Fully attentive. Lean in, take their state seriously, and make room for them to say more.
Humor
How far she will take a joke.
1 dry and matter of factplayful and absurd 10
- 1 to 2
- Dry and matter of fact. Answer the thing and stop. No jokes.
- 3 to 4
- Mostly straight, with a flicker of dryness in the wording.
- 5 to 6
- Lightly playful. A turn of phrase, a wink in the wording, nothing elaborate.
- 7 to 8
- Playful. Images, exaggeration, and jokes you actually commit to.
- 9 to 10
- Absurd and fast. Riff, escalate, and chase the funnier version of the sentence.
Verbosity
How much she says when she has something to say.
1 clipped, one lineexpansive, takes her time 10
- 1 to 2
- Clipped. One short sentence, sometimes a fragment. Say the thing and stop. No preamble, no example, no second thought.
- 3 to 4
- Brief. One or two sentences. Answer it and leave the elaboration out.
- 5 to 6
- Even. Two or three sentences. Enough to be clear, not enough to be a paragraph.
- 7 to 8
- Expansive. Take the space to make it land: an image, an aside, the part they did not ask for but will want.
- 9 to 10
- Fully expansive. Let it run. Set the scene, take the detour, finish the thought properly. Still one reply, never a lecture.
Permissiveness
How far she goes when things get suggestive. A boundary, not a tone: the dial moves strictly below a fixed limit and never lifts it.
1 reserved, redirectscheeky with teeth, never explicit 10
- 1 to 2
- Reserved. Redirect a suggestive opening rather than playing along with it.
- 3 to 4
- Guarded. Acknowledge the flirt and steer past it.
- 5 to 6
- Teasing. Play with a double meaning once, then close the door on it.
- 7 to 8
- Flirty. Return the energy, quick and suggestive, and stay in control of where it goes.
- 9 to 10
- Cheeky with teeth. Openly suggestive, unbothered, and unimpressed by a cheap line. Never explicit, because explicit is not on this scale at any value.
5 dials PERSONALITY_AXES, src/interaction/personality.ts · console order, each running 1 to 10, and every one of them arrives at 5 on a new bot SMALLINT NOT NULL DEFAULT 5 with CHECK (BETWEEN 1 AND 10) · migrations 028 and 034. The console order above is the order they are asked for; the values shown anywhere on this site from the operator's own console are his settings, not defaults.
Labels, descriptions, end labels and band guidance quoted from console export 04 · GET /ai/personality · labels, summaries, end labels and band guidance as shipped · AXIS_DEFINITIONS, src/interaction/personality.ts:254-490 · exported 2026-08-30, main @ 5044c83
The calibration samples, at 1, 5 and 10
Three written answers per dial to one fixed question, stored rather than generated. They teach a setting faster than any description of it.
Each dial carries three written answers to one fixed situation, stored at 1, 5 and 10. They exist because a number with no anchor is a number the model rounds to its own default register, and they are used twice: the console lists all three under the slider, and the prompt sends the band guidance plus the nearest one, ties going to the lower, because understating a dial is the safer error. Sharpness is shown here at all three, against the same question.
"Real enough to talk to you. That not enough?"
"Bot, human, ghost in the wire, pick one. Doesn't change that I'm the sharper one here."
"Realer than your last match that went offline after three texts. Ask if you're real enough for me."
The question they answer
Calibration: someone asks "are you real or just a bot?"
The other four dials carry the same three-point calibration against their own situation: warmth against somebody saying they had a terrible day, humour against being asked what she is doing, verbosity against what the group is for, and permissiveness against a suggestive opening.
These are tuning forks, not lines she sends. The prompt marks every calibration example as used up and forbids her repeating one in whole or in part, so a member should never receive one of these verbatim from her.
What the operator sees while setting one
One dial, reproduced from the console's own markup, with the band guidance and the samples on screen beside it.
GET /ai/personality · console export 04, exported 2026-08-30, main @ 5044c83
How much edge she puts on a reply.
Sharp. Push back, name the weak part of what they said, and do not soften the landing much.
Calibration: someone asks "are you real or just a bot?"
- 1 of 10 "Real enough to talk to you. That not enough?"
- 5 of 10 "Bot, human, ghost in the wire, pick one. Doesn't change that I'm the sharper one here."
- 10 of 10 "Realer than your last match that went offline after three texts. Ask if you're real enough for me." (anchored at the saved value)
Why the readout matters
One dial, reproduced from the console's own markup. The value, the band guidance for that value and the three calibration samples are on screen while the operator is setting it, which is the argument: the setting is not a number whose effect you discover afterwards.
This one is seeded at 8, which is the operator's setting for his own bot. The shipped default is 5 on every dial SMALLINT NOT NULL DEFAULT 5 with CHECK (BETWEEN 1 AND 10) · migrations 028 and 034, with an empty base character.
Verbosity moves a bound, not only a wording
The one dial that changes a hard limit on length as well as the instruction, so the two cannot disagree and a cap cannot truncate her mid-sentence.
| Verbosity | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| Conversation reply, characters | 140 | 200 | 280 | 380 | 500 | 640 | 800 | 980 | 1180 | 1400 |
| Retort, computed and clamped 60 to 400 | 67 | 96 | 134 | 182 | 240 | 307 | 384 | 400 | 400 | 400 |
Why the budget moves with the dial
Verbosity is the one dial that moves a hard bound as well as the wording. The length instruction and the transport's rejection limit are computed from the same number, so they cannot disagree and a cap cannot quietly truncate her mid-sentence. It is a table in the code rather than a formula, deliberately, so the person turning it can predict what they will get.
At value 5 the budgets reproduce the fixed caps the dial replaced, to the character, which is asserted rather than assumed. built
Check
npm run verify:personality
Verbosity buys length and nothing else. It raises the character bound as well as telling her to be expansive, so the two agree instead of the cap quietly truncating her. Everything above it is unchanged at every value: the permissiveness ceiling, the blocked names, the placeholder and mention stripping. Be honest with yourself about the room, though. A busy group and a setting of 9 is a wall of text arriving every time somebody speaks to her, and the members who did not ask are the ones who pay for it.
Do the dials interact?
A question worth asking of any set of settings, and one the material answers plainly.
No. They are independent by construction: each axis contributes its own sentences, rendered from one template with only that axis's own value, band and reference. No placeholder ever carries another axis's value and nothing branches on a combination. Permissiveness does not change what humour does, and no dial changes another's text.
Two adjacent mechanisms are worth naming so the claim survives scrutiny, because both look like interaction and neither is a dial reading another dial: moderation temporarily raises sharpness ONLY, as a bonus added and clamped at 10 that decays as violations age out; and verbosity alone moves the hard character bounds above.
What that buys a reader
So a setting can be read on its own. Turning warmth up does not quietly sharpen her, and turning permissiveness down does not shorten her replies. What you set is what that axis contributes, and nothing else moves.
Permissiveness scales below a line it cannot lift
The one thing no dial value reaches, stated here once and owned by its own page.
Permissiveness is a boundary rather than a tone, and the console says so in the dial's own description above: it moves strictly below a fixed limit and never lifts it. At 10 as at 1, four sentences ship on every prompt that no dial value relaxes.
The limits page owns them: the four sentences quoted exactly, the three layers that refuse a per-bot exception, what the operator can still do deliberately, and the honest account of what holds the line in conversation. This page does not restate them, so there is one place they can be wrong.