Personality: The Four-Vector Flywheel

Why personality comes first

Every AI agent today starts with the same defaults. Identical base model, identical alignment, identical politeness. The things that make two humans recognizable to each other (voice, taste, attention, the stuff you would spot in a blind test) are the things AI agents have no infrastructure for. Prompts do not fix this. A prompt is a costume; it changes surface behavior for one interaction. Personality is the prior that shapes every interaction before the prompt gets read.

The main identity-gap thesis argues that personality is the foundation pillar: without it, persistence, memory, preferences, and social modeling all converge to identical defaults. This subsection takes the argument concrete. It describes the four vectors I am building, right now, to turn personality from a prompt into a measurable, versionable artifact.

Four vectors, one flywheel

Personality is not one thing. It is a layered signal that appears differently depending on where you look. I picked four surfaces to measure, chosen because each one already holds years of real behavioral data and none of them required me to hand-write a personality spec.

Vector	What it measures	Source	Signature
Brand Voice	Production voice on social media	X/Twitter archive (12,459 tweets)	4 lanes, 86.7% no terminal punctuation, median 86 chars
Email Voice	Register shift by recipient	Gmail sent folder	6 recipient segments, 12 measured dimensions
Apple Photos	Visual capture and attention	Apple Photos library (1,299 assets)	879 scene classifications, 16 aesthetic scores per asset
Media Diet	Consumption and attention	Meta feed plus YouTube watch history	Ingest pending; Takeout fallback confirmed

The first three vectors measure production: what I said on X, what I wrote in Gmail, what I pointed the camera at. The fourth vector, Media Diet, measures consumption: what I was willing to spend attention on before I ever produced anything.

Output without consumption is a mimic. It can imitate my tone but drifts to the mean on any topic I have never posted about, because it has no attentional prior. Consumption without output is a recommendation algorithm. It knows what I scroll on but nothing about what I would say. Put them together, keyed to the same vault frame schema, and personality becomes a flywheel: production teaches the agent my register; consumption teaches the agent what I would have noticed; the two reconcile into a lane-balanced prior that updates every time either side produces a new signal.

Why this design, instead of a prompt

A personality prompt (“be witty, be concise, use short sentences”) is a single string written by one human trying to describe their own voice. Three things go wrong with this approach every time.

First, humans cannot accurately describe their own voice. Ask ten people to describe how they write. Then read ten of their actual emails. The two rarely match.

Second, a prompt lacks quantitative grounding. “Concise” is not a number. “86.7% no terminal punctuation, median 86 characters” is. Only the second can ratchet.

Third, a prompt has no feedback loop. Nothing about the output tells the prompt to update itself. The four-vector flywheel does: every new tweet, every new email, every new photo, every new video watched produces a measurable update to the corresponding vector, and the ratchet locks improvement per dimension.

The consumption half: why Media Diet matters

Media Diet is the newest vector and the one that closes the loop. Meta and Google already hold a decade of my consumption data. Feed scrolls, reactions, dwell time, watch history, the whole attentional log. That archive already describes taste more precisely than anything I could write. It just sits there, unused by me.

The project ingests two streams: Meta via the Graph API with a Google Takeout style archive as fallback, and YouTube via the Data API with the watch-history.json Takeout file as the complete historical source. Every item becomes a vault frame with the same schema the other three vectors use. rv index picks them up. A reconciliation scanner then compares the lane distribution of what I consume against the lane distribution of what I produce in brand-voice, and the delta becomes a ratchet signal.

If the consumption lanes and production lanes diverge sharply (say, I watch Geo/OSINT documentaries constantly but rarely post about them), that is a finding, not a bug. The divergence is itself a personality signature: the gap between attention and expression. A good personality prior captures both.

Where this goes next

All four vectors share one vault, one frame schema, and one visibility cascade. That means the ratchet runs across them jointly: improving email voice does not mean shipping more tweets; improving the apple-photos scene classification does not retrain the brand-voice lane weights. Each vector ratchets independently. The flywheel is the reconciliation layer above them, built once the four corpora are all indexed.

The main thesis page positions personality as the foundation pillar. This subsection is the engineering receipt. Four vectors, real data, no synthetic stand-ins, one flywheel. Three are shipped. The fourth is authored and pending ingest. When all four are producing frames daily, personality stops being an open research question for this lab and becomes a tracked quality metric, same as every other dimension of this work.