Predictive Profiling at LIFE

Predict the user's reaction, then personalise around it

LIFE's teacher, Wisdom, takes an action every turn (a practice, a tone, a question). The predictive‑profiling programme builds a per‑user profile that predicts how that specific person will react — settle or get activated, accept or push back, complete or drop, return or churn — so Wisdom can adapt safely and so we can measure what actually helps.

The strategic claim. The most valuable predictive companies in the world (Meta, TikTok/ByteDance) win by modelling each user's own behavioural history as a sequence and predicting the next action — not by leaning on static demographics or one‑time questionnaires. Decades of psychometrics show why: traits predict your average tendency, not how you'll respond right now. LIFE is making the same behaviour‑conditioned bet, in a domain (contemplative practice) where no one has yet measured it.

Across the full scope of life outcomes, the best tests explain a minority of the variance

If a personality test or IQ score reliably told us how someone will behave, we wouldn't need this programme. The evidence — from decades of meta-analyses spanning work, school, physical health, mental health, relationships, wealth and wellbeing — says they don't. Here is the predictive validity (the correlation r, and the share of variance it explains, r²) of the best psychometric predictor in each domain, from the largest available meta-analyses.

Most r-values are "corrected" (operational) validities — the field's optimistic, measurement-adjusted estimates; the raw correlations you'd see in live data are typically lower still. †Mortality is reported as a hazard ratio per standard deviation (its native metric); the individual-level effect is small (rough r-equivalent < .10). IQ is the strongest psychometric mortality predictor — far larger than any personality trait.

By trait: the Big Five across outcomes

Drilling into personality specifically (since that's what a profile would carry): only Conscientiousness predicts across the board, and even it is weak. The average Big Five trait predicts job performance at just r ≈ .13 — near the floor of usefulness.

Job-performance ρ: Wilmot & Ones 2021 (corrected; "Emotional Stability +.13" shown here as Neuroticism −.13). University grades r: Poropat 2009 (corrected). Relationship satisfaction r: Malouff et al. 2010 (observed; own-trait → own satisfaction). Wellbeing/mortality roles summarised in the last column rather than tabulated, as those come from different metrics (multiple-R and hazard ratios).

The pattern is unmistakable. Outside one or two exceptions, psychometric predictors of real life outcomes cluster at r ≈ 0.10–0.30 — roughly 1–9% of the variance, leaving 90%+ unexplained (Roberts et al. 2007; replicated at ~¾ strength by Soto 2019). The canonical comparative review found personality, IQ and family background "indistinguishable" in their (modest) power to predict mortality, divorce and career attainment — a low ceiling shared by all the static predictors at once.

The two "big" numbers are not what they look like

• IQ → school grades (r ≈ .54) is the genuine high point — but it largely predicts an outcome (graded tests) built from the same cognitive skills the IQ test measures.
• Personality → wellbeing (~39% of variance; Steel et al. 2008) only reaches that height when you combine all five traits and predict a self-report (life satisfaction) from self-reports (personality) — shared method, an optimistic corrected ceiling, not a behavioural forecast.
• Neuroticism → mental health looks enormous concurrently (r ≈ .64 for "who is anxious/depressed right now"; Kotov et al. 2010) but collapses to r ≈ .1–.2 as a forward predictor of who will become depressed, once you adjust for what you already know (Jeronimus et al. 2016, N = 443,313). The gap between "describes the present" and "predicts the future" is the whole game.

And the missing variance is real, not measurement noise

The natural objection — "the tests are just noisy" — is wrong, and that's the crux. IQ tests are among the most reliable instruments in psychology (test-retest ≈ 0.90–0.95). So the 90%+ they leave unexplained is genuine situational, in-the-moment variation. The landmark finding: 62–93% of the variability in how "conscientious" or "extraverted" someone actually behaves is within a single person — the same individual acting differently hour to hour — not between people. Fleeson, 2001/2007. A person differs from herself across occasions more than she differs from other people.

We are not claiming psychometrics are worthless — their aggregate effects rival those of socioeconomic status (Roberts et al., 2007), and they're genuinely useful for nudging priors across many people over long horizons. We're saying they are population-level actuarial instruments with a low ceiling, near-useless for calling one person's next reaction — so we complement stable traits with a behaviour-conditioned layer that targets exactly the in-the-moment variance they structurally cannot reach.

In our own domain the static-trait story is even weaker — and the field is wide open

Narrowing from "life outcomes" to meditation, the honest headline is a null, and that is the most important finding: the best meta-analysis of who-benefits-from-meditation found that personality traits did not reliably moderate outcomes at all. The only robust baseline predictors were how distressed someone already was and how motivated they were — not their personality.

Who benefits, who sticks — the evidence is thin and mostly null

• Who benefits: personality was the only non-significant predictor category; the robust signals were baseline depression severity (r≈.22, → worse) and motivation (r≈.23, → better). Buric et al. 2022, Br. J. Health Psych. A few tiny RCTs (n=56) hint at personality×meditation-type matching, but they're too small to trust. Furnell et al. 2025
• Who drops out: the largest attrition meta-analysis (114 studies, n=11,288) found essentially no baseline, demographic, or personality variable predicts who quits. Lam, Kirvin-Quamme & Goldberg 2022
• The one modest real signal: Conscientiousness predicts home-practice adherence (β≈.21–.33, but explains only 13–24% of the variance, single n=96 study; Canby et al. 2021), consistent with Conscientiousness (r≈.32) and Neuroticism (r≈−.4) being the strongest trait correlates of dispositional mindfulness (Giluk 2009). These are correlations and small single studies — not a deployable predictor.

The engagement cliff — the commercial reality

Regardless of personality, meditation/mental-health apps shed almost all users fast: ~3.3% day-30 daily-active retention across 93 apps (Baumel et al. 2019); ~25–39% attrition even inside controlled trials (Linardon 2023); a Calm research cohort had ~42% still meditating by day 350 (Fowers et al. 2022). This is the number a behaviour-conditioned signal would need to move — and one nobody has published a fix for.

Median daily-active retention across 93 mental-health apps (Baumel et al. 2019). The cliff is near-universal and personality does not explain who falls off it.

Has Headspace, Calm or anyone else published this? No.

The published meditation-company science is overwhelmingly outcome RCTs ("does the app reduce stress/anxiety?"), heavily concentrated in Headspace. On the question that matters for us — has any company published a predictive model of meditation behaviour or a deployed personalization system on its own users — the answer is essentially no.

The only published "who-benefits" personalization model in the space (a Personalized Advantage Index, r²≈.10) is for the nonprofit Healthy Minds Program, not a commercial app. Webb et al. 2022

How the world's best predictive companies actually do it

This isn't a speculative bet. The two companies with arguably the largest behavioural datasets on earth independently converged on the same answer: model the user's own action sequence and predict the next action; adapt fast from live behaviour.

Meta's documented progression — the winning move was modelling each user's own action sequence (the same mechanism as a language model predicting the next word); TikTok reached the same place via real-time behavioural adaptation. Zhai et al. 2024.

Meta — from static features to "Actions Speak Louder than Words"

• 2019: Meta open‑sourced DLRM, the classic "predict engagement from features" model. Naumov et al., arXiv:1906.00091.
• Production ranking became multi‑task: one model predicts many engagement events at once — P(click), P(like), P(see‑less)… — blended into a single value score. Engineering at Meta, 2023.
• 2024 — the consequential shift: Meta's ICML paper "Actions Speak Louder than Words" recasts recommendation as predicting the next action from a user's behavioural sequence — the same mechanism as an LLM predicting the next word. It obeys LLM‑style scaling laws, reached 1.5 trillion parameters, and delivered a +12.4% lift in online A/B tests at a platform serving billions of users. Zhai et al., ICML 2024, arXiv:2402.17152.

The title is the thesis: a person's behaviour is a stronger signal than their stated attributes. (Note: the +12.4% is the real online business lift; a separate 65.8% figure in the paper is an offline benchmark — we keep those distinct.)

TikTok / ByteDance — predict from live behaviour, adapt by the minute

• ByteDance's Monolith system is built for online learning: it refreshes the model from user feedback at minute‑level cadence, explicitly because user behaviour is non‑stationary ("concept drift") and recent behaviour predicts behaviour change better than stale models. Online training beat batch training in every case they evaluated. Liu et al., ByteDance, arXiv:2209.07663, RecSys 2022.
• TikTok states publicly that the feed is driven primarily by behaviour — finishing a video matters a lot; follower count is "not a direct factor." Every interaction teaches the system your interests. TikTok Newsroom.

Careful scope: Monolith is ByteDance's training/serving infrastructure, not a published spec of the full "For You" ranker; the public scoring formula is a simplified internal explainer. The directional lesson is rock‑solid and on the record: behaviour beats declared preference, and fast adaptation wins.

Behaviour is far more predictable than language — and the individual matters

People are predictable where it counts. Human routine behaviour sits near a high predictability ceiling, and a stable model of a person captures most of the achievable signal — strong evidence that a persistent profile (not just the live conversation) carries real value. The newest and most direct result: Centaur, an AI fine‑tuned on ~10.7 million real human choices, out‑predicted bespoke scientific models of human cognition in 159 of 160 experiments on people it had never seen, and generalised to new tasks. Binz et al., Nature 2025. Conditioning on real individual behaviour has genuine headroom.

Two honest caveats. These ceilings are for regular behaviour (where someone goes, whether they return); free‑form text is much less predictable, which is exactly why our headline targets are behaviour and reaction type, not the literal words. And the "95% from your social network" result is also a privacy red flag we take seriously: a platform could profile someone from their contacts even if they share nothing.

Persona collapse: sounding personal ≠ being personal

The obvious shortcut — "just ask a frontier model to be this user" — fails in a documented, measurable way. Asked to role‑play a specific person, today's best models collapse toward a generic, average persona: they flatten individual quirks and minority views, and resist deviating from their default voice.

• On a benchmark of evolving individual preferences, several frontier models scored below random chance (one well‑known model at ~13.5% where chance is 20%) — they confidently predicted a stale, average preference. The best model reached only ~53%. HorizonBench, 2026; landscape review.
• Independent work shows complete "persona collapse" on cognitive tasks for some frontier models, while others retain individual variation — it's model‑ and method‑dependent. Wang et al., Nature Machine Intelligence 2025; Santurkar et al., ICML 2023.

The lesson is decisive for our roadmap: gains come from method and representation — conditioning on a person's real history — not from a bigger base model. A profile that reads beautifully ("Sarah, a cautious beginner who values gentleness") is worthless if swapping in a different user's profile predicts Sarah's reactions just as well. That failure is invisible to the eye and visible only to the right test — which is why we built one.

Rich profile, narrow probes, brutal baselines

The design (reviewed independently, including by an external frontier model that caught three real flaws) rests on a few decisions that a non‑specialist can hold onto:

What we predict — and where we aim

Words → Deeds → Meaning → Felt outcome → Better choice. Language (L1) is near-unpredictable, so we aim at behaviour and reaction-type (L2/L3); felt-outcome and action-selection are deliberately later phases.

The testbed we've specified (built on tooling we already have)

To prove the approach on realistic, controllable users before touching production data, we'll extend the existing review UI ("Data Lab") into a persona‑driven testbed. As a reviewer works a conversation against a chosen persona, the profiler predicts the persona's next reaction; the prediction is logged and graded against the reviewer's human‑validated ground truth. Crucially this also produces four high‑value data products from one workflow: prompt/context‑engineering data for our agents, model‑training sets (long‑term), profiler labels, and Wisdom regression tests. Status: specified, not yet built; most of the infrastructure (personas, branching, model routing) already exists.

A subtle but essential discipline carries through: the persona's true disposition is the answer key used to generate ground truth — it is shown to the reviewer but never to the profiler, which must work only from the observable history. In production you never have the answer key; the whole point is recovering it from behaviour.

How it works, end to end (in the LIFE app)

Two coupled loops. A live loop runs inside the app on every message; an offline loop runs in the background between sessions to build and update the profile, which then feeds back into the live loop. Nothing changes for the user in Phase A — the profile is observed and scored, not yet used to steer Wisdom.

The same loop, worked through one exchange

A returning user with a profile that has learned two patterns: "reacts against directive tone" and "prefers short practices when drained."

In the live product (the later, ethics-gated phase) Wisdom would act on this — offering the 2-minute grounding it has learned this user completes when drained, in invitational language, instead of the 10-minute directive scan. In Phase A we simply prove we could have predicted the reaction.

The measuring instrument works — proven on synthetic data

We built the eval harness and ran it through a three‑way self‑test on data where we control the truth:

On the signal world (120 users, 6,328 held‑out events), the headline results:

Stress‑tested. Five independent "skeptic" reviews tried to prove the GO was an artefact and found no blocking flaws; they independently confirmed the confidence intervals are honest (a naïve method would have badly overstated certainty), and caught and fixed a genuine "peeking at the future" bug in the data adapter. The leak audit confirmed the result was not explained by cheating (the edge was just as strong on ordinary events as on "spoiler" ones), and a deliberately‑rigged world confirmed the audit catches leakage when it's there.

One honesty note we insist on. We do not headline the eye‑catching "+1.0 bits" individuation figure — it's inflated because a wrong profile is confidently wrong, which the scoring punishes heavily. The figures we stand behind are the conservative own‑vs‑own number (+0.063 bits) and a fairness‑corrected individuation (~+0.40 bits).

The five terms behind every number above

What it takes to get the real answer — and what could go wrong

The three unblocks between "instrument built" and "product‑validated"

• Authorised production reads. The real data lives in production mental‑health records; our safety tooling correctly blocks ad‑hoc access. This needs an explicit, governed approval.
• A consented cohort. Only users who have opted in to research data sharing — the lawful basis. The gate already exists in the data model.
• A human labelling pass. Automated reaction labels are a useful proxy, not ground truth. A reviewer pass (which the testbed is designed to produce as a by‑product) is the gate to a trustworthy first readout. This is the main human‑time cost to scope; it doubles as training/eval data.

What would change our mind (the NO‑GO criteria)

We've pre‑committed to killing or redesigning if, on real data: the profile's edge vanishes against the cheap per‑user baseline; a wrong profile predicts as well as the correct one (collapse); the lift comes only from "spoiler"/leakage events; or the profile improves engagement while worsening the wellbeing/activation‑risk signal. Stating these up front is what keeps the programme honest.

The bet, the proof, and the ask

The bet: behaviour‑conditioned, continually‑updated profiles predict a person's reactions far better than static traits — the same bet Meta and TikTok validated at planetary scale — and in contemplative practice no one has measured it yet, so a clean result would be genuinely novel and defensible.

The proof so far: we've built and independently stress‑tested the measuring instrument and shown it is rigorous and honest — it detects real signal, reports a true null, and catches cheating — on synthetic data.

The ask: a governed green‑light for the real‑user readout — authorised access to a consented cohort plus a scoped human‑labelling pass — to convert "the ruler works" into "here's how predictable our users actually are, and where personalising will and won't help."