Odessa RCT — does personalized fiction change lives beyond journaling alone?

The claim under test: personalized AI fiction, layered on top of journaling, improves well-being more than journaling alone. Not "fiction helps" (we know that). The sharp, fundable, falsifiable question is the marginal effect of Odessa on top of an already-active control.

This is a study protocol, written to be pre-registered. It runs on our Citizen Science Platform rails (see "Reuse vs build"). Upstream: ODESSA_NORTH_STAR.md.

The scientific lineage (why this isn't starting from zero)

We are not inventing the idea that writing heals. James Pennebaker's expressive writing paradigm is a 35-year, hundreds-of-studies literature showing that writing about emotional experience measurably improves psychological and even physical health. That is the giant whose shoulders we stand on — and it is also our control condition. Our novel contribution is narrow and testable: does turning that writing into personalized fiction the person can see themselves inside add measurable benefit over the writing alone? If yes, we've extended a canonical literature into the AI era. If no, that's a real and publishable finding too.

Hypotheses (pre-registered before any data)

H1 (primary): Participants who receive Odessa show greater improvement in well-being (T1→T2) than control, between-arms.
H2 (secondary): Odessa increases self-insight (you see your life more clearly) more than control.
H3 (mechanism / exploratory): Within the Odessa arm, "braver" branch choices predict subsequent shifts in journal language toward agency and positive reappraisal — the North Star's loop-closing signal, measured.

Design — 2-arm RCT with a journaling run-in

        Day 0            Day 30                       Day 60
 ┌────────┬───────────────┬────────────────────────────┬────────┐
 │ CONSENT│  both arms     │  TREATMENT: receives Odessa │  both  │
 │ +intake│  journal       │  story + episodic loop      │ final  │
 │  T0    │  (run-in)  T1  │  CONTROL: keeps journaling   │  T2    │
 └────────┴───────────────┴────────────────────────────┴────────┘
   measure         measure                                measure

Why the run-in (days 0–30, both arms journal) is the smart part — and it's your instinct: it gives every participant a within-person baseline trajectory before Odessa enters at day 30. That controls for journaling's own effect, for novelty, and for regression to the mean. The primary test is the difference in slopes after day 30 — does the Odessa arm bend upward relative to a control that's doing the exact same journaling.

Outcome measures

Validated instruments at T0 / T1 / T2, plus continuous in-product signal:

Primary: WHO-5 Well-Being Index (5 items, free, validated, ideal for repeated measures) — or WEMWBS if we want a richer well-being construct.
Secondary: Self-Reflection & Insight Scale (insight subscale); a short psychological-flexibility measure (AAQ-II); a connection/loneliness item set.
In-product, objective, continuous (our edge): language markers in the journals themselves — positive-emotion words, cognitive-processing words, and self-distancing pronoun shifts. This is exactly the Pennebaker measurement tradition, and it's a cheap, objective, high-frequency outcome our competitors can't collect because they don't hold the journal.
Exploratory mechanism: branch-choice "bravery" → journal-delta agency (H3).

Deliberately NOT clinical (a strategic + ethical call)

v1 measures well-being / flourishing, not disease. We do not use PHQ-9 / GAD-7 or claim to treat depression or anxiety. Reasons: (1) it keeps us out of FDA/FTC medical-claim territory; (2) it lowers the ethical/safety burden (no routine screening for suicidality); (3) "helps people flourish" is true to the ethos and still a major result. If a later, IRB-heavier study wants clinical endpoints, that's a separate protocol with a crisis-referral safety plan.

Sample size — and why we pilot first

Honest math: to detect a medium between-arms effect (d≈0.5) at 80% power, α=.05, two-tailed, you need ~64 completers/arm. A more realistic small-medium effect (d≈0.35) needs ~130/arm. With 30–40% attrition over 60 days, a well-powered confirmatory study is ~150–400 enrolled. That's real money in paid recruitment.

So: pilot first. A feasibility pilot (~30–40/arm) does three things a big study can't yet: proves the rails work end-to-end, measures our actual effect size and attrition, and de-risks the spend. Then we pre-register a confirmatory RCT powered on the pilot's observed effect. This is the credible path and the cheaper one.

Randomization, blinding, integrity

Randomization: simple or stratified (by baseline well-being) at enrollment, seeded + reproducible, stamped on the enrollment row.
Blinding: participants can't be blinded (they know if they got a story); analysts are. Outcome scoring + the primary analysis run on arm-masked data.
Pre-registration on OSF before data collection — hypotheses, primary outcome, and analysis plan locked. This is the single biggest credibility lever we have, and it's nearly free.
The integrity line we hold even though we're invested: we are testing Odessa, not confirming it. A pre-registered null result — "Odessa did not beat journaling alone" — is an honest, publishable outcome, and we ship it as readily as a win. The moment we'd p-hack toward the story we want is the moment the result becomes worthless. Pre-registration is how we protect ourselves from ourselves.

Ethics

Human-subjects psychological research on highly sensitive data (personal journals) with compensation. Non-negotiables: explicit informed-consent flow (not just a row stamp), an independent IRB review (commercial IRBs exist for non-academic research; required for publication and the right thing regardless), research-use + de-identification covered in consent, and the right to withdraw + delete. Distress-handling resources surfaced even in the well-being framing.

Analysis plan (pre-specified)

Intention-to-treat. Linear mixed-effects model on the repeated well-being measures (time × arm interaction is the H1 test), participant random intercepts, baseline as covariate. Dropout handled by the mixed model / multiple imputation, not casewise deletion. Effect sizes + CIs reported, not just p-values.

Reuse vs build (on the Citizen Science rails)

Reuse: participants (enrollment + consent stamp + fuzzed privacy), competitions/enrollment pattern, satisfaction_quick_values (daily mood sampling), family_routine_checkins schema shape (scheduled, dated, status), cron + heartbeat infra, Stripe Connect payout plumbing, is_seed real/test split, journal_entries timestamps for the language outcomes, PostHog captureServerEvent.

Build (RCT-specific):

study_enrollments (study_id, participant_id, assigned_arm, enrolled_at, stratum, eligible) + seeded randomization at enroll.
A validated-instrument survey primitive — multi-item Likert capture + auto-scoring (WHO-5 etc.) + per-response timestamps. (surveys, survey_questions, survey_responses.)
Active outreach cron: "who is due for their T1/T2 instrument on study-day N of their enrollment" → email via Resend; track completed/missed.
Participant compensation ledger (platform → participant; distinct from creator earnings) + payout on milestone completion.
Cohort export admin endpoint → one row per participant, ITT-ready CSV.

Phased plan

Phase A — Protocol + pre-reg: finalize instruments, write the OSF pre-registration, IRB submission, consent + intake copy. (No code.)
Phase B — Study rails: build items 1–3 above; dry-run with is_seed participants end-to-end (enroll → T0 → run-in → T1 → Odessa → T2).
Phase C — Pilot (~30–40/arm): recruit, run 60 days, compute effect size + attrition. Build items 4–5.
Phase D — Confirmatory RCT: re-power on pilot data, pre-register v2, run.
Phase E — Publish + tell the world. The result becomes the marketing the ethos always wanted: not "we believe," but "we measured."

How it could fail (named on purpose)

Attrition over 60 days could swamp the signal; the effect could be real but smaller than a pilot can detect; an active-control purist will note that v1 control lacks an attention-matched placebo (a later arm: non-personalized generic story). None are fatal; all are reasons to pilot, pre-register, and stay honest.

Status: draft protocol for review. The design (2-arm, journaling run-in, marginal-effect question, pre-registered) is the considered recommendation; the instrument choices and N are the live decisions.