Pairing users into mutual accountability partnerships for habit goals using multi-feature compatibility scoring
Every downstream matching decision is only as good as the intake data feeding it. Before any scoring or pairing happens, the system needs a compact, honest picture of who a user is as an accountability partner: what they are trying to achieve, how often, when they are actually awake and available, how committed they say they are, and — for returning users — how reliably they have shown up in the past. Habit-formation research consistently shows that structured commitment at the point of goal-setting predicts adherence far better than the goal itself.
Behavioral economics has repeatedly shown that intentions alone are weak predictors of follow-through — the gap between what people plan and what they do is the central problem habit apps are built to close. Commitment-device research (Ayres & Karlan's StickK platform, Katy Milkman's work at Wharton on temptation bundling and commitment contracts) demonstrates that adding stakes, observability, or a second party to a goal measurably increases completion.
An accountability partnership layers several of these mechanisms onto a single goal simultaneously:
• Social commitment — stating a goal to another person increases the psychological cost of abandoning it • Reciprocal obligation — each partner's consistency is implicitly a promise to the other, not just to themselves • Observability — a partner who can see your check-ins removes the private, low-stakes nature of solo goal tracking • Loss aversion — letting a partner down is processed as a more acute loss than missing a personal target alone
None of these mechanisms activate, however, unless the two people paired together are actually compatible enough to sustain the relationship long enough for the mechanisms to compound.
The intake form is deliberately short (four core fields beyond the goal itself) because completion rate drops sharply after the second minute of onboarding — but every field is chosen because it feeds a specific matching feature downstream, not because it is easy to ask.
• Goal category + free-text goal description: mapped to one of 14 taxonomy categories and embedded as a short text vector for similarity scoring • Target frequency: sessions/week, used to align cadence expectations between partners • Timezone + typical active hours: required for computing a shared availability window • Self-reported commitment level (1–10): a calibrated Likert item shown to correlate with early-partnership dropout risk • Prior app usage / reliability history: for returning users, pulled from historical check-in logs rather than self-report
New users with no history hit the cold-start problem: the system assigns a population-average prior reliability score (≈0.62) with wide uncertainty, which is shrunk toward observed behavior via Bayesian updating after the first five check-ins — preventing an untested user from either being penalized or over-trusted on day one.
Raw intake fields are not directly comparable — a goal description, a timezone, and a Likert score live in completely different spaces. Feature extraction converts each into a normalized signal, and a weighted composite function turns those signals into a single 0–1 compatibility score for every candidate pair in the pool.
Each candidate pair is scored on four weighted sub-features, combined into a single composite compatibility score:
• Goal similarity (weight 0.35) — cosine similarity between 64-dimensional goal embeddings; two users both training for a 10K score far higher than a runner paired against a meditation goal, even within the same broad category • Timezone overlap (weight 0.20) — hours per day both users are typically active; pairs below a 4-hour floor are heavily penalized since same-day mutual check-ins become logistically unlikely • Commitment alignment (weight 0.20) — a small delta between self-reported commitment scores; pairing a 9/10-committed user with a 3/10 user tends to produce early resentment and asymmetric effort • Reliability score (weight 0.25) — historical check-in consistency; the composite score is capped so that no amount of goal or schedule similarity can compensate for a candidate with a very low reliability estimate
Weights were tuned against retrospective partnership-survival data — pairs above a 0.65 composite score were roughly twice as likely to still be mutually checking in at week 6.
Self-reported commitment is a weak predictor on its own; the system instead tracks an exponentially weighted moving average (EWMA) of observed check-in behavior:
R_t = α·X_t + (1−α)·R_{t−1}
where X_t is 1 if the user completed their scheduled check-in and 0 otherwise, and α = 0.3 gives roughly an 8 check-in half-life — recent behavior dominates, but a single missed day does not collapse the score.
New users without check-in history are assigned a Bayesian-shrunk prior: R_0 = population mean reliability, weighted as if it were a small number of pseudo-observations, so the estimate moves quickly toward a user's true behavior once real check-ins start arriving but is not distorted by the very first data point.
With a pairwise compatibility score computed for every candidate in the pool, the system needs an assignment procedure that turns a dense score matrix into a clean set of one-to-one partnerships — maximizing joint compatibility while respecting hard constraints that a pure top-score greedy pass would violate.
The classic Gale–Shapley algorithm (1962) solves stable matching for two distinct sides of a market (e.g. students and schools). Accountability partnering is a one-sided market — any user can in principle be paired with any other — which is closer to the "stable roommates" problem (Irving's algorithm, 1985).
In production, each user builds a ranked preference list from their compatibility scores against the pool. Proposal rounds run iteratively: unmatched users "propose" to their top remaining candidate; a candidate holds the best offer received so far and rejects weaker ones; rejected users move to their next preference. The process terminates in a stable configuration — no two users would both prefer each other over their assigned partners — typically within a handful of rounds for a pool in the low thousands.
At this scale, O(n²) pairwise scoring (≈1.4M comparisons for 1,204 users) plus a bounded number of proposal rounds comfortably fits in a nightly batch job.
A pure score-maximizing match can still produce unusable pairs, so several hard constraints filter the candidate list before proposals are generated:
• Mutual availability — timezone overlap must clear the 4-hour floor regardless of how well goals align • No repeat pairing — users are never re-matched with a partner from a prior partnership that churned or was flagged as failed • Reliability floor — high-commitment seekers (9–10 self-rating) are only proposed candidates above a minimum reliability threshold, since an unreliable pairing there produces outsized disappointment • Single-partner cap — this design is strictly 1:1; group accountability pods are excluded from this pipeline • Cold-start preference — brand-new users with no history are preferentially proposed to a smaller subset of proven high-reliability partners for their first pairing
For the current cycle, 602 stable pairs are produced from a pool of 1,204 users, with an average matched-pair compatibility score of 0.74. Users left unmatched at the end of a cycle (odd pool size, or too few compatible candidates within constraints) are rolled into the next cycle with slightly relaxed timezone and commitment-delta thresholds rather than left waiting indefinitely.
Once a pair is finalized, the match is "locked": both users are notified simultaneously, and the pairing is written to the partnership table that drives the check-in loop in the next stage.
A stable match is only the starting point. The behavior-change effect of an accountability partnership comes entirely from what happens after activation: a recurring, visible, mutual check-in loop that turns a one-time algorithmic decision into a sustained social contract.
Each partner logs a lightweight check-in (done / not done, optional short note) tied to their shared goal cadence. The check-in is visible to the partner in near real time, and a shared streak counter tracks consecutive successful cycles for the pair as a unit, not just per-individual.
An escalation ladder handles silence:
• 0–24 hrs quiet: no action, normal variance • 24–48 hrs quiet: a gentle nudge to the silent partner only • 48+ hrs quiet: both partners are notified — the silence itself becomes visible, which is often the strongest lever in the system • Extended stall (7+ days): the partnership is flagged for the outcome/rematch pipeline in the next stage
The escalation is intentionally social rather than purely transactional: it surfaces the partner's awareness of the gap rather than simply repeating a generic reminder.
Reciprocal nudges — where the message effectively says "your partner is waiting to see your check-in" — are opened at roughly double the rate of a generic solo reminder in the same product (≈78% vs. ≈34%). This tracks with loss-aversion research: the prospect of visibly disappointing another person is processed as a more acute, immediate cost than the diffuse, delayed cost of missing a personal goal.
Reciprocal accountability compounds several mechanics at once:
• Visible partner disappointment — the silent partner knows the other party can see the gap • Public/dyadic commitment — a promise made to another person is harder to quietly abandon than one made only to an app • Streak-as-shared-asset — a broken streak is a joint loss, not an individual one, which increases the felt cost of lapsing • Implementation-intention reinforcement — the recurring check-in slot functions like a scheduled "if-then" cue (Gollwitzer's implementation-intention research), strengthened by a partner's expectation
Structured buddy systems in exercise-adherence literature are consistently associated with roughly double the session attendance of matched solo controls over 8–12 week windows — the effect is not from the goal itself but from the recurring, mutual visibility of follow-through.
The matching system does not end at pairing — every partnership generates data that measures whether the pairing actually worked, and that data flows back into the reliability scores and constraint set used for the next matching cycle. This closed loop is what separates a one-time algorithmic assignment from a system that improves over time.
Outcome measurement compares completion rates between matched pairs and a randomized holdout of otherwise-similar solo goal-setters over an 8–12 week window. Across cycles, paired users complete their target frequency at roughly 1.9–2.1× the rate of the solo cohort — broadly consistent with the wider commitment-device and exercise-buddy literature (StickK-style commitment contracts, buddy-system adherence studies), which repeatedly finds that adding a social or financial stake to a goal roughly doubles follow-through versus an unstructured intention.
Because self-selection is a serious confound — people who seek out a partner may already be more motivated — the holdout group is drawn from users who requested a match but were randomly withheld for one cycle, isolating the effect of the partnership itself rather than the act of wanting one.
Every check-in updates each user's EWMA reliability score in real time, and that updated score is what the Stage 2 compatibility function reads for the next matching cycle — a user who shows up consistently becomes a more attractive candidate for future high-commitment seekers, while a user who repeatedly lapses is proposed more cautiously.
Stalled partnerships (7+ days without a mutual check-in from either side) are flagged for review rather than left to decay silently:
• Both users are offered an explicit "rematch" option rather than being auto-reassigned without consent • A churned pairing is recorded and permanently excluded from being re-proposed to either party • Users who consistently stall across multiple partnerships are routed to a lower-cadence, lower-stakes matching track rather than repeatedly failing at the standard cadence • Healthy, still-active pairs are never disturbed — the rematch loop only acts on partnerships that have gone quiet
Closing the loop turns matching from a one-time decision into a continuously-improving system: every check-in is training data for a better next match, and proactively rematching stalled pairs prevents "zombie" partnerships — technically matched, functionally inactive — from silently dragging down platform-wide completion metrics.