How updateme could rank a date

Rank people by how likely a date is to actually happen.

The path to a card

Nearby people
Hard filters
Cooldowns
Cap of 3
Score
Daily shuffle
Top 10

The score: five chances, multiplied

Full math
S = f × A × C × g × ρ f = sigmoid(−1.60 + P + 0.50(Q − 0.5)) i proposes to j A = 1 − exp(−w / 6) j answers before expiry C = (1 + n)^(−0.6) j isn't already taken g = sigmoid(−1.10 + P) j accepts i ρ = ((6.5 + H) / (8 + H + N)) / 0.8125 j shows up (1.0 at launch) P = 0.12·z − 0.30(δ − 0.5) − 0.30·[age gap 5–9] − 0.70·[age gap 10+] φ = (K·o − kᵢ·kⱼ) / √(kᵢ·kⱼ·(K − kᵢ)·(K − kⱼ)) z = φ·√(K − 1) Rank by ln S + 0.25·Gumbel, show the top 10 K interests both saw · k kept by each · o kept by both · δ distance ÷ smaller travel limit Q profile completeness · w waking hours to expiry · n pending incoming · H happened dates · N no-shows

Try it: 1,000 cards through the funnel

Score S
vs worked example

Interests: more than chance, not more in total

Both keptOnly one keptNeither

Who gets the proposals: people keeping 30+ interests

Current
37.6%
Gemini × window
21.0%
This formula
9.2%
Line = their fair share (9.0% of users).

Time: sleep eats the answer window

Chance of answering in time

Proposals that expired (simulated)

≤ 3 waking h
61%
> 8 waking h
5.4%

Sharing attention

Load: the more waiting, the lower j ranks

1.00
0.66
0.52
hidden
0 waiting123

Daily shuffle: strong cards lead, others get a turn

What the simulation found (synthetic)

Dates vs the current ranking, 95% interval, by users and gender split

0%+10%+20%+30%
Grey: that gain comes from cooldowns, not ranking.

People nearby matter most: candidates in a typical session

Shown (10-card limit)Didn't fit

Lift in dates per user

100 → 1,000 users
2.6×
Best-case ranking
1.49×

What happens to Gemini's formula

Keep
  • Cooldowns
  • Load term
  • Cap of 3
  • Intent + smoking
  • "See them again"
Change
  • Weighted sum → probabilities
  • Interests 50% → φ
  • Quality → proposing only
  • Reliability → proper prior
  • Jitter → Gumbel
Drop
  • Harmonic mean
  • Urgency bonus
  • Recency
  • Stars, related credit
  • 300 / 2,500 thresholds

Ship order

Logging + who no-showed
Cooldowns
This formula
Waking-hours reminder
Intent + smoking

Then learn from real data