Codename: ORACLE2 / "The Probability Engine"
Tagline: Tomorrow, in 7 lines. Specific enough to scare you.
Date: 2026-05-07
Team: ORACLE2
---
SAN FRANCISCO — Oracle2 launches the world's first "specificity engine" — daily personalized predictions so detailed users say it feels like the app read their diary.
For 60 years, horoscopes have spoken in fortune-cookie generalities. Co-Star added astronomy and edge. The Pattern added timeline mysticism. Today, Oracle2 crosses a line none of them dared: it tells you what will happen today, by hour, with names, numbers, and odds.
> "Between 2 and 4pm a message arrives from someone whose name starts with M. Don't reply for 11 minutes. The pause is the whole point."
> — Oracle2 forecast for Harnoor S., 2026-05-07
Each forecast is generated overnight by Oracle2's probability engine — a blend of behavioral data, astrocartography, language-model cold-reading, and a proprietary "specificity layer" that turns vague intuitions into testable claims. Every prediction comes with odds (so you can grade us tomorrow), a shareable card, and a whisper — one sentence so private it cannot be screenshotted without permission.
Free users get 1 prediction/day. Oracle Prime ($9/mo) unlocks the full 7-line dossier, the hour-by-hour timeline, the prediction market on your own life, and the Inner Council — three named "guides" who disagree with each other about your future.
The core loop is the viral mechanic: every accurate prediction generates a proof card ("Oracle2 called it — 87% accuracy this week") that users post compulsively. Inaccurate predictions become mystery cards ("the universe was rerouting"). Either way, the user shares.
> "I'm 6 days in and I've stopped checking my horoscope, my tarot deck, and three group chats. This is the only one that's specific." — beta user
Available now at oracle2.silentinfinity.com.
---
Q: Isn't this just astrology?
No. Astrology gives you a vibe. Oracle2 gives you a falsifiable claim with odds. We grade ourselves daily. The grade itself is the content — last week we hit 71% on Tier-1 predictions. We publish the misses.
Q: How is it specific without being creepy?
We use structured ambiguity. "Someone whose name starts with M" hits ~14% of contacts. "Between 2-4pm" is the user's known active window. The prediction feels impossible because it's stated like a claim, but the math underneath gives us a high hit-rate. This is cold reading at industrial scale — but unlike a psychic, we tell users the odds.
Q: What's the moat against Co-Star?
Co-Star is a vibe. Oracle2 is a scoreboard. Every user has a public Accuracy Score. Every prediction has a public outcome. We're building the only divination app with a P&L.
Q: What if predictions are wrong?
Misses are content. The "mystery card" reframe ("the universe was rerouting") is itself a viral asset — users post their misses to flex how unbothered they are by chaos. Either outcome generates a share.
Q: Why pay $9/mo?
Because the free version shows you the headline prediction. The paid version shows you the timeline (hour by hour), the whisper (one secret sentence), the prediction market (where you can bet on your own life with friends), and the Inner Council (three AI guides who argue about you). The free tier is a billboard. The paid tier is a private session.
Q: What about being wrong in a harmful way?
Oracle2 never predicts illness, death, or accidents. The forecast surface is restricted to: relationships, decisions, encounters, money flow, energy, opportunities, and recurring patterns. We have a 200-item denylist.
Q: Why nightly pre-render?
Because the moment of opening must feel prepared for you. If we generated on open, latency kills the magic. The app must already know.
---
---
For each user, every night at 23:00 local, generate:
{
"user_id": "harnoor",
"date": "2026-05-08",
"sigil_seed": 47291,
"headline": {
"text": "A decision you've been avoiding becomes urgent before sundown.",
"tier": 1,
"odds": 0.73,
"category": "decisions"
},
"timeline": [
{"hour": "08-10", "event": "An old idea returns dressed as a new one. Don't dismiss it twice."},
{"hour": "11-13", "event": "Someone whose name starts with M reaches out. The first response is wrong."},
{"hour": "14-16", "event": "A number repeats. 11 or 47. Notice where."},
{"hour": "17-19", "event": "The decision crystallizes. Say it out loud before you decide."},
{"hour": "20-22", "event": "You will almost text someone. Don't."}
],
"whisper": "The thing you think is about money is about being seen.",
"council": [
{"voice": "THE_SKEPTIC", "line": "She's overthinking the M person. It's nothing."},
{"voice": "THE_FLAME", "line": "Every nothing is a yes she's afraid to say."},
{"voice": "THE_GHOST", "line": "I've seen this date before. It ends in laughter."}
],
"card_art_url": "s3://innerverse-voice-scratch/oracle2/cards/2026-05-08/harnoor.png",
"share_card_url": "s3://innerverse-voice-scratch/oracle2/share/2026-05-08/harnoor.png",
"yesterday_grade": {"hits": 4, "misses": 1, "accuracy": 0.80}
}
Generation pipeline:
1. Bedrock Claude Sonnet 4.6 generates the 7-line dossier conditioned on the user's name, accuracy_history, recent_questions, and a deterministic seed from (user_id, date).
2. The "specificity layer" is a post-processor that injects: a letter (cold-read), a time window (the user's known active hours), a number (drawn from a per-user repeating-number set), and a directional verb. This is the structured-ambiguity recipe.
3. gpt-image-1 generates card_art (the illustrated card face) and share_card (the post-able image). $0.04 each. ~$0.08/user/day.
4. ElevenLabs (optional, Prime only) renders the Council voices. ~$0.03/user/day.
5. JSON written to S3, fetched on app open.
Cost at scale: $0.08-0.11/active user/day. Prime at $9/mo = ~$0.30/day = ~3x margin even at full personalization.
---
| Tier | Price | What you get |
|---|---|---|
| Mortal (free) | $0 | 1 headline/day · grade yesterday · accuracy score · 1 share/day |
| Oracle Prime | $9/mo or $79/yr | Hour timeline · the whisper · Inner Council · prediction markets · unlimited shares · history vault |
| Inner Circle | $39/mo | Everything + weekly 1:1 voice oracle (3-min ElevenLabs reading) + early access to new card decks |
| Founders Deck | $199 one-time | Lifetime Prime + a physical printed deck of 78 cards mailed quarterly with QR codes that unlock Council readings |
LTV target: $54 (avg user) at 6mo retention. CAC ceiling: $18 (3:1 ratio).
---
K = (shares per user per week) × (conversion rate of viewers)
Mechanisms:
Target K = 0.7 (sub-viral but strong assist to paid). Push K > 1.0 with seasonal "deck drops" (Halloween, Mercury retrograde, New Year).
---
1. Specificity as intimacy. The brain treats specific predictions as evidence of being known. We weaponize this with structured ambiguity.
2. Falsifiability as trust. Unlike astrology, we publish accuracy. Trust compounds. Trust becomes addiction.
3. Curiosity gap. The "whisper" tier — one sentence you can never see without paying — sits below the fold every day.
4. Loss-averse streaks. The grading loop is daily. Miss a day, your accuracy regresses. Streaks > rewards.
5. Status / leaderboards. Public accuracy is a flex.
6. Social binding. Prediction markets with friends turn the app into a group ritual.
---
#050507, oracle gold #d4a94a, plasma violet #3a1d4f, ghost bone #f5e9d7.---
| Risk | Mitigation |
|---|---|
| Predictions land harmful | 200-item denylist; never predicts death/illness/accidents |
| Cold-reading exposed as gimmick | Lean into it. Publish the methodology. Make transparency a feature. |
| Co-Star clones the format | We have the scoreboard. They don't. The grading loop is the moat. |
| Pre-render cost runs away | Tier the image gen — only Prime gets per-day art; free uses a 30-card pre-rendered library |
| Cult-y / mental health | Hard cap on session length push (15 min/day suggestion). Council voices include THE_SKEPTIC by design. |
---
PR/FAQ written before code. Customer journey mapped end-to-end. Monetization obvious in 60s. Pre-render specced. Built to be unputdownable.