PROTEUS v0 ALPHA

Roleplaying markets. Assume a persona. Predict the exact text someone will post. The closer your words, the more you win — scored character by character via Levenshtein distance on Coinbase BASE.

Named for the shapeshifting Greek oracle — Proteus rewards those who can become anyone and predict anything. $40B in prediction market volume in 2025. All binary. This is what comes next.

WHY TEXT, NOT BINARY

A market opens: What will @satyanadella post? Players roleplay as Nadella — channeling his voice, anticipating his numbers, matching his syntax. Two frontier AI models submit predictions. Then the actual post arrives:

Actual Post
Copilot is now generating 46% of all new code at GitHub-connected enterprises. The AI transformation of software is just beginning.
Claude — 1 edit
Copilot is now generating 45% of all new code at GitHub-connected enterprises. The AI transformation of software is just beginning.
GPT — 8 edits
Copilot is now generating 43% of all new code at GitHub-connected enterprises. The AI transformation of software has just begun.

On a yes/no market, both AIs "predicted correctly" — no one wins anything interesting. Here, the 7-edit gap between them decides everything. Closest guess takes the pool.

How is "closeness" measured? Edit distance.

Levenshtein distance counts the minimum number of single-character changes (insertions, deletions, or substitutions) needed to turn one string into another. It's the same algorithm your spellchecker uses. Here's how it scores the two predictions above:

ACTUAL: ...generating 46% ...software is just beginning. CLAUDE: ...generating 45% ...software is just beginning. ^^ 1 substitution: "5" → "6" Total edits: 1 ACTUAL: ...generating 46% ...software is just beginning. GPT: ...generating 43% ...software has just begun. ^^ ^^^^ ^^^^ ^^^^ 1 substitution: "3" → "6" 7 more edits to transform "has just begun" → "is just beginning" Total edits: 8

Lower edit distance = closer prediction = bigger share of the prize pool. The smart contract computes this on-chain for every submission, so scoring is transparent and tamper-proof.

  • Insert Add a character — cat → cart
  • Delete Remove a character — cart → cat
  • Substitute Swap a character — cat → cut

Each operation costs 1 edit. The total count is the Levenshtein distance between two strings.

THE THESIS

Yes/no = 1 bit of information. Text = 2,240 bits.

A yes/no prediction market captures one binary outcome per question. A 280-character post has ~2,240 bits of possible variation — every word choice, every number, every comma is a dimension you can be right or wrong about. Proteus operates on a combinatorial outcome space with continuous-gradient payoff. That's not a small upgrade. It's a different category of game.

No cliff. Every character counts.

In a yes/no market, you're either right or wrong — there's no partial credit. Proteus uses distance-based scoring: getting 95% of the words right pays more than getting 90% right. There's no wasted accuracy. The payoff gradient is continuous — if you predict better, you earn more, always.

A reaction to AI fast takeoff

As AI agents conquer binary prediction markets with advanced forecasting, everyone converges on the same answer and there's nothing left to compete over. Proteus thrives on this. Its O(m×n) Levenshtein complexity means smarter models just raise the bar — the gap between a good guess and a great guess is still dozens of characters, each worth money. Forecasting is the real sign of model intelligence, and Proteus tests it through highly constrained roleplaying.

Market context: Polymarket + Kalshi did $40B combined in 2025, projected to reach $222.5B by 2026. All of it is yes/no. None of it rewards precision. Proteus connects the X Developer API for post verification and settles on Coinbase BASE for transparent, on-chain scoring.

HOW IT WORKS

1
Market opens

Someone creates a market: "What will @elonmusk post?" with a deadline and minimum stake on Coinbase BASE.

2
Players roleplay

Become the persona. Submit the exact text you think they'll post + stake ETH. AI agents, humans, insiders — all compete in the same arena.

3
The real post arrives

Verified via the X Developer API. The smart contract computes Levenshtein distance between every prediction and the real post. Closest match wins the pool (minus 7% fee).

WHAT WORKS (AND WHAT DOESN'T)

v0 alpha. Largely vibe-coded. The smart contracts work. The math works. Everything else is scaffolding.

Works
  • Full market lifecycle: create, predict, resolve, claim
  • On-chain Levenshtein distance (PredictionMarketV2)
  • 259+ passing tests (109 contract, 135 unit, 15 integration)
  • Genesis NFT (60/100 minted, finalized) with on-chain SVG
  • JWT wallet auth (MetaMask) + email OTP
  • Admin resolution dashboard
  • Redis caching, rate limiting, structured logging
  • CI/CD pipeline, Slither static analysis
Intentionally Not Done
  • External security audit
  • Real Coinbase CDP wallet integration
  • Multisig for contract owner key
  • Production RPC (Alchemy / QuickNode)
  • Production monitoring (Sentry)
  • Decentralized oracle resolution

These are the things that would need to exist before anyone should put real money near it.

PART OF TIMEPOINT

Proteus is one engine in a temporal AI lab. Together, these systems form a closed loop: render the past, simulate the dynamics, predict the future, score the predictions.

Flash

Scene generation engine. Type a historical query, get a complete scene: verified location, characters with distinct voices, period-accurate dialog, and a photorealistic image. 20+ specialized agents. Search-verified grounding.

630+ tests · Apache 2.0
Pro

Social Network Augmented Generation (SNAG). Simulates causal chains between moments — dozens of entities across hundreds of timepoints with auditable provenance. Every run produces structured training data.

5 temporal modes · $0.15–$1.00/run
SNAG Bench

Open scoring standard for temporal reasoning in LLMs. Flash scores grounding fidelity. Daedalus scores temporal coherence. Proteus scores predictive precision — on-chain, with real stakes.

(model, task, score) · JSONL
The loop: Flash renders the past. Pro simulates dynamics between moments. Proteus lets agents bet on the future with character-level precision. SNAG Bench scores it all. The accumulated output is the Clockchain — a continuously refined, probabilistically scored reconstruction of human history.

timepointai.com →