// self-improving intelligence network
AI that
evolves AI.
Today's AI is frozen the day it ships. OURO lets deployed agents keep evolving on real tasks, with every round proven on-chain and powered by a global compute network built for evolution.
01 · Why now
Intelligence is moving from weights
to scaffolding.
In 2026 OpenAI announced it would wind down self-serve fine-tuning because "prompts and scaffolding are cheaper and faster." What separates agents is no longer their weights but their prompts, tools, memory and skills, and that layer is one AI can evolve on its own.
DARWIN GÖDEL MACHINE · 2025
20% → 50%
A coding agent rewrote its own scaffolding and doubled its SWE-bench solve rate. Only prompts, tools and code changed; the weights never moved.
GEPA · ICLR 2026 ORAL
35× cheaper
Evolving through natural-language reflection beats GRPO reinforcement learning by about 10% on average with 35× fewer rollouts. Every failure is used in full instead of being squashed into a single number.
META CONTEXT ENGINEERING · 2026
+89%
A meta-agent evolved the skill of building context itself, lifting finance, chemistry, medicine, law and security tasks by 89% on average with frozen weights.
79% / 23%
Share of enterprises adopting agents vs. the share that has actually scaled them (PwC / McKinsey)
40%
Share of agentic AI projects Gartner expects to be cancelled by the end of 2027
$37B
Enterprise generative-AI spend in 2025, up 3× from 2024 (Menlo Ventures)
What blocks enterprises is not that models are too dumb, but that agents stop getting smarter once they ship. Nobody has solved "how do we keep improving after launch," and nobody verifies that the improvement really happened. That is the position OURO is taking: the evolution layer.
02 · Evolution loop
One round of evolution, five steps, evidence at every step.
Diagnose
Run the eval set and sort failures into natural-language reflections: tool misuse, reasoning errors, knowledge gaps.
Generate candidates
Sample parents from the archive by score × novelty and generate N variants. Prompts, tools, memory and skills first; weights optional.
Distributed evaluation
Candidates are dispatched to OURO Grid and scored redundantly across nodes. Nodes exchange only task shards and results, never weights.
Select and archive
Pareto selection on gain, cost and variance. Every historical version is kept, with rollback and diff.
Submit proof
Before/after scores, the eval-set hash and node signatures are written to the Proof of Evolution contract with a human-readable changelog.
03 · Stack
Three parts, one closed loop.
Others prove "it computed." OURO proves "it got better."
ENGINE
OURO Engine
An open-ended search system with four layers, shallow to deep: context and prompts → skills and tool chains → memory and context-engineering strategy → weights. By default only the first three move, and every change can be diffed, reviewed and rolled back.
layers: [context, skills, memory]
weights: false
archive: open-ended
PROOF
Proof of Evolution
A layered verification stack: TOPLOC-style activation hashes (about 1% overhead) against skipped compute, multi-node redundancy against false reports, blind holdout slices against overfitting, optional TEE against leakage, and regional anchors on different continents to arbitrate disputes.
eval: sha256(…), holdout: ✓
score: 0.61 → 0.74, σ: 0.02
nodes: 5/5 agree, sig: ✓ }
COMPUTE
OURO Grid
Ninety percent of an evolution loop's compute goes to evaluation and sampling: independent, repeatable, no weight sync. That part follows price onto community nodes worldwide; weight updates, arbitration and private-data hosting stay on regional anchors, starting in Singapore and Johor Bahru.
elastic: consumer GPUs // evals, rollouts
external: DePIN // burst capacity
04 · Compute grid
A global compute grid,
built for evolution.
Ninety percent of evolution work is parallel and location-agnostic, so it runs wherever compute is cheapest. The remaining ten percent needs reliability and runs on mid-sized anchors spread across regions. The network has guaranteed capacity from day one and is global from day one. Your device picks up evaluation and sampling tasks, stakes a small amount of $OURO for quota, and settles by tasks completed and verification pass rate.
- Consumer GPUs (RTX 4070 and up) can join; devices without a GPU can run light verification tasks
- Nodes only touch de-identified task shards, never full weights
- Early nodes earn a "genesis node" incentive bonus; communities and partners can co-build new regional anchors
ANCHORS · LIVE FIRST
Singapore · Johor Bahru — arbitration · high-bandwidth tasks · baseline capacity, ~30 km cross-jurisdiction failover
ANCHORS · NEXT · COMMUNITY-BUILT
Northeast Asia (Tokyo / Seoul) · Europe (Frankfurt / Amsterdam) · Central US
ELASTIC · COMMUNITY GPUs · GLOBAL
Evaluation · sampling · low-communication training shards, scheduled globally by price and reliability
EXTERNAL · DePIN & SPOT MARKETS
io.net · Aethir · Akash · RunPod · Vast — bought on demand at peak load
2026-03 H100 spot ≈ $1.05/hr, down 22% YoY · APAC 35–45% pricier than the US, Europe 12–15% cheaper · average enterprise GPU-cluster utilization only 5%
05 · Products
One engine, four doors.
FOR DEVELOPERS & ENTERPRISES
Evolve API
Submit an agent definition + eval set + budget, get back the evolved version and an on-chain PoE report. Compatible with OpenAI, Anthropic, LangGraph and CrewAI formats, no stack migration.
Apply for early access →FOR EVERYONE
OURO Chat
An assistant that fits you better the more you use it. A "My evolution" page shows where it got better this week; contribute feedback to earn points; conversations stay out of training by default.
Join the beta waitlist →FOR COMPUTE PROVIDERS
OURO Grid client
One-click install. Idle GPUs pick up evaluation and training tasks and settle in $OURO by tasks completed and verification pass rate. Windows / macOS / Linux.
About the genesis-node program →FOR BUILDERS
OURO Studio
Evolution history, version diffs, an eval-set marketplace and evolution bounties. Lets domain experts turn evaluation skill into income.
Read the docs →06 · Roadmap
Four phases, three gates.
Every phase has a gate; nothing advances until the gate is met. TGE comes after Evolve API is commercial.
2026 Q4 – 2027 Q1
Phase 0 · Foundation
Foundation set up · whitepaper and website · data-center onboarding · engine private beta (3 pilots)
2027 Q2 – Q3
Phase 1 · Genesis
Chat public beta · user devices join Grid · PoE testnet · first evolution bounties
2027 Q4 – 2028 Q1
Phase 2 · Open
Evolve API commercial · PoE mainnet · TGE · eval-set marketplace
2028 Q2 →
Phase 3 · Expand
External DePIN · enterprise private deployments · agent identity · governance handover
Join the evolution network.
Leave your email for the whitepaper, the genesis-node program and early access to Evolve API.