← → / Space to navigate · E to edit
Bittensor Subnet · Routing Intelligence

ProviderRank

Verifiable routing intelligence for the same model across every provider endpoint — mined, replayed, and verified.

We don't rank models. We rank provider endpoints for the same model.

A decentralized routing-intelligence market — not a dashboard.

01 / The Problem

Same model. Different providers. Materially different behavior.

Developers access models through multi-provider gateways. The same model name can perform very differently depending on which endpoint serves it.

latency timeout / failure rate structured-output reliability repeat-call consistency effective cost-performance

Model selection became a routing decision — still made on stale benchmarks, anecdotes, and brittle heuristics.

02 / The Solution

Turn unstable provider performance into a continuously refreshed routing commodity.

The output isn't a leaderboard. It's production-usable routing guidance the same model can be served by.

01

Recommended primary

The best provider endpoint to send traffic to right now.

02

Ordered fallback chain

A ranked backup sequence for graceful degradation.

03

Routing config

A production-ready policy that drops into the gateway.

03 / The Mechanism

Miners produce it. Validators verify it. Emissions reward it.

Mine

Measure

Per cell, miners submit telemetry, a routing signal, and a proof bundle (prompt/response hashes, timing, schema checks).

Verify

Replay

Validators replay sampled tasks under pinned params. Out-of-tolerance trips the Gate to 0 — no LLM judge.

Reward

Emit

Emissions flow proportionally to the most accurate, fresh, and useful routing signals.

Unit of work: cell = model × provider-endpoint × task-type
04 / Incentive Design

Deterministic scoring. No LLM judge.

Scell = Gate × ( 0.45·Accuracy + 0.25·Usefulness + 0.20·Freshness + 0.10·Coverage )

Success-first gate

Out-of-tolerance replay invalidates the whole submission. Gate ∈ {0,1}.

Coverage × Accuracy

Coverage is multiplied by accuracy — scale without truth earns nothing.

Proportional, capped

Rewards are proportional, not winner-takes-all, with a ~15% weight cap.

41% miners 41% validators + stakers 18% owner
05 / Why This Fits Bittensor

A time-sensitive, verifiable intelligence commodity.

Continuous demand

Provider quality changes constantly, so routing intelligence must be refreshed continuously.

Partial verifiability

Validators can replay challenge tasks and verify latency, success, schema, and routing quality.

$

Direct economic utility

Better routing lowers cost and latency and raises reliability in production.

Competitive supply

Many miners compete to generate better signals while validators continuously verify them.

06 / Market

Buyers already feel this pain in production.

Who buys

  • OpenRouter power users comparing providers for the same model
  • Agent & AI application developers needing task-specific routing
  • Inference teams optimizing cost, reliability, and fallback

What they pay for

  • Lower failure rates
  • Lower effective latency
  • More stable structured outputs
  • Safer fallback behavior

Not a leaderboard subscription — a shared intelligence layer that replaces fragmented private benchmark stacks.

07 / Go-To-Market

A three-sided market that reinforces itself.

Miners

Recruited from LLM infra, eval/benchmark, and competitive builder communities. Early emission upside on high-value cells.

🛡

Validators

Technically credible operators who run replay infra. Early position, staking economics, governance influence.

🚀

Consumers

Agent builders & inference teams. Free/discounted early access in exchange for the demand that makes signals worth producing.

Miners compete → validators make signals trustworthy → consumers create the demand. That loop is the go-to-market.

We don't rank models.

We rank provider endpoints for the same model — then turn that into routing decisions.

ProviderRank · A Bittensor Subnet for Verifiable Provider Routing Intelligence