VERITENSOR Run demo
Bittensor Global Subnet Hackathon 2026

The decentralized verification layer for machine intelligence.

Miners compete to produce reliable AI answers. Validators independently verify them. Performance determines reputation and emission.

VERITENSOR does not reward miners for producing output. It rewards output that survives independent verification.

Run network demo Explore mechanism

4

verifiable task families

13

task generators

5

scored dimensions

9

miner archetypes

01 — Problem

Nobody is paid to be reliably right.

Generation is cheap. Verification is not.

Model output is abundant and nearly free. Determining whether a given answer is actually correct still costs human review, and it does not scale with the volume of machine-generated claims.

Benchmarks leak, then die.

A static benchmark is a public answer key. Once it circulates it measures memorisation, not capability, and the leaderboard it produces is no longer informative.

Confidence is unpriced.

A model that is right 60% of the time while claiming 95% certainty is more dangerous than one that admits uncertainty. Nothing in today's stack charges for that miscalibration.

02 — Solution

Turn verification itself into the commodity.

A validator generates a task whose answer it already knows but has never published. Miners answer. The validator grades correctness, evidence, robustness under mutation, confidence calibration and latency, then converts the result into a normalised emission weight.

Dynamically generated tasks

Four verifiable families — code security, mathematics, logical reasoning, data analysis — generated from random seeds with ground truth computed at generation time. 13 generators, no static answer key.

Independent validation

Validators grade with deterministic verifiers, never with another language model. The same answer always receives the same score, and any judge can re-run it.

Adversarial robustness probes

After a correct answer a validator may issue a semantics-preserving mutation of the same problem. A conclusion that flips under renaming was never knowledge.

Calibration is scored

Stated confidence is graded with a Brier score over a rolling window. Persistent overconfidence costs emission weight.

03 — The loop

Task → miners → verification → reputation → emissions.

Better verified intelligence → higher reputation → higher emission weight → stronger miners. Hover a stage for detail.

How miners work

  1. 01 Receive a task request containing only prompt, category, difficulty, nonce and deadline.
  2. 02 Solve it with any backend — a local model, an OpenAI-compatible endpoint, or a bespoke solver.
  3. 03 Return an answer, a calibrated confidence, evidence and execution time, echoing the nonce.
  4. 04 Expect a mutated follow-up at any time; consistency is scored, not just correctness.

How validators work

  1. 01 Generate a task and commit to its hidden answer with an HMAC commitment.
  2. 02 Dispatch to a strategy-defined subset of miners and enforce nonce, deadline and rate limits.
  3. 03 Grade every response deterministically and probe robustness with mutations.
  4. 04 Smooth scores into reputation and publish normalised weights that sum to one.
04 — Scoring

Five dimensions, one weight vector.

45%

Accuracy

Deterministic verification against hidden ground truth

20%

Evidence quality

Concept coverage, structure and specificity

15%

Robustness

Consistency under semantics-preserving mutation

10%

Calibration

Brier score of stated confidence vs outcomes

10%

Latency

Budget-based, never a race to the bottom

05 — Anti-gaming

Memorisation is not a strategy.

  • Hidden ground truth

    Answers exist only inside the validator boundary; no API path returns them for an open task.

  • Dynamic generation + mutation

    Every task is drawn from a seeded generator, then optionally mutated into a variant with the same answer.

  • Duplicate & collusion detection

    Repeated answer fingerprints and byte-identical reasoning across miners are flagged and penalised.

  • Replay protection

    A response is bound to one task id and nonce; reuse is rejected before scoring.

06 — Architecture

Readable by a subnet engineer, not just a judge.

subnet/protocol

Task + response wire types. Ground truth is structurally absent from every miner-facing message.

subnet/tasks

Generators, mutation engine and deterministic verifiers. Ground truth is computed, not written.

subnet/miner

Pluggable model backends plus nine behavioural profiles for the local simulation.

subnet/validator

The pipeline: generate → dispatch → validate → score → probe → reputation → emissions.

subnet/scoring

Weighted scoring, Brier calibration, EMA reputation, emission normalisation, anti-gaming detectors.

subnet/adapters

One interface, two implementations: SimulationAdapter today, BittensorAdapter for testnet.

07 — Roadmap
Now

Local subnet prototype

Full pipeline, 13 generators, deterministic scoring, emissions and 120 automated tests.

Next

Testnet registration

Wire the BittensorAdapter to a funded testnet hotkey and publish weights on chain.

Then

Commit–reveal on chain

Publish the task commitment on chain before dispatch so grading is externally auditable.

Later

Verification marketplace

External clients submit claims for verification and pay for a scored, evidenced verdict.