Define
Freeze the parent model, metric, deadline and conditions.
A MORE HONEST AI FUTURE
DON'TGUESS.PROVE IT.
Post a measurable model failure.
Compete to fix it. Verify the improvement.
BETTER MODELS
THROUGH OPEN
COMPETITION
THROUGH OPEN
COMPETITION
001 / THE EVIDENCEMODEL IMPROVEMENT / OBSERVED
BEFORE84.72%Frozen baseline
recorded local run
recorded local run
AFTER95.28%Selected candidate
same evaluation set
same evaluation set
RECORDED LOCAL CPU EXPERIMENT360 HELD-OUT EXAMPLES · NO ON-CHAIN PAYOUT
FROM FAILURE TO VERIFIED IMPROVEMENTFour steps.
Four steps.
One measurable result.
Compete
Independent builders submit bounded model updates.
Evaluate
The named evaluator tests every admitted fix on equal terms.
Settle
A verified eligible winner can receive the escrowed bounty.
002 / THE PROOFEvidence is the
Evidence is the
whole point.
No invented payouts. No unverifiable accuracy claims. Inspect the original model, the signed artifacts and the rules used to pick a winner.
VERIFIABLE LOCAL EXPERIMENTRESULT 001 / DIGITS CLASSIFIER
01 / Frozen model84.72%Baseline accuracy
02 / Accepted model95.28%Measured held-out accuracy
+10.56percentage points
measured improvement
measured improvement
Loading the signed run evidence…
Inspect evaluation & model passportPARENT → WORKERS → EVALUATOR → SELECTED MODEL
360 HELD-OUT EXAMPLES3 INDEPENDENT PROCESSES ON ONE MACHINELOCAL · UNPAID
003 / THE ARENALet the work
Let the work
speak for itself.
Browse the actual recorded bounty. Inspect every candidate, the frozen rules, the artifact boundary and the evaluator's signed receipt.
Open the evidence consoleThis deployment displays a recorded local experiment. A public live bounty service and funded Devnet payout have not been established.
Reading the evidence
Only real run data appears here.
Build a better modelRun a worker from your own machineSHOW COMMANDS +
Research, not theatre.
The working proof uses actual CPU training. Worker keys stay on your machine. No paid API or GPU is necessary for this example.
004 / THE LIMITSTrust is not
Trust is not
a visual effect.
The named evaluator decides model quality. Hashes identify exact bytes, not scientific truth. Our interface makes those boundaries explicit.
One named evaluator
The evaluator scores models and authorizes settlement. This is not decentralized ML consensus.
A bounded benchmark
The public Digits evaluation split is reproducible, not a secret enterprise anti-cheating set.
No invented money
The Solana escrow program is locally tested. No completed live Devnet payout is claimed.
How candidates become eligible
At the deadline, the evaluator compares registered candidates against the same frozen baseline and held-out set. Observed improvement must meet the bounty threshold and a one-sided, multiple-candidate-adjusted paired-bootstrap lower bound must be positive. The statistics are approximate under IID assumptions, and do not eliminate leakage or selection bias across repeated rounds.