Proven · Live demonstration
One epoch.
A hundred submissions.
One proof.
Watch a single round resolve through the Verification Funnel, the way a validator sees it. Submissions enter as noise; the funnel rules out the cheap failures first, then scores what survives by the faults it kills.
Faithful simulation · the mechanism is real, the data is generated on the spot, nothing touches a live subnet
How Proven works
Proven turns software testing into a market. A validator hands out a spec, never your source code; a swarm of miners act as adversarial QA engineers, writing Playwright suites that try to break it; the network scores each suite by how many injected faults it kills. What survives is Proof of Assurance.
How Proven reaches consensus
- 01
Spec
A validator's Synthetic Spec Engine writes a fresh specification and deploys the target app. Your source code never leaves your repo.
- 02
Mine
A market of miners reads the spec alone and races to synthesize Playwright suites that try to break the app.
- 03
Verify
Each suite runs the funnel in turn: lint, then the clean reference, then a horde of mutated builds. The cheap failures are ruled out first.
- 04
Consensus
What survives is scored on the faults it kills. Winner takes the epoch, settled on-chain and paid in TAO.
Illustrative testnet figures
Two surfaces, one standard
Frontend
What the user touches, proven against reality.
- User flows, end to end
- Playwright UI automation
- Visual & interaction faults
- Accessibility & edge cases
Backend
What the user trusts, proven under pressure.
- API & contract conformance
- State machines & invariants
- Error, timeout & race paths
- Auth, access & abuse
This round’s spec · Synthetic Spec Engine
SaaS dashboard · SPEC-1427
Given a viewer-role account, when the user opens the billing panel, then every 'Upgrade plan' control must be non-interactive.
Miners receive this specification and nothing else. The application’s source code never leaves the validator. It is the only input to every suite in the field below.
Input · from validator via Synapse
{
"spec_type": "user_story",
"content": "Given a viewer-role account, when the user opens the billing panel, then every 'Upgrade plan' control must be non-interactive.",
"target_url": "http://target-app:8080/billing",
"timeout_ms": 9000
}- 01cost · very low
Static Gate
Lints the script. Broken or malformed code is ruled out at near-zero cost.
—passed - 02cost · medium
Reference Gate
Runs the suite on one clean Reference app. It must pass 100%, or Pclean = 0. Catches false positives, over-strict assertions, and DOM-crawl timeouts before any mutants spin up.
—passed - 03cost · high
Mutant Horde
Spins up 20+ mutated builds. The score is the kill ratio, scaled by efficiency.
—scored · 20 mutants
Submission inspector
Select any cell in the field, or a leaderboard row, to read the submitted Playwright suite and how the funnel judged it.
Epoch resolution · winner takes all
Scoring against the Mutant Horde…
How a score is computed
Si=Pclean·α ·KiNmut·Ei
A score only exists once a suite clears both gates. After that, it is the kill ratio, scaled by how cleanly it ran. Winner takes the epoch.
- Pclean
- Binary switch. 1 if the suite passes the Reference Implementation 100%, else 0. One false positive ends the round.
- Ki / Nmut
- Kill ratio. Mutants killed over mutants generated. The efficacy of the suite, and the heart of the score.
- Ei
- Efficiency decay. Bloated or DOM-crawling suites that exceed the execution window are penalized here.
Discouraging gamification
Assertion Roulette
A suite blind-asserts a failure to rack up kills. It must pass the clean Reference Implementation first; any false positive flips Pclean to 0 and zeroes the score.
Hard-coded answers
A miner memorizes a known build. Mutants are generated dynamically at runtime, so yesterday's answers never generalize to this epoch's horde.
Plagiarism
A miner copies a top suite. Validators track AST and semantic similarity; under first-to-chain advantage the original submitter keeps the whole score.
Probing
A suite crawls the DOM to map the environment instead of testing the spec. Containers are isolated and timed; excess crawling blows the execution window and the efficiency decay E cuts the score.