How VERITENSOR prices trust
The mechanism, end to end, with the live configuration values the running subnet is using.
Why VERITENSOR?
The gap this subnet fills
AI output is abundant; verified AI output is not.
Anyone can generate a plausible answer. Establishing that an answer is correct still requires either a human or a deterministic checker, and neither scales with generation volume.
Static benchmarks decay into answer keys.
The moment a benchmark is public it measures recall. A verification network must generate its problems continuously and keep their solutions private.
Bittensor prices exactly this kind of work.
The subnet needs a scarce, objectively measurable output. “Answers that survive independent verification” is measurable, hard to fake, and useful to anybody deploying models in production.
The loop
Validator → task → miner → verification → scoring → emissions
Why miners compete
The economic argument
Emission weight is a monotone function of reputation above a floor, sharpened by a temperature exponent of —. A miner that improves its verified accuracy does not gain a linear amount of reward — it gains a superlinear amount, because the surplus above the floor is raised to that power before normalisation.
At the same time the per-miner cap of — prevents a single dominant miner from absorbing the subnet, which keeps it worthwhile for new entrants to register.
Net effect: better verified intelligence → higher reputation → higher emission weight → capital flows to the miners that are actually right.
Why validators matter
Integrity of the measurement
Validators are the only holders of ground truth. They commit to the hidden answer with an HMAC over (task id, nonce, answer) before dispatch, so they cannot retrofit an answer once responses arrive.
Each validator runs an independent task engine, guard and scorer with a different strategy — coverage, probe rate and category mix all vary — so a miner cannot overfit to one evaluator's habits.
Consensus is reported for observability but is deliberately not used as truth for programmatically verifiable categories. If a colluding majority could define correctness, the subnet would measure agreement, not accuracy.
Scoring formula
Live weights from the running configuration
Calibration, interactively
Move the sliders: this computes the exact Brier formula the backend uses.
brier = p·(1 − c)² + (1 − p)·c² · calibration = 1 − min(brier, 0.25) / 0.25
0.3625
Calibration component0.000
A miner claiming 95% while being right 60% of the time scores 0.000 on this dimension. Honest, discriminative confidence — high when right, low when unsure — is the only way to earn it.
Anti-gaming
What each defence actually stops
Hidden ground truth
Answers live only in the validator process. No public endpoint returns them for an open task; the admin reveal route refuses non-closed tasks.
Dynamic generation
Each task is drawn from one of 13 seeded generators, so the surface text is new every time.
Mutation probes
A correct answer may be followed by a semantics-preserving variant. Consistency is 15% of the score.
Replay protection
A response is bound to one (task id, nonce, miner). Reuse is rejected before scoring, with a full-score penalty.
Duplicate detection
Repeated answer fingerprints on open-ended tasks and reused evidence bodies are penalised; enum answers are exempt up to a high threshold to avoid false positives.
Sample requirements
No emission below 10 scored tasks, and reputation is shrunk toward a prior below 20 tasks.
Limitations are documented rather than hidden: none of these defeat a genuinely capable colluding cartel, and duplicate detection is a heuristic. See docs/ANTI_GAMING.md.
Adaptive difficulty
Keeping the subnet in the informative band
The network is currently scoring —, so validators with adaptive policies are drawing from the easy band. Thresholds are configuration, not constants in the code.
Why this becomes a marketplace
Beyond the hackathon
Verification is a measurable commodity.
A reputation number backed by thousands of independently graded, mutation-tested tasks is a price signal for reliability — something no public leaderboard currently provides.
Demand exists outside the subnet.
Any team shipping model output into production needs an answer to “how often is this right, and does it know when it is not?” The same pipeline can grade externally submitted claims.
The mechanism scales with new families.
Adding a verifiable domain means writing one generator plus one deterministic verifier. The scoring, reputation and emission machinery is unchanged.