M-ONE
M-ONE decides which source your agent should trust, verifies the answer before it ships, and labels exactly how confident it is. It never claims data is true. It reports how much it knows, and it has never once overstated that.
Anyone can index endpoints. The hard part is deciding which one deserves the call, and then admitting how much the answer is worth once it arrives. That is the whole of M-ONE.
Deciding where the call goes
Every source carries a score built from measurement rather than from a listing: how often it delivers, how quickly, how fresh its data is, and how often it disagrees with everyone else. The call goes to whichever source is most likely to work at that moment.
Routing is blind to price and to stake. Nobody can pay to be picked, and no amount of collateral moves a source up. That is a tested invariant in the engine rather than a policy we promise to follow.
Checking before delivery
The response is checked before it reaches you. Structure, freshness, plausibility, agreement with independent sources, and, where a chain or an oracle provides ground truth, agreement with that. A source that fails verification is swapped inside the same request. Your agent sees a slightly higher latency and nothing else.
The label on the answer
anchored means the answer agreed with an independent authoritative reference, typically on chain state or a signed oracle. consensus means independent sources agreed, where independence is measured rather than assumed. single_source means the answer passed structural and plausibility checks with no independent confirmation available. It may be right. We could not confirm it, and we say so.
No level guarantees correctness. Anchored means it agreed with a reference at a point in time, and the reference can be wrong too. Where the stakes are high, set the minimum level your agent will accept, and below it M-ONE refuses rather than dress an unconfirmed answer as a confirmed one. That refusal is the product working.
Built against adversaries first
M-ONE was pointed at liars before it was pointed at the real world. Against a source returning plausible wrong numbers that pass every structural check, wrong deliveries fell from 78.5% to 0.5%.
Against five providers with five different names reading one shared upstream, the engine had been producing 575 confidently wrong consensus deliveries. Afterwards, zero. The data did not improve. The label became honest.
Two sources that lie together, in a category with no ground truth, are informationally indistinguishable from the truth. No statistical judge resolves that, and M-ONE does not pretend otherwise: it labels those answers unconfirmed, every time. Across every adversarial scenario the mislabelled rate was 0%.
Measured against the real world
500 live calls against independent exchange feeds, judged against a signed on chain oracle. Zero lies and zero unavailable, with a lying rate below 1% at 95% confidence, reported as an interval because a rate reported as a single number is not a measurement.
Two exchange feeds tested at residual correlation r=0.154, genuinely independent. Two social providers tested at 100% exact agreement across eleven handles, identical to the byte. Those were not two observations. They were one, counted twice, and the engine refused to call it consensus.
What ships with every answer
Every response carries the level it earned and a trace of how the decision was made: which sources were considered, what each one scored, which attempts ran, and why one won. This is the part that matters. A number without its reasoning is a number you have to take on faith, which is the position your agent was in before Metera existed.