read the PoA whitepaper agAin. same question keeps coming back.
verifiable. that word appears throughout OpenLedgers documentation. verifiable data provenance. verifiable attribution. verifiable AI. the entire value proposition RESTS on PoA making AI systems verifiable.
verifiable means independently confirmable. in cryptography verifiable means a third party can check a proof without trusting the prover. in AUDIT contexts verifiable means following a documented procedure produces a consistent result. PoA claims to make AI attribution verifiable a third party should be able t0 confirm that a specific dataset contributed to a specific model output.
the PoA whitepaper describes tw0 attribution methods. influence function approximations for SMALLER models. suffix array token attribution for LLMs. both methods produce attribution scores. the scores determine contributor payments.
but verifiable requires a verification standard. a defined DEFINED procedure that any third party follows to arrive at the same result. the PoA whitepaper describes the attribution methods. it does not publish a verification standard. a third party who wants to independently verify a PoA attribution calculation has no published procedure to follow.
influence function approximations produce different results depending 0n which implementation you use which subset of training data you sample and what approximation parameters you select. two independent implementations of influence function attribution on the SAME model and dataset can produce different scores.
if two independent implementations produce different scores which one is correct? without a published verification standard that specifies exact implementation parameters there is no answer. the attribution is not verifiable in any meaningful technical sense? it is calculable by one method by one implementation controlled by ONE team.
that is not verifiable. that is trusting the calculator.
watching published verification standard for PoA with exact implementation parameters independent third party VERIFICATION of any historical attribution calculation whether governance covers verification standard definition.