{"study":"SETAEI Error Learning Lab","version":"toy-error-lab-v1","lockedAtUtc":"2026-10-24T15:06:08.212Z","runnerSha256":"42b18c75ca2b69a6e167b008ec12fcf8fe7deb8ad1ab69d394a740c0c9029997","studySeed":"0bbfc6d9aae8b018bd238c2548749c8af9a6d0e8c6559a7c6698afd2b71b97d3","wordBits":16,"featureBits":80,"counts":{"train":512,"validation":512,"test":2048},"epochs":24,"plannedRounds":14,"controls":["learnable_permutation","independent_random","fixed_hash"],"nightlyTime":"03:15 Europe/Berlin","procedure":"Mistake-driven batch perceptron. Only incorrectly classified training bits contribute updates. A training word weight is 1 + fraction of its wrong bits. Select the earliest best validation epoch; open the test set once after selection.","split":"One seeded permutation of 65,536 synthetic 16-bit words. Each round reserves 3,072 disjoint words. No prior train/validation/test word reappears in another round.","inference":"Paired 95% normal-approximation CI on correct-bit differences per test word. Primary comparator is an independently generated random word. All daily estimates are exploratory; no real-wallet deployment or efficacy claim is permitted.","restart":"Per-round initial model snapshots remain fixed across retries. Interrupted attempts are recorded; only a published completed round advances the counter.","horizon":"14 scheduled rounds including pilot. No favorable-result early stopping, no extension based on results. A final review must include every completed and failed round.","privacy":"Four-hex-character toy words only. No wallet addresses, real private keys, funds, balance APIs or live-agent job data are inputs.","limits":"A successful learnable control only validates this pipeline. The fixed-hash control has a tiny exhaustive word space and is not a security proof. Current Ethereum model errors are described separately and are not training inputs here."}