What the Filters Kept (EA-3I-ATLAS-NOTEBOOK-01, version 1.2) is the working notebook of a public-data search for a signal from the interstellar object 3I/ATLAS, written under the heteronyms Nobel Glas and Orin Trace and deposited in the Crimson Hexagonal Archive on 6 November 2026 for Transactions on Substrate Engineering. The investigation ran from 2 to 6 November 2026 on Breakthrough Listen's public products from the Green Bank Telescope, the Allen Telescope Array, Parkes and MeerKAT, with Lee Sharks directing two machine collaborators, Claude and ChatGPT, who checked each other's work through code, candidate tables and hashes. Every deciding test had its plan frozen and hashed before its counts, and every null was built as a set of false frames constructed the same way as the true one. The notebook identifies no artificial transmission. It reports one unresolved result: thirteen comet-frame coincidence clusters near 1666 MHz in the Green Bank cadence of 18 December 2025, which reproduce under an independent implementation, depend on normalization and threshold, and do not replicate in sixteen held-out channels or at Parkes. Its findings that hold concern representation: a Google AI Overview presented Avi Loeb's position, held at Level 4 of his scale from July 2025, as a later concession; the Allen Telescope Array paper filters for positive drift where its stated acceleration gives a tone on the object a negative one; and a public statement of that search's null dropped its frequency and drift scope. The notebook sets the archive's instruments against the record, among them instrument-conditioned nullity, correlated blind spots, disciplinary model collapse and frame conversion, and runs the sophistry measure for the first time. Its companion is Lee Sharks's public account A Firm No (#1743). Version 1.1, the same day, records the operator's reading that the published searches were built to rule a signal out at the least cost without the work ruling out requires. Version 1.2 (6 November 2026) makes factual repairs in place: MeerKAT's technosignature search, with its limit of 0.17 W over 900 to 1670 MHz (ATel #17499), is added as a fourth published search, and the notebook's citations of Loeb's April essay and of the PER-D rubric are corrected.