The personal archive of Konstantin E. Tsiolkovsky (1857-1935) is held as fond 555 of the Archive of the Russian Academy of Sciences. It has been scanned and put online, but without a queryable catalogue, full-text search or dataset. This record contains a machine-readable catalogue of all 2,019 files of the fond across its five inventories, covering 51,008 scans, with the real archival inventory and file numbers, so that each row is a citable archival reference. It also contains the retrieval and measurement pipeline that produced it: a resumable downloader that resolves each file's inventory and number from the scan paths, a page classifier that separates handwriting from typescript without a neural network, and tools that score transcriptions both against published texts on Russian Wikisource and, where no published text exists, against the typed copy of the same text inside the same archival file. Measured over the whole fond: 51,008 scans retrieved and structurally verified with none corrupt; 68% handwritten, 29% typewritten or printed, 3% notes and covers. This release contains the fond transcribed whole: all 2,019 archival files, all 51,008 scans, with uncertainty and authorial deletions marked explicitly - 303,977 uncertainty marks and 38,465 passages struck out. On accuracy, three figures that measure different things and should not be merged. Against published editions, on the part transcribed by the earlier pipeline: 98.1% of characters on typescript and 81.1% on handwriting. Against published editions across the finished corpus, on the ten files where the archival file holds the redaction that was printed: 92.4% of characters. Where the file holds a draft or the materials for an article rather than the printed text, agreement falls to about 24%, which is the distance between manuscript and edition rather than misreading, and those files are reported separately rather than averaged in. Where no published text exists, accuracy is bounded by calibration: two readings of one and the same text - a handwritten sheet against the typed copy of it in the same file - agree on 37% of words over 1,371 pairs from 194 files. That figure is the same for the part read by the earlier pipeline and for the part read later by a different model, which is the evidence that the two parts are of one quality. The corpus is a machine transcription with uncertainty marked. It has not been checked by a human against the scans and is not a scholarly edition. The catalogue and transcriptions are released under CC0. The code in the repository is MIT.