Everything on this site descends from one body of physical evidence: Olaf Stapledon's own lecture notes, held in the Olaf Stapledon collection at Liverpool University Library. The project holds scans of 223 sheets across fifteen sets of notes — pencil manuscripts, mostly abbreviated, underlined and margin-annotated for the platform — centred on Humanism and Superhumanism, delivered to the South Place Ethical Society on 30 December 1934 (catalogue reference OS/F11/4). Alongside the working scans, the project later received the University of Liverpool Digital Heritage Lab's 72-megapixel archival captures of all fifteen sets, with IIIF catalogue metadata that independently supplies each lecture's official date and reference. This page explains, step by step, how those sheets became the transcriptions, editions and analysis you can read here — and where the uncertainty honestly remains. The full technical narrative is in the project report; the public correction trail is the errata page.
No single machine reader was trusted with the manuscripts. Every one of the 223 sheets was transcribed independently by three frontier vision models from different families (OpenAI's GPT‑5.6‑sol, Anthropic's Claude Opus 5, and Google's Gemini 3.1 Pro), each producing a formatting-preserving transcription: indentation levels, single and double underlining, strikethrough, insertions, red and teal pencil, boxed and marginal text, and Stapledon's private shorthand (θ/ð for “the,” ∴, ∵). A fourth pass then put the original scan beside all three transcriptions and adjudicated every disagreement against the image.
The point of the ensemble is that models fail differently. On the very first page transcribed, one model read “Birmingham West” where two correctly read “Birkenhead West” — Stapledon's home ground — and “workers wanted higher wages” against the correct “workers wrested higher wages.” Where the models genuinely could not settle a word from the page, the disagreement was not hidden: disputed readings are flagged inline, with the alternatives preserved (in the manuscript decks, hover over a dotted-underlined word to see them), and illegible spans are marked as such. Across the whole corpus only 116 words needed a disputed flag — about one per two pages; the Humanism lecture carries just ten.
The consensus transcription was then attacked from three further directions.
Reconstructing a spoken lecture from outline notes involves editorial judgement, and this project's answer is not to hide that judgement but to publish it three times, under three explicit contracts:
The method itself was calibrated against ground truth. One lecture, Interplanetary Man? (1948), survives as both manuscript notes and Stapledon's own published text. Comparing the two established how he actually expanded outline into speech — roughly 10–20× per note-line, with real, load-bearing citations and a plainer platform voice than his literary prose — and the reconstructions were tuned to those findings.
The deep analysis reads the lecture against everything Stapledon wrote that the project holds. Its quotations are not on the honour system: every substantive quotation from the held corpus carries an invisible locator marker that the build resolves against the corpus text itself — an exact match, whitespace aside, or the build fails. The full build currently passes with all 1,156 corpus citations across the five analysis pages (the three monographs and the two lecture dossiers) resolving; each renders as a link into the anchored reader editions with an expandable excerpt of the quoted passage in context. The handful of quotations that could not be located verbatim in the held corpus are deliberately left unlinked and listed in the build report rather than approximated.
In July 2026 the project was submitted to three independent audits by three
different model families — GPT‑5.6, Kimi K3 Max, and Fable 5 — whose
reports are preserved unaltered in the project's review/ folder. The project's owner
adjudicated their findings under a fixed rule: no audit-claimed correction was applied on the
audit's authority alone. Every claim was re-verified against a primary source — the
consensus transcription, the 72-megapixel captures, or the published corpus — before being
applied.
The outcome: roughly 274 applied corrections (about 200 distinct findings), and 24 audit claims rejected with evidence — including one where the audits were confidently wrong in an instructive way: three auditors asked the project to honour a strikethrough on the word “other” that, at full archival resolution, turned out not to exist (the only crossbar on the page is the “t” of “other”). The correction ran the other way, into the transcription. Two further re-audit rounds followed: the first found twelve findings that the fix pass had claimed but not landed — all repaired — and the second verified all seventeen claimed repairs against their primary sources and ruled the artifacts converged, with no corpus-checkable content error remaining. A permanent propagation checker now scans every project-authored page for the known-bad strings from both rounds and fails on any reappearance. The user-facing record of the whole episode — what was wrong, what was fixed, what was rejected and why — is public: the errata page.
In August 2026 the faithfulness audit itself went corpus-wide. Every prose section of the model-generated reconstructions was audited against the verified notes by the same three model families, the findings clustered, and every cluster on which two or more providers agreed was adjudicated against the manuscript sheets themselves — never on the auditors' authority alone. That adjudication is now complete for all thirteen generated lectures: 604 ensemble-agreed clusters produced 572 sheet-verified repairs and 32 rejections with counter-evidence (one early rejection was conceded and repaired in a confirmation round), with the patched text re-audited in confirmation rounds until the flags converged; the pages were then rebuilt from the repaired text. Together with Humanism's full pedigree and Interplanetary Man? — where Stapledon's own published text is the ground truth — all fifteen reconstructions now carry an audited and adjudicated faithfulness status. The remaining honesty note is itself on record: the roughly 1,290 findings flagged by only a single auditor remain recorded but unadjudicated (beyond ~50 string-level fixes). The full per-lecture records — every correction with its sheet evidence, every rejection with its counter-evidence — are preserved in the project's working folder, and the headline corrections are on the errata page.
One book got the manuscript treatment end to end. The corpus text of Latter-Day Psalms (1914) — previously a scan's embedded OCR layer, riddled with running heads and garbling — was replaced by a fresh transcription of all 68 page images at 300 dpi, reconciled word-by-word against the old text with every non-trivial divergence adjudicated on the page image. The result is a poetry-native edition: linked contents, exact verse lineation, the 1914 pagination in the margin, and a citation anchor on every verse. The reconciliation even preserved the book's own printer's error (“happines,” p. 32), now quoted with [sic] rather than silently tidied.
This project was built end to end by AI models — the transcription, the reconciliation, the reconstructed prose, the analysis, and the audits of all of the above were machine work, with the models' disagreements used as the primary quality instrument. The human role was to supply the materials, set the rules, adjudicate the audits, and spot-check: a few key Humanism sheets were verified word-by-word by eye (yielding two corrections and the decisive reading of the date), but there has been no independent human collation of the full 223-sheet corpus. Anyone with access to the Liverpool archive could perform one, and the apparatus is designed to make that easy rather than embarrassing.
Uncertainty that survived every pass is still uncertainty, and it is left visible:
If you find an error, the trail is public: the errata record what has already been caught, the report documents the method in full, and every page of the manuscript deck puts the original sheet beside the transcription so you can check the project's readings against the evidence yourself — which is, in the end, the whole design.