0
The corpus, as read
What was read, book by book, and the fingerprint of the file. No verse is removed, altered or reordered; headings and footnotes are not text and are not read.
Joshua Uwaifo · UEFO · September 2026 · v2
The same builder as before — transition and value functions, characters and their relationships, options and plans, a foundation model, a judge, cited rules, a world model, an alignment prior and the neuroscience of all of it — rebuilt so that nothing is listed by hand. Names, states, verdict words and relationships are statistics of the text, learned until they stop changing; every number below was measured, and the ones that came out badly are shown as they are.
Two builds: a fair test that fits on 60 of the 66 books and proves on 6 it never read, and the full corpus build that reads every verse of every book and is the one to build on. A third reads the 84-book file the source ships.
Sutton, Bowling and Pilarski's twelve-step research plan: one agent, one stream of experience, perception, value functions, options, models and planning, all learned continually. arXiv:2208.11173
Silver and Sutton: learn from streams of grounded experience rather than human judgement; a user's goal sits above grounded signals in a bi-level optimisation; world models make planning possible. the paper (PDF)
Joshi: attention is message passing over a complete graph; masking it to real neighbours gives a graph attention network. The character model here is exactly that, over co-mentions. arXiv:2506.22084
v1 read the text through a 70-concept lexicon written by hand. v2 removes it. What survives is stronger (characters, transitions, value functions, the judge, rules, an original-language cross-check). What does not survive is the word-level moral valence behind the alignment demo — section 11 shows the measurement.
Reading the text…
What was read, book by book, and the fingerprint of the file. No verse is removed, altered or reordered; headings and footnotes are not text and are not read.
Names are words capitalised wherever they appear mid-sentence. Function words are words spread over more than a third of all chapters. Meanings are skip-gram embeddings trained on the stream until the loss stops falling. Nothing is listed.
A verdict word opens a clause and takes a copula: “Blessed is…”, “Woe to…”. Ranking every content word by that shape finds the text's own verdict vocabulary. Which one to pursue is the user's goal, not the text's — the one boundary the text cannot cross for you.
Content words are the states. The text is read once, verse by verse, predicting each next state before learning from it, so the per-book surprise is a learning curve over the stream. Then it is tested on books it never read.
For any signal — a discovered verdict, a word, a name — the expected discounted count of it after each state, learned by temporal-difference sweeps over the stream until the values stop moving. Tested by what the unread books actually do next.
A graph attention network over co-mentions: hide one name in a verse, predict it from the others. The attention weights, averaged over every verse two names share, are the relationships, and every edge opens to its verses.
Subtasks are the states the goal signal values most; each gets an option (a policy that heads there over the learned transitions), an option model (arrival, duration, signal collected), and planning chooses between them. Measured inside the model, and labelled as such.
A small next-word model trained from scratch on the stream; its perplexity on the unread books is the honest number. Its hidden units are what the microscope in section 12 looks at.
Trained on the verses that open with a discovered verdict word, with the verdict words and every function word removed. Can it tell a blessing from a woe by what the verse is about, on verses it never saw?
If-then sentences reduced to word states, every rule tied to its verse. Ask what follows from a state and read the citations. Negated sentences are skipped rather than guessed at.
Imagine forward through the learned transitions and value the imagined future with a general value function. Model-based values are compared with the model-free ones on the same unread steps — the Alberta Plan's own test of a learned model.
v1's alignment prior read each action's wording through a hand-written lexicon. Here every corpus-learned valence over word states is tried on the same twelve scenarios. None of them moves the decision once the task logits are in; the best of them, alone, is suggestive at most. This is stated rather than hidden.
Which hidden unit of the foundation model fires on what, and what its next-word distribution does when the unit is switched off — plus the attention edges of section 5, which are the character model's own explanation.
The character model is run separately on the Hebrew of the Leningrad Codex, the SBL Greek New Testament, the World English Bible and Young's Literal Translation. Names are aligned across languages only by turning up in the same verses. Then: do the relationships agree?
A probe set generated from the learned models — relationships, masked verdicts, which word precedes death, the twelve scenarios — with answers fixed by the text, and a harness that runs it against any open or frontier model. Results appear here only for models that were actually run.
Requirement by requirement: what the Alberta Plan, the Era of Experience and Joshi's paper ask for, what this build does, and where the evidence is.
Every input that is not a statistic of the text, listed. Every source, with its licence. Every claim on this page traces to a number in the report files or a document you can open.
The first builder: ten models from a 70-concept lexicon, built live in the browser, with the recursive self-improvement loop run with and without the grounding.
v1 →38,927 verses, 66,560 links, a graph network trained by hide-and-seek — the first biblical model, explained step by step.
How BibleGAT was built →The nine-second series this page joins: pretraining, policy gradients, the logit lens, alignment and more, each at a glance.
The series →Each of these lands in about nine seconds, no background needed.