Colophon

How a Plainspoken Classic is actually made

Most publishers would hide this. We think it's the most interesting thing about these books.

Every volume in this library is translated by a fleet of AI translators working in parallel — dozens to hundreds of them, each assigned a short passage of the Greek. They are orchestrated by a stronger frontier model that acts as the editor-in-chief: it writes the translation charter every agent must follow (one voice, one register, one set of conventions across the whole library), assigns the work, audits the results, and dispatches repairs when anything falls short.

The rules every translator-agent works under

1. Greek only. Each agent sees the source text and nothing else — it is forbidden to consult any existing English translation, and for scripture it may not even open the English columns of our own database.
2. Honest seams. Passages are cut mid-sentence at batch boundaries; agents must translate exactly what is in front of them and never invent a completion — the neighboring agent supplies the rest.
3. The charter. Contemporary scholarly register, readable aloud; Greek gods keep Greek names; ancient units stay ancient; epithets rendered freshly but consistently; no archaisms, no paraphrase-drift.

Then the checking starts

1. Adversarial accuracy review. Independent referee-agents re-read the Greek behind a rolling sample of batches, hunting specifically for mistranslation, omission and invention — quoting the Greek for every check.
2. Mechanical independence scanning. The finished text is scanned, phrase by phrase, against every prior translation we hold. Long verbatim runs get flagged and re-translated until the wording is verifiably its own. The resulting numbers are published on the proof page.
3. An outside judge. A separate comparative review graded famous passages against the classic translations and the Greek — the verdict is quoted, unedited, on the proof page.

Why work this way?

Because it produces something the old publishing model never could: an entire ancient library in one consistent voice, finished in months rather than lifetimes, at prices measured in dollars rather than hundreds — with its accuracy and originality measured and published rather than assumed. A single human translator is a lifetime per corpus. A fleet with a good editor is a season.

On updates — a plain-spoken promise

As AI translation gets better, these editions can get better with it: errors fixed, phrasing sharpened, whole volumes re-rendered by stronger fleets — and improvements ship to everyone, because everything here is free. In the same spirit of honesty that runs through this whole page: updates are an intention, not a guarantee. This is a small independent press; if the library never finds its readers, it simply stays as it is — complete, accurate to the best of the checks published here, and freely downloadable in full, forever.

Who does the work: Claude Fable 5 holds the pen. Full transparency — here is the entire setup, role by role. The fleet is built from Anthropic's Claude models, directed by Claude Fable 5, Anthropic's most capable model. Fable 5's specific jobs:

• Coordination — designs and runs the whole pipeline: batching the Greek, dispatching the fleets, assembling the books, deciding what ships.
• Arbitration — every finding a referee raises goes to Fable 5, which verifies it against the Greek and fixes it or dismisses it (referees overcall; several "defects" turned out to be fine on inspection).
• Hard passages — anything the fleet flags (mid-sentence fragments, rare vocabulary, tangled referents) is re-translated by Fable 5 directly.
• Independence rewrites — when the mechanical scan finds wording that echoes an earlier translation, Fable 5 rewrites those sentences fresh from the Greek until the scan reads zero.
• Whole short works — completeness-critical texts are Fable 5's alone (the Crito in this library is its work end to end: in our three-model head-to-head on the same Greek, Fable 5 was the only model that never silently dropped a line).

Claude Sonnet — the translation fleet. Under Fable 5's direction, hundreds of Sonnet agents each own one ~1,500-word batch of the source text:

• Bulk translation — every batch is translated from the Greek or Latin alone, under the same charter prompt: contemporary register, familiar names, never consulting an existing English translation.
• Aligned output — each agent must return exactly one English section per source section; a count mismatch fails the batch automatically and it is re-run.
• Why Sonnet — in our trials its prose quality matched the larger models on ordinary narrative, and it follows batch instructions most reliably; its known weakness (rare silent omissions) is exactly what the mechanical gates catch.

Claude Opus 5 — the independent referee. A separate model, deliberately not the one that translated. Opus 5 took this seat on 24 July 2026, the day of its release; volumes finished before that date were refereed by its predecessor, Opus 4.8, under the identical protocol:

• Blind re-reading — Opus agents re-read the original language behind samples of every book, sentence by sentence, hunting for omissions, inventions, reversed meanings, and wrong referents.
• Findings only — referees never rewrite; they file findings, which Fable 5 verifies against the source and either fixes or dismisses (referees overcall — several reported "defects" proved fine on inspection, and dismissing those is part of the design).
• Why Opus — in head-to-head trials it showed the sharpest eye for hard philology: rare vocabulary, mid-sentence fragments, quotation shape.

Did the referee upgrade actually matter? We tested it.

When Opus 5 replaced Opus 4.8 in the referee seat, we didn't just assume the new model was better — we ran the experiment, and we're publishing the result either way. On 29 July 2026 both models were given the identical passage (Marcus Aurelius, Meditations 2.1–9, 574 Greek words), the identical prompt, through the identical harness, and the two translations were judged line-by-line against the Greek by Fable 5.

The result: both were accurate and fluent — but Opus 5 was consistently sharper on exactly the skills a referee needs. Concrete examples from the head-to-head:

• Anatomical νεῦρα: 4.8 wrote the anachronistic "nerves"; 5 wrote "sinews", correct for ancient usage.
• The adverb στιβαρῶς ("stoutly"): 4.8 silently absorbed it; 5 kept it as its own word.
• ὑφορᾶσθαι: 4.8 paraphrased "shrink from"; 5 gave the exact "eyeing with suspicion."
• The legal metaphor προθεσμία: 4.8's flat "appointed term" vs 5's living "taken an extension from the gods."
• ἐμούμενον: 4.8 softened to "spat out"; 5 kept Marcus's blunt "vomited out."

And the honest part: both models made the same mistake. At 2.6 both read ὑβρίζει as an imperative ("Abuse yourself, my soul!") where the Greek is indicative ("you are abusing yourself, my soul") — an identical inherited blind spot. That shared miss is why this library never relies on any single model: it is exactly the class of error that a different pair of eyes, a mechanical gate, or a reader writing in can still catch.

The same difference showed up in live production the same day: the Opus 5 referees on our newest volumes caught silently "corrected" regnal numbers, an omitted opening sentence, and reversed passive constructions — failure classes the earlier audits surfaced more rarely. Conclusion: the upgrade is real, it is incremental rather than transformative, and volumes refereed by Opus 4.8 remain sound — each book's own page says which referee checked it, and we do not relabel old work.

Fable 5 fleets — the specialists. Beyond directing, Fable 5 also works in fleets of its own, reserved for the jobs the trials showed only it does reliably:

• Escalation — any batch that fails the mechanical gates twice is automatically re-translated by a Fable 5 agent rather than retried on the fleet model.
• Independence rewrites — when the plagiarism-style scan finds wording shared with an earlier translation, Fable 5 fleets rewrite exactly those sentences fresh from the source until the scan reads zero.
• Re-translation of compromised text — when an audit found one book's early draft echoing a copyrighted modern translation, Fable 5 fleets re-translated all 42 affected letters from the Latin the same day.
• Repairs — every defect an audit surfaces (a truncated paragraph, a reversed meaning) is fixed by a Fable 5 agent working from the source, never by patching the English blind.

The models, plainly: what each one is, and what its scores do and don't tell you.

Model	Its job here	Released	Price / 1M tokens (in · out)
Claude Sonnet 5	the translation fleet — nearly all first-draft text	30 Jun 2026	$2 · $10 → $3 · $15 after 31 Aug 2026
Claude Opus 5	independent referee (Opus 4.8 in this seat before 24 Jul 2026)	24 Jul 2026	$5 · $25
Claude Fable 5	direction, arbitration, hard passages, whole short works	9 Jun 2026	$10 · $50

Why we don't quote their benchmark scores at you. These models are marketed on public evaluations — Frontier-Bench, ARC-AGI, CursorBench, OSWorld, Humanity's Last Exam. Anthropic reports that Claude Opus 5 more than doubles its predecessor's Frontier-Bench score and scores roughly three times the next-best model on ARC-AGI 3, and those numbers are real. But every one of them measures software engineering, agentic tool use, or scientific reasoning. Not one of them measures translation. Anthropic's own announcements for Sonnet 5 and Fable 5 make no claim at all about prose quality. Quoting a coding benchmark as evidence that a Greek genitive was construed correctly would be a category error, and we would rather not commit one on a page about honesty.

So the evidence we offer for the translations is evidence about the translations: a source-paired accuracy audit of hundreds of passages checked against the original languages, mechanical completeness gates on every batch, an independence scan proving no run of eight words is shared with any earlier translation, and our own head-to-head trial in which three models translated the same Greek and only one never silently dropped a line. Those tests are narrow, repeatable, and about the actual job. We publish the numbers either way, including the ones that did not flatter us.

Provenance varies by edition, and we say so. The roles above describe how the library was built when Claude Fable 5 was available to us. Model availability is not constant, and an edition made without a model is not the same artifact as one made with it — so we record which models actually made each book rather than implying a single uniform process. Editions produced while Fable 5 was unavailable were directed by Claude Opus 4.8 in Fable 5's place, with Sonnet fleets translating and Opus refereeing as usual; the mechanical gates (completeness, ratio, independence) are identical either way, since they are code and do not depend on which model is on the pen. Where an edition differs, its own page says so. This matters for anyone comparing editions: the pipeline is constant, the hand on the pen is not.

Which books, and what we intend to do about it. As of now the Opus 4.8-led editions are Lucretius, On the Nature of Things, and Ovid's Metamorphoses (in progress). Their cards on the shelf say so too. When Fable 5 is available to us again we intend to re-translate both from the Latin with it, publish that as the new edition, and keep the Opus 4.8 version alongside it rather than deleting it — so the two can be read against each other, passage by passage, with the same Latin in front of you. We think that comparison is worth publishing in its own right: it is a rare chance to see what actually changes when the model on the pen changes and everything else is held constant. If it turns out very little changes, we will say that too.

How we decided who does what: we tested. The division of labor is measured, not assumed. Before assigning roles we ran the same Greek passages — five genres, then a complete dialogue — through all three models under identical instructions, and graded the outputs against the source:

• Sonnet won on prose naturalness, familiar-name conventions, and instruction discipline — and was the only model that never added a preamble it was told not to add. Its failure mode: rare silent omissions. → bulk translation, guarded by mechanical completeness gates.
• Opus won on hard philology — rare vocabulary, mid-sentence fragments, quotation shape — but followed formatting instructions least reliably. → referee, where sharp eyes matter and formatting doesn't.
• Fable 5 was the only model that never silently dropped a line across the entire test, and scored best on the hardest cruxes (referent tracking, preserved rhetorical structure). → final pen: arbitration, hard passages, repairs, and completeness-critical texts.

The harness: Claude Code. All of this runs on Claude Code, Anthropic's agent harness — the fleets are its subagents and workflows, the gates are ordinary scripts, and the whole pipeline is reproducible by anyone with the same tools. A step-by-step recipe, the source texts, and a runnable replication notebook are on the open data page — we want this copied.

When the tests surprised us, the pipeline changed to match: the referee layer exists because omissions were invisible without it; the mechanical gates exist because they catch for free what the fleet model gets wrong; the escalation path exists because retrying a failed batch on the same model rarely fixed it. This three-tier design is why a small press can ship a library: each model does only what it measured best at, Fable 5's judgment goes exactly where it matters, and the checks are mechanical everywhere else.

Sources: the public-domain Greek texts of the Perseus Digital Library and Open Greek & Latin (Homer, Plato, Plutarch), Niese's Josephus, and Cohn-Wendland's Philo. The audiobooks are narrated by a locally-run neural voice. Models improve and editions may be re-rendered — the charter, the checks, and the voice are what stay constant. The Greek always remains the authority.

← back to the library