Unearthing What Might Have Been

Epictetus wrote nothing. His lectures at Nicopolis survive because his student Arrian recorded them, and Arrian later condensed their essentials into a short manual, the Enchiridion, or Handbook. Four books of the longer Discourses remain. The rest are lost.

Because the Handbook is a distillation, it preserves traces of lectures that no longer exist. Most of its chapters can be matched to passages in the surviving Discourses. Some cannot. Chapter 27 is the best-known example:

Gregory Sadler, a philosopher who teaches and writes widely about Stoicism, notes in an essay for Modern Stoicism that the Discourses passage behind this sentence does not survive. He draws on the work of Gerard Boter, the classicist whose critical edition of the Handbook is the standard modern text. The chapter is a single compressed sentence, and its meaning is still debated.

This project began with a simple question. Starting from a Handbook chapter, could a language model give us a glimpse of what the fuller Discourses passage might have been like? We did not want a restatement or a generic Stoic reflection. We wanted a passage that develops the chapter's argument the way Epictetus did in the classroom: with direct address, a student's objection, a correction, and images that do real work.

A glimpse is an informed, bounded impression of what a lecture could have contained, and one plausible reading among several. It is not a replacement for what was lost, and no model output restores the lost books. Every generated passage in this post is modern writing and is labeled as such. Our aim was to make these passages as careful as possible, and to be honest about how far they fall short.

1How we got here

We did not begin with prompting. No dataset of Handbook-to-Discourses pairs exists, so we built one and tried to teach a model the task directly, mostly with small local models. Models, data and rewards changed between attempts, so this is a history, not a controlled comparison. The figure uses standard shorthand: SFT is supervised fine-tuning on example pairs, GRPO is a kind of reinforcement learning (RL), LoRA is a lightweight way to fine-tune on modest hardware, and 1B to 8B gives model size in billions of parameters.

Every attempt ran into the same problem. At various points, similarity scores, rewards and loss all moved in the right direction, but none of them told us whether the argument had survived. Chapter 27 exposed that gap more than once. A passage that gets the argument wrong is worse than none, because it puts the wrong teaching in a familiar voice. None of this shows that these techniques fail in general. It shows only that they did not deliver reliable results at an acceptable cost here. Our own summary of this period is:

We found that better metrics, more optimization and more inference complexity were unreliable substitutes for inspecting the actual philosophical argument.

So we changed direction. Every attempt so far had asked a model to infer Epictetus from examples and rewards. What if we described him instead: who he was, what he taught, where his doctrines stop, and how he taught?

2Defining the teacher

Every request in the study went to one model, GPT-5.6-luna, with no fine-tuning. Only the guidance around it changed, arranged as a ladder of seven levels:

Levels 1 and 2 are baselines: a two-line instruction (zero-shot), then the same instruction with a real Discourses sample to imitate (one-shot). Level 3 swaps the sample for a written account of the teacher. It sets out the distinctions a faithful passage must keep, the methods he used (direct address, an objection answered, analogies whose details do work), and what his voice is not:

His forcefulness should come from precise reasoning, not ornamental language, generic encouragement, or repeated declarations of certainty.

Above all, it puts the chapter first:

The supplied chapter governs the interpretation. Preserve its particular argument, examples, and qualifications. Do not replace it with a familiar Stoic lesson merely because that lesson seems related.

Level 4 adds the sample back, now paired with the Handbook chapter it relates to (chapter 47) as a demonstration. Level 5 splits the work in two. The model first writes short notes on how it reads the chapter, an interpretation worksheet, and then writes the passage from those notes. Levels 6 and 7 run two or five Level-5 pipelines and let a selector choose among them, using medium reasoning effort at Level 6 and high at Level 7.

3Testing against what survives

An imagined lecture cannot be checked against the real one. So we tested first on chapters whose sources do survive, where we can ask a concrete question: does the expansion develop the argument the real lecture developed?

The test set had seventeen chapters: four real ones (3, 8, 20 and 42) paired with their surviving Discourses passages, one control whose Handbook text nearly copies its source, and twelve synthetic Handbook-style chapters. The writing model never saw the surviving passages. We kept 100 of the results and had each one reviewed twice: once by the same model acting as a judge (labeled Luna in the figures), and once by a second model, the direct reader, that read every passage closely beside its source, without knowing which level wrote it. Every score had to quote the passage it judged. Each reviewed passage has an ID from R001 to R100. Neither reviewer is a human scholar, a limit we return to below.

4What we found

Describing the teacher helped

On the real chapters, the direct reader scored the written account (Level 3) at 4.72 out of 5, against 4.28 for the bare instruction and 3.94 for the style sample. Adding the sample was the weakest setup in the study. Showing the model what Epictetus sounds like strengthened the rhetoric and sometimes pushed the argument aside. Telling it what he taught, and where those teachings stop, kept the passages closer to the chapter.

Interpreting before composing helped most

On the 13 chapters both levels completed, Level 5 beat Level 4 on 10, tied on 3 and lost none. Having the model settle what the chapter says before it starts writing was the most consistent improvement we found.

More candidates did not

Level 7 wrote five candidates and had a careful selector pick one. It beat Level 5 on 4 of 13 shared chapters, tied on 4 and lost on 5, and on the real chapters it never came out ahead. It also used about eight times as many tokens and failed to finish more of its runs. Level 5 is our default, and Level 3 is the single-call alternative.

5Where automated judgment fell short

A generated passage is only worth having if it does not misrepresent the teacher. That makes evaluation central, and it is where we found the most important weakness. On average the automated judge and the direct reader nearly agreed. On a few passages they diverged sharply, and those gaps show how fluent prose can hide a broken argument:

Each of these passages needs only one repair. But in Epictetus' voice, a single reversed pronoun can invert his teaching while still sounding like him, and the automated judge did not reliably catch it. A short adversarial check, aimed at negations, pronouns and whether conclusions follow, would suit the problem better than another round of blanket scoring.

6Beside the lectures that survive

The real chapters let us set each generated passage beside the lecture behind its chapter. Handbook 3, on the jug and the kiss, shows both what the method can do and where it stops. The strongest passages handled a delicate analogy well. A broken jug and a dead child can be compared only in a narrow sense, and one two-candidate passage said so directly:

The surviving lecture behind this chapter, Discourses 3.24.84–88, is richer than any of the expansions. It compares a loved one to a jar or a crystal goblet. It recalls the attendant who stood behind a general in his triumph, reminding him that he was mortal. It warns that longing for someone no longer given to you is like hankering for a fig in winter. None of the six reviewed passages on this chapter used any of these images. They kept the argument, some with real care, but the particular life of the lecture could not be recovered from the Handbook's two sentences. That is the clearest evidence we have of the gap between a generated passage and the text it gestures toward.

7Where nothing survives

Chapter 27 was the reason for the project, and it is where the difference between an impression and a replacement matters most. With no surviving lecture, no score can say how close a passage came. So we gave the chapter's single sentence to four levels, once each, with no commentary on the chapter and no scoring. The passages sit outside the 100 reviewed above.

Levels 4 and 5, working separately, arrived at almost the same sentence: "You have confused what happens with what a thing is for." Both prompts carry the same account of Epictetus, which may be what drew them together. With one trial each, we cannot say more. The simplest prompt with an example, Level 2, produced the sharpest image in the set:

Three of the four passages also bring analogies of their own: a surgeon's knife, an eye made for seeing, a physician who fails to heal. Level 5 keeps to the mark alone. Those images are the model's additions. Nothing connects them to a lecture Epictetus actually gave, and only the mark and the missed aim come from the Handbook.

The passages also lean toward different readings. Levels 4 and 5 stress what nature is ordered toward. Level 2 treats evil as a failure to reach the good. Scholars disagree too: Albert Watanabe, in his commentary on the Handbook for Dickinson College Commentaries, sets a providential reading beside Boter's relational one. We did not try to settle that question, and we do not think a model should. Showing several readings side by side is more honest than presenting one as what Epictetus said.

The model's notes end on what the chapter leaves open: what nature means here. Every passage written from chapter 27 should carry that uncertainty with it.

8What a generated passage is, and what it is not

The Discourses and the Handbook have been read for nineteen centuries, often by people facing real hardship. Their authority rests on what they actually say, and a passage mistaken for the original borrows that authority without earning it. A model can do damage here in two ways.

The first is substitution. A model trained on the open web has absorbed a great deal of popular Stoicism, and asked for Epictetus it will readily supply a familiar maxim in his voice. Several reviewed passages swapped their chapter's particular argument for a nearby, more familiar lesson.

The second is quietly adding our own values. Many outputs carried modern caveats about grief and safety. Sometimes these protect a distinction the chapter needs. They can also, in the direct reader's words, "smooth away some of the severity and dramatic texture of the supplied ancient material." Epictetus is sometimes severe. A version that makes him gentler for a modern reader shows us ourselves, not him.

That is why the account of the teacher was written to mark the limits of his doctrine as well as its content, and why the worksheet asks the model to state what it is inferring rather than assert it. Any presentation of this work should keep ancient and generated text visibly distinct, never call a generated passage recovered or restored, keep uncertainty notes attached, and show the surviving source first wherever there is one.

Used this way, a generated passage can be a useful companion to study. It can show how a compressed chapter might unfold, and send a reader back to the surviving Discourses with sharper questions. It remains modern writing, and the lost books remain lost.

9Limitations

• The test set is small and public. Four real chapters, all likely seen by the model in training, make this a development comparison, not an unbiased test.
• Surviving chapters are not lost ones. The model may partly draw on lectures it has already read. A truly missing lecture offers no such support.
• Neither reviewer is a human scholar. Both are language models, and each level was reviewed on a single run.
• The comparisons are between whole setups, not isolated causes. Each rung of the ladder changed more than one thing.

10Recommendations

1. Establish a prompting baseline before training. It also defines what any training run must beat.
2. Describe the author and the limits of his teaching. It beat both a bare instruction and a style sample.
3. Interpret before composing. It produced the most consistent gains.
4. Treat extra candidates with caution. It cost far more and showed no clear gain on real chapters.
5. Read closely. Check negations, pronouns and whether conclusions follow; automated scores miss them.
6. Present generated passages for what they are. Label them, keep their uncertainty attached, and offer more than one where readings diverge.

11Looking forward

Next we need broader inputs, especially more chapters with distinctive images whose sources survive, and review by human experts. If we return to training, it should start from carefully reviewed interpret-then-compose examples.

The method that worked best asks the model to stop and examine what a chapter says before it writes, much as Epictetus asked his students to say, "Wait for me a little, O impression," before giving assent. Readers of any generated passage should apply the same discipline. These passages offer a glimpse, bounded and uncertain, of lectures we will never hear. The Discourses that survive remain the only Epictetus we have, and they deserve to be read first.