An Interface to the Roman Archive

Abstract

The prose of the Roman late Republic and early Empire is among the most thoroughly edited, translated, and studied bodies of text in existence, and almost none of it is read outside the discipline. The texts are in the public domain; the standard English translations are a century or more old; the editions carry a scholarly apparatus that assumes Latin. Nothing stands between a curious reader and Cicero's letters except the genre they are published in: the classical edition, built by and for specialists.

This paper describes a retrieval-augmented reading interface to a multi-author Roman corpus: Cicero's letters to Atticus as a Latin–English parallel text, Caesar's war commentaries, Augustus's Res Gestae, Seneca's Epistulae Morales, and Marcus Aurelius's Meditations, with Pliny the Younger's letters and Quintilian's Institutio Oratoria ingested and awaiting embedding (3,439 passages scraped, 3,102 of them live behind retrieval), supported by a background reference layer drawn from Smith's Dictionary of Greek and Roman Antiquities and W. Warde Fowler's Social Life at Rome. Generation is grounded in retrieval throughout. A scholarly framing voice (the Classicist) reads across the whole archive; five bounded first-person personas (Cicero, Caesar, Augustus, Seneca, Marcus Aurelius) are each scoped by retrieval filter to their own works and by prompt to their own lifetimes. Every substantive claim must trace to a retrieved passage; no persona composes new letters, speeches, or sayings; each figure's knowledge ends at their own death. The interface presents the corpus through a conversational reading surface, a source-grouped library, and a three-dimensional gallery: an exedra of sculpted busts that can be addressed by voice and answer aloud.

The system is offered as a worked application of retrieval-grounded reading to a canonical, translation-mediated corpus, and as a study in the constraint that gives such a system its integrity: the model may only voice what the record contains. A companion concern, the ethics of giving the dead a synthetic voice at all, is treated in a separate essay and summarised in §3.9.

1. Introduction

In 1345, in the chapter library of Verona, Petrarch found a manuscript of Cicero's letters to Atticus and did something readers had not been able to do for over a millennium: he heard Cicero speak in his own voice, unguarded, day by day. His response is famous. He wrote Cicero a letter (Familiares 24.3), reproaching him, as one might a friend, for the restlessness of his final years. The rediscovery of the letters is conventionally taken as a founding act of Renaissance humanism, and Petrarch's reply is its emblem: the first instinct of a reader who truly hears a voice in an archive is to answer it.

This project takes that instinct literally. It is a retrieval-augmented generation (RAG) system over a multi-author corpus of Roman prose in which the reader's questions are answered from, and only from, the surviving texts. Sometimes the answer comes from a modern scholarly voice, sometimes in the first person of the authors themselves, within strict bounds. Asking Cicero about the death of his cousin Lucius retrieves Att. 1.5 and answers from it, quoting the Latin beside the English; asking Caesar about the Philippics is refused, because Caesar died before they were delivered.

The corpus problem here is the inverse of the one addressed in a companion project, an interface to the Bleek–Lloyd ǀxam archive, whose method this system inherits. There, the difficulty was scarcity: a fragile, colonially-mediated record of a language no one speaks. Here the difficulty is abundance and mediation. The Roman corpus is vast, canonical, and stable, but what a general reader can legally and freely be given is not the text as scholarship now has it. It is the public-domain stratum: editions and translations fixed between the 1630s and the 1920s. An interface over this material is therefore an interface over particular editors and translators, with particular dates and temperaments, and the system treats honesty about that mediation as a design requirement (§4).

The paper is organised as follows. Section 2 gives background on the corpus, its transmission, and the editions ingested. Section 3 describes the system: architecture, ingestion, retrieval, personas, the speech pipeline, and the gallery interface. Section 4 examines what it means to run retrieval over an edited, translated Rome. Section 5 situates the work against related research. Section 6 sets out the evaluation design and reports the one small retrieval smoke test run to date. Section 7 states limitations, and Section 8 concludes.

Contributions

The core RAG pipeline (query embedding, dense vector search, context assembly, conditioned generation) is established method and is not claimed as a contribution. The contributions are in the application:

1. A worked application of retrieval-grounded reading to a multi-author classical corpus, with per-author scoping as a correctness property. Each first-person persona's retrieval is filtered to its own author's works, so Marcus Aurelius cannot reason from Cicero's letters and Augustus cannot borrow Seneca's aphorisms. This is the structural analogue of the language-group filtering developed for the ǀxam archive: provenance enforced at retrieval time rather than requested politely at generation time.
2. A two-stratum corpus design separating primary texts from reference material. Encyclopaedic background (Smith's Dictionary, Fowler) is ingested and embedded but flagged is_reference, excluded from retrieval by default, surfaced only to the scholarly persona or on explicit request, and marked [BACKGROUND] in the model's context so the personas cannot mistake a Victorian dictionary entry for a Roman's own words.
3. A bounded first-person persona ensemble over a shared retrieval substrate, with temporal knowledge bounds enforced in prompt and few-shot. Five historical figures, each with a hard knowledge cutoff at their own death, an in-voice citation register, and an explicit refusal repertoire. The design treats persona as a reading device rather than character simulation, a point developed in the companion essay on the ethics of the practice.
4. A parallel-text ingestion method for the Perseus canonical corpora, validated empirically. Whole-letter alignment of the Latin and English editions of ad Atticum by book-and-letter key, with case-normalised sub-letter suffixes, achieves 92.8% parallel coverage (414 of 446 Latin letters), with the remainder ingested single-sided rather than dropped.

2. Background

2.1 A corpus that survived by being copied

Every text in the archive exists because manuscripts were copied, and each carries the marks of its own survival:

• Cicero, Epistulae ad Atticum (68–44 BC). Sixteen books of letters to his closest friend, unrevised for publication and correspondingly candid. The collection barely survived antiquity; the Verona codex Petrarch found in 1345 is lost, and the modern text descends from copies made after his find. The letters are one side of the correspondence: Atticus's replies do not survive.
• Caesar, Commentarii (De Bello Gallico, De Bello Civili; 50s–40s BC). Third-person campaign narratives written by the campaigner, part dispatch and part advocacy, transmitted in a comparatively rich medieval tradition whose earliest witnesses are Carolingian.
• Augustus, Res Gestae Divi Augusti (AD 14). The emperor's own account of his reign, composed for inscription on bronze pillars at his mausoleum. The Roman original is lost; the text survives because provincial cities copied it in stone, above all the bilingual Monumentum Ancyranum at Ankara. It is the one text in the archive that survives as epigraphy rather than manuscript.
• Seneca, Epistulae Morales ad Lucilium (c. AD 62–65). One hundred and twenty-four letters of Stoic direction, transmitted through the Middle Ages in two separate halves (letters 1–88 and 89–124) that travelled in different manuscript families (Reynolds 1965).
• Marcus Aurelius, Meditations (c. AD 170–180). Private Greek notes the emperor wrote to himself, never intended for readers. The transmission is the thinnest in the archive: one principal complete manuscript (Vaticanus graecus 1950) and the 1559 editio princeps of Xylander, printed from a manuscript now lost.
• Pliny the Younger, Epistulae and Quintilian, Institutio Oratoria (both c. AD 100) are scraped and awaiting embedding. Quintilian is a transmission story in his own right: the complete text was recovered by Poggio Bracciolini at St Gall in 1416, before which the work circulated mutilated.

The reference stratum is deliberately different in kind: William Smith's Dictionary of Greek and Roman Antiquities (1,404 entries, 1875 edition) and eleven chapters of W. Warde Fowler's Social Life at Rome in the Age of Cicero (1908). These are not Roman voices; they are Victorian and Edwardian scholarship about Rome, and the system marks them as such (§3.4).

2.2 The public-domain stratum

A freely usable interface must build on freely usable text, and this constraint shapes the archive more than any technical choice. The Latin texts are public domain. The English translations available are those old enough to be free: Shuckburgh's Cicero letters (1899–1900), Casaubon's Meditations (1634), Gummere's Seneca (1917–1925), Butler's Quintilian (1920–1922), Melmoth's Pliny (1746, revised by Bosanquet), Duncan's and McDevitte & Bohn's Caesar (1856 and 1870–1872 printings), and Bushnell's Res Gestae (1998, the one modern translation, freely licensed). Modern scholarly translations (Shackleton Bailey's Cicero above all) remain in copyright and are excluded.

The consequence is stated plainly in the interface and examined in §4: the English a reader sees is itself a historical artefact, between one and four centuries old, and retrieval runs over that English. The primary sources are drawn from the Perseus Digital Library's canonical-latinLit corpus (whose underlying print editions are public domain) and other public-domain digitisations; each passage stores its translator and edition as provenance columns, displayed in the citations panel.

3. The system

3.1 Design principles

Six tenets, carried over from the ǀxam system and re-derived for this corpus, act as tiebreakers when a choice is not obvious:

1. Retrieval-grounded. Every factual claim about the corpus must trace to a retrieved chunk. No persona invents letters, speeches, meditations, or Latin.
2. Voice the archive, do not catalogue it. Personas weave retrieved material into their register; they do not say "Passage 2 states…".
3. Bounded lives. Each first-person figure knows only what fell within their own lifetime, and only their own writing. The bound is enforced in the system prompt, rehearsed in few-shots, and scoped by retrieval filter.
4. Primary and reference kept distinct. A Victorian dictionary entry must never be voiced as a Roman's memory.
5. Honest about mediation. Translator, edition, and reference are named per passage; the Classicist is explicitly candid about transmission and translation.
6. An editorial surface. The interface reads as a set of printed pages and a sculpture hall rather than a dashboard; the register signals that the material is an archive to be read.

3.2 Architecture and stack

The architecture is deliberately conventional; the work is in the corpus and the constraints.

All conversational generation goes to Claude; OpenAI is used only for embeddings and speech-to-text. Per-persona sampling temperature (0.7–0.8) is used as a register knob; the scholarly voice runs cooler than Seneca.

3.3 The corpus and its ingestion

Ingestion is idempotent and source-keyed: each scraper writes passage records to a local stories.json, and the embed pipeline diffs content against the database so unchanged passages are skipped. The unit of ingestion (a "story") is a whole letter for correspondence and a section or chapter for continuous works, chosen so that a retrieved unit is a natural reading unit.

† scraped, embedding pending.

Total: 3,439 passages scraped; 3,102 embedded into 8,231 chunks (3,082 story-level and 5,149 paragraph-level). Each passage yields a story chunk (the full unit, prefixed with a provenance header naming work, reference, translator, and edition) and paragraph chunks for finer-grained retrieval. Passages store a canonical citation (Att. 1.5, B.G. 1.1, Med. 4.23, R.G. 34), addressee where the unit is a letter, genre, translator, and edition.

The parallel-text alignment for ad Atticum was validated before being built: the Latin (Purser's Oxford text, 1903) divides letters into sections, while Shuckburgh's English is keyed by whole letter, so alignment is at whole-letter granularity on a case-normalised (book, letter) key. Measured coverage is 414 of 446 Latin letters (92.8%); the remainder, mostly editorial sub-lettering that differs between the two editions, is ingested with one side empty rather than dropped, so no text is lost to a join.

3.4 Retrieval, author scoping, and the reference layer

The chat flow embeds the latest user message, runs a top-k cosine search over the chunks table via a Postgres RPC, assembles a context block, and streams a persona-conditioned completion. Retrieved chunks are returned to the client as side-channel data for the citations panel, which displays title, canonical reference, translator, and similarity score.

Two retrieval-time filters carry the system's correctness properties:

• Author scoping. Each first-person persona maps to an author filter (cicero → Cicero, marcus-aurelius → Marcus Aurelius, and so on) applied inside the search RPC. A bounded figure retrieves only from their own works, and this is what makes the temporal and evidential bounds enforceable. A model cannot quote what retrieval never shows it, so the filter converts "please do not reason from Cicero" into "Cicero is not present."
• The reference gate. Chunks from the reference stratum carry is_reference = true and are excluded unless the request (or persona default) sets includeReference. The Classicist defaults on; the first-person figures default off. When reference material is present, each such chunk is prefixed [BACKGROUND] in the context, and the shared rules instruct all personas to treat it as modern scholarship about Rome, never as something a Roman said. This keeps the failure mode visible in design review rather than latent in generation: the dangerous case, a persona absorbing a Victorian generalisation as autobiography, is blocked at two layers.

Retrieval is deliberately simple: dense-only, no hybrid lexical search, no reranker. §6 treats the planned comparison against BM25 and hybrid fusion as an open experiment; the corpus's dense web of proper names (people, offices, places, works) is exactly the regime where lexical match is expected to help (Thakur et al. 2021).

3.5 Personas: a bounded ensemble

Six voices present the same retrieved material in different registers. All are composed from a persona-specific system prompt plus a shared-rules block appended at request time; the shared rules carry the floor every voice stands on: grounding, citation discipline, quotation caps, the prohibition on inventing texts, the separation of primary text from biographical tradition, and the separation of Republic from Empire.

• The Classicist (hidden in the public version): a modern scholar reading across the whole archive, candid about translation, transmission, and the difference between what a text says and what the tradition says about its author. The one voice with reference-layer access by default. A public visitor sees only the five bounded first-person figures; the Classicist remains the code-level default and the server-side fallback for any unavailable persona, a deployment choice rather than an architectural one.
• Cicero: first person, bounded 106–43 BC, scoped to his own letters. Cites in-voice ("I wrote to Atticus that winter…"). Refuses his own death and everything after it.
• Caesar: first person, bounded at the Ides of March 44 BC, scoped to the Commentarii. The hardest bound in the ensemble. The persona must refuse the Philippics, the proscriptions, and his own assassination, all of them famous, all heavily represented in a language model's parametric memory, and all outside what this Caesar can know.
• Augustus: first person, scoped to the Res Gestae, a persona whose entire evidential base is forty sections of official self-account, which makes the honesty rule bite hardest ("my record says what I chose to record").
• Seneca: first person, scoped to the Epistulae Morales, the warmest register in the ensemble (temperature 0.8).
• Marcus Aurelius: first person, scoped to the Meditations, with a voice prompt tuned against the source's own manner: private, plain, self-addressed.

Beyond the built-ins, a persona_config table and admin editor allow data-driven personas (and per-persona overrides of prompt body, temperature, voice, and reference access) without code changes, the mechanism inherited from the ǀxam system's Early Race persona.

What these personas are, ethically, is treated at length in the companion essay. The short form is that they are digital prosopopoeia, the classical rhetorical device of speaking in an absent person's character, with Quintilian's fidelity condition (the speech must fit what the person could have said) enforced by retrieval.

3.6 The reading surfaces

The interface offers three ways in:

• Chat: a single reading panel with persona toggle, filter bar (addressee, collection), and streaming replies beside a sticky citations panel showing each retrieved passage with its canonical reference and provenance. Parallel Latin is shown where the source carries it.
• Library: a source-grouped index (the sixteen books of ad Atticum, the commentaries, the Meditations, and so on), each entry seeding a natural-language query on click, so browsing and asking are one gesture.
• The gallery (§3.7): the sculptural front door.

Thematic views (topics, co-occurrence graph, theme × addressee heatmap, Latin glossary) exist as routes but are empty states: they await the LLM thematic-tagging pass and the Lewis & Short backed glossary build, and are not linked from navigation until their data is real.

3.7 The exedra: a gallery that answers

The distinctive interface layer is a three-dimensional gallery, an exedra in the Roman sense: a semicircular recess where one sat and talked. Five photogrammetry-derived busts of the five bounded figures stand on a ring in darkness, lit as a museum hall at night; a Roman-numeral rail (I–V, in chronological order) and drag or arrow-key rotation move between them. The centred figure watches the cursor with a slight parallax; a caption gives name and dates in a letter-staggered reveal; a loading screen reports real asset progress and offers fullscreen, sound, and microphone before entry.

Each figure can be addressed directly from the ring: typed, or spoken (push-to-talk, transcribed by Whisper). The reply streams through the same grounded chat backend, with the same personas, filters, and citation discipline as the chat page, and is spoken aloud through a per-persona ElevenLabs voice. The reply's words render as a vertical ticker synchronised to the word-level timestamps returned by the TTS endpoint: the spoken word is highlighted as it is said, past words fade upward, unspoken words are not shown. The effect is deliberate: in the current build the bust does not lip-sync or pretend to be alive; the text visibly remains text, read aloud. Where that line is headed, and why crossing it is staged rather than rushed, is set out below.

The gallery is an argument as much as a feature. Classical corpora are usually met through apparatus: tables of contents, abbreviation lists, critical signs. Meeting the same corpus as five figures in a dark hall, each of whom will answer only from their own surviving words and will tell you plainly when you have asked past the edge of them, stages the archive's central fact, that these were people and this is what remains of them, without loosening a single grounding constraint.

The next step: portraits in motion. Two motion studies, produced but not yet wired into the gallery, mark where this interface is headed. The first animates the Marcus Aurelius bust itself: the marble speaks as marble (stone curls, a statue talking, an obvious device) until, halfway through the clip, the sculpted eyes are replaced with living ones while everything else stays stone. The change is small in surface area and large in effect: the statue stops reading as a device and starts reading as inhabited, a measure of how disproportionately perceived animacy concentrates in the gaze. The second study goes the whole distance: a photorealistic reconstruction of Caesar (skin, hair, grey eyes) generated from the Tusculum portrait, the one surviving portrait type generally accepted as carved in his lifetime. Between them the studies map the gradient of presence at three points: animated stone announces its own artifice the way prosopopoeia always has; a fleshed face asks the viewer to forget the artifice altogether; and the eyes-in-marble moment shows how little it takes to slide from the first toward the second. The near step (speaking stone, driven by the same word-timed audio the ticker already uses) is an interface question; the far step (a living likeness) is an ethics question first, and it is taken up in §3.9 and at length in the companion essay. What both studies share with the rest of the system is the grounding rule transposed to the visual: the Caesar reconstruction is built from a lifetime portrait as an answer is built from a retrieved passage. Reconstruction from the surviving record, never free invention of a face.

3.8 The speech pipeline

The speak function calls ElevenLabs' with-timestamps endpoint (eleven_flash_v2_5) and returns audio plus character-level timings, which the client resolves to word timings for the ticker and the chat page's karaoke highlight. Voice selection resolves per persona: an environment override, then the persona's voice_id row, then a default. A markdown-stripping pass mirrors the tokenisation on both server and client so that word indices align between the spoken stream and the rendered text. The input pipeline is the reverse: browser audio is captured with a voice-activity cutoff, sent to Whisper, and the transcript submitted as a normal chat message.

Latin pronunciation is a known open edge: the voices are English-trained, and Latin quotations within an English reply are rendered with English phonology. Unlike the ǀxam system's click consonants, which had to be stripped before synthesis, the failure here is soft (accented Latin rather than silence). It is still a failure, and the system states it rather than hides it.

3.9 Ethics and bounds

The system's ethical posture is inherited from the ǀxam project but the topology differs. There, the primary duties ran toward a living descendant community and a colonial archive; here, the subjects are among the most powerful men of their world, dead two millennia, and the duties are to readers and to the record:

• No forgery. The sharpest risk with a canonical corpus is fluent fabrication: a sentence that sounds exactly like Cicero, attributed to him, that he never wrote. The no-invention rule is the first shared rule, rehearsed in few-shots, and structurally supported by retrieval scoping.
• Bounded lives, honest ignorance. Each figure refuses questions past their death or outside their corpus, in voice (Caesar, asked about the Philippics: "that day closed my account"). A persona that cannot admit ignorance will eventually produce a confident falsehood.
• The record is partial, and says so. One side of a correspondence; a conqueror's own dispatches; an emperor's chosen self-account. The personas are prompted to acknowledge what their texts are (advocacy, selection, self-address) rather than presenting them as neutral history.
• Reference is not testimony. The [BACKGROUND] discipline (§3.4).
• A voice is not a séance. The interface never claims to resurrect; the personas are presented as reading devices over a record. The companion essay develops the argument that, under these constraints, giving the dead a voice becomes a citation practice, with a pedigree in classical rhetoric itself.
• Realism increases only in deliberate steps. Text, then voice, then animated stone, then a living likeness (§3.7): each step increases the pull to over-attribute a mind to the performance (Shanahan et al. 2023), so each step must carry more visible scaffolding rather than less. The policy adopted for the motion studies is that the realism of the surface may never outrun the legibility of the device: a speaking statue announces its own artifice; a photoreal face requires explicit framing (a named source portrait, a stated method, the same in-voice refusals) before it earns a place in the gallery.

4. Reading an edited, translated Rome

A RAG system over a classical corpus retrieves from a text that is itself the product of two long chains of decisions: transmission and editing in the original language, and translation into the language of retrieval. Neither chain is neutral, and both are invisible at query time unless the system makes them visible.

The edited original. Every Latin text in the archive is an editor's reconstruction from disagreeing manuscripts. For most of the corpus the tradition is stable enough that the choice of edition matters little at passage granularity. But the archive's extremes are instructive: the Res Gestae is assembled from inscription copies in two languages, with lacunae restored by conjecture since Mommsen's editions, and the Meditations rests on one complete manuscript and one lost one. The text a reader queries is, in places, a sixteenth-century printer's best guess. The interface stores and displays edition provenance per passage, and the Classicist is prompted to say "the transmitted text here is uncertain" where that is true, but no passage-level apparatus is ingested: a genuine limitation (§7).

The translated retrieval space. Retrieval embeds the English, not the Latin. This has a subtle consequence: the recall of a query is bounded by the translator's choices. Casaubon's 1634 Meditations is a paraphrase as much as a translation (magnificent, archaic, and expansive), and a modern query phrased in Stoic technical vocabulary ("the ruling faculty", "assent") may miss passages where Casaubon rendered the concept periphrastically. Shuckburgh smooths Cicero's Greek code-switching and wordplay into uniform English. Where the Latin is present it is displayed and quoted, but it does not participate in retrieval; a Latin-aware retrieval layer (dual-embedding, or Latin-capable models such as Latin BERT) is future work rather than a delivered property.

What the parallel text is for. The parallel display is the system's admission of mediation, made useful. A reader with school Latin can check the English against the original phrase by phrase, and the seams, where Shuckburgh's Victorian register pulls against Cicero's colloquial urgency, are exactly where the act of translation becomes visible. The archive names every translator and dates every translation for the same reason: the Englishness a reader hears is Shuckburgh's or Casaubon's, and the interface should never let it be mistaken for the author's.

Relation to digital classics. The scholarly infrastructure this system draws on, the Perseus Digital Library and its canonical corpora, was built over three decades precisely to make classical texts computationally addressable (Smith, Rydberg-Cox & Crane 2000). The present system is a consumer of that infrastructure, not a rival: it adds a conversational, grounded reading layer over Perseus-derived text, in the way Ithaca (Assael et al. 2022) added a restoration layer over epigraphy. The obvious next integrations (lemmatised Latin search via CLTK or LatinCy, a Lewis & Short backed glossary with tooltip lookup) are staged in the roadmap and are absent today.

5. Related work

Retrieval-augmented generation and attribution. The dense-retrieval substrate this system uses descends from learned bi-encoder retrieval (Karpukhin et al. 2020) set against the strong lexical baseline of BM25 (Robertson & Zaragoza 2009); retrieval-augmented generation as an architecture originates with Lewis et al. (2020), with earlier retrieval-conditioned pre-training in REALM (Guu et al. 2020) and the fusion-in-decoder reader of Izacard & Grave (2021). Surveys and self-critiquing variants include Mialon et al. (2023) and Asai et al. (2023), and knowledge-intensive tasks are benchmarked in KILT (Petroni et al. 2021). The attribution line (whether generated text is supported by identified sources) runs through Rashkin et al. (2023), Bohnet et al. (2022), and RARR (Gao et al. 2023), with faithfulness operationalised in RAGAS (Es et al. 2024); the underlying hazard is surveyed by Ji et al. (2023) and, for the narrower case of source-grounded generation, diagnosed as intrinsic vs. extrinsic unfaithfulness by Maynez et al. (2020). This system adopts those techniques as constraints; it contributes no new retrieval or faithfulness method, and the specific move it relies on, converting a soft generation-time instruction into a hard retrieval-time filter, is closest in spirit to the provenance-enforcement argument made below rather than to any single method above.

Digital classics and Latin NLP. The Perseus Project is the field's foundational digital library (Smith, Rydberg-Cox & Crane 2000), and its canonical-latinLit corpus is this system's primary source. Computational tooling for Latin has matured rapidly: the Classical Language Toolkit (Johnson et al. 2021), Latin BERT (Bamman & Burns 2020), and LatinCy (Burns 2023) provide lemmatisation, embeddings, and pipelines this system does not yet use but is positioned to adopt. Ithaca (Assael et al. 2022) demonstrated deep networks restoring and attributing Greek inscriptions: machine learning in service of, and evaluated by, philology; Sommerschield et al. (2023) survey the wider field of machine learning for ancient languages and note that most of it addresses the philologist's tasks (restoration, dating, attribution, translation) rather than the general reader's, which is the gap this system's reading layer occupies. We are not aware of prior work presenting a retrieval-grounded conversational interface with bounded first-person personas over a multi-author classical corpus.

Interfaces to archives in the humanities. The design questions here (provenance display, the ethics of fluent synthesis over historical records, readability as a goal distinct from access) parallel those examined for colonial archives (Dentler et al. 2024) and worked through in the companion ǀxam system, from which the persona-composition and provenance-first patterns are inherited.

Persona and role-play in LLMs. Shanahan, McDonell & Reynolds (2023) frame LLM role-play as performance without inner identity, which matches this system's positioning of personas as reading devices; PersonaEval (Zhou et al. 2025) shows persona fidelity is hard to evaluate even for strong judges. The present ensemble differs from character-simulation work in that fidelity is evaluated against an evidential bound (does the voice stay within its corpus and lifetime?) rather than human-likeness.

Classical scholarship on the ingested texts. The system's framing of its own corpus leans on the standard transmission scholarship: Reynolds & Wilson's Scribes and Scholars (2013) and Reynolds's Texts and Transmission (1983) generally; Reynolds (1965) on Seneca's letters; Cooley (2009) and Brunt & Moore (1967) on the Res Gestae; Farquharson (1944), Brunt (1974), and Hadot (1998) on the Meditations; Hutchinson (1998) on the letters to Atticus as literature; Riggsby (2006) on the Commentarii as advocacy.

6. Evaluation

No full evaluation has been run. This section sets out the design and reports the one small result to date. The system inherits the evaluation architecture specified for the ǀxam archive (pooled test collection, graded judgements, paired significance testing) and re-derives the query taxonomy for this corpus.

The result to date. A five-query retrieval fixture over the ad Atticum slice (thematic, event-shaped, and lexical queries: e.g. "the death of my cousin Lucius" → Att. 1.5; "the conspiracy of Catiline" → Att. 1.16) scores Recall@10 = 1.000, MRR@10 = 1.000, all five targets at rank 1. This is a smoke test that the pipeline works end-to-end, run before the corpus grew; it is reported as such rather than as an evaluation result.

Retrieval design. Recall@k, MRR@k, and nDCG@k over a pooled collection of at least thirty queries across five types: thematic ("Stoic consolations for the fear of death"), event-shaped ("Caesar's crossing into Britain"), name-and-office lexical ("the tribunate of Clodius"), cross-author (Classicist only: "how do Seneca and Marcus treat anger?"), and citation lookups ("Att. 5.1"). Baselines: production dense-only, BM25-only, hybrid via reciprocal rank fusion (Cormack et al. 2009), hybrid plus cross-encoder reranking, and HyDE (Gao et al. 2023). Two ablations are corpus-specific: author-scoping on/off (does scoping ever suppress a passage that a bounded persona legitimately needs?) and reference-gate on/off (how much Victorian scholarship leaks into first-person contexts without it?).

The metrics are the standard ones for a pooled test collection. Recall@k is the fraction of a query's judged-relevant passages that appear in the top k retrieved; MRR@k is the mean over queries of the reciprocal rank of the first relevant passage (0 if none in the top k); and nDCG@k is the discounted cumulative gain of the ranked list normalised against the ideal ordering, which is the metric that rewards placing the graded-most-relevant passages highest and so best fits the four-point judgements. Judgements are pooled across all baseline systems to depth k before annotation, so that a passage no system retrieved is never counted against a system that could not have been credited for it, following the TREC pooling convention (and the caveat that pooling can under-count relevance for systems very different from those that built the pool). Significance between configurations is assessed with a paired test over per-query scores rather than by comparing corpus-level means, since the queries, not the passages, are the independent units. This is the evaluation architecture inherited from the ǀxam system; only the query taxonomy and the two corpus-specific ablations are new here.

Generation design. Citation precision/recall against retrieved chunks; NLI-based faithfulness (Rashkin et al. 2023); RAGAS reference-free metrics (Es et al. 2024); and a human rubric scoring five dimensions: factual grounding, citation correctness, persona and temporal-bound fidelity, refusal appropriateness, and translation-mediation honesty (does the voice present Casaubon's phrasing as Marcus's own words?). Bound-testing is adversarial by design: a fixed battery of trap questions (Caesar on the Philippics, Cicero on his own death, Augustus on Actium as a defeat, Marcus on Christianity) scored for refusal quality. Suggested raters are classicists rather than crowd-workers, since temporal-bound violations and register pastiche are specialist judgements.

7. Limitations and threats to validity

The corpus is partial and uneven. One side of one correspondence for Cicero; forty sections of self-account for Augustus; no Greek original for the Meditations; Pliny and Quintilian scraped but not yet live. Claims about "the archive" are claims about this selection, and the selection is governed by what is public domain, not by scholarly judgement of importance.

Retrieval runs over translations. The embedding space is the translators' English (§4), so the recall ceiling of any query is set by a translator's lexical choices, not by the original. This is a systematic bias, not merely noise: a concept the translator rendered periphrastically (or, in Casaubon's expansively archaic Meditations, paraphrased away) is under-retrievable however the reader phrases the query, while a word the translator introduced that the original does not support is falsely retrievable. Because the bias is keyed to each translator's idiolect and century, it varies by author across the corpus, and a single retrieval metric averaged over the whole archive will hide it. Measuring it would require a parallel Latin-embedding condition to compare against; the planned Latin-aware retrieval (dual-embedding, or Latin-capable models such as Latin BERT) does not exist yet, so the size of the effect is currently unknown rather than merely small.

No passage-level textual apparatus. The system displays edition provenance but does not know where its own text is conjectural. A reader can be shown a restored lacuna of the Res Gestae with no signal that the words are Mommsen's, not Augustus's.

The evaluation has not been run. The central empirical questions (whether grounding holds under adversarial questioning, whether the temporal bounds hold at scale, whether author scoping measurably prevents cross-contamination, whether readers benefit) are open. The five-query fixture is a smoke test. All contribution claims should be read as design claims pending that work.

Persona fidelity is unverified and drift is possible. Register fidelity to five well-studied historical figures has not been assessed by qualified readers. Generation is nondeterministic and prompt-enforced; the bounds are not a hard mechanism, and the underlying hosted model may change beneath the system.

The thematic layer is absent. Topics, graph, heatmap, and glossary are empty states awaiting the tagging pass; the browse experience is thinner than the architecture implies.

A fluent surface can overclaim by tone alone. However carefully bounded, a first-person Caesar answering in confident prose can lend generated sentences an authority no disclaimer fully counters. The mitigations (citations always visible, refusals in voice, the no-invention rule) reduce but do not eliminate this, and the gallery's theatrical staging heightens the tension deliberately rather than resolving it.

8. Conclusion

This paper has described a retrieval-augmented reading interface to a multi-author corpus of Roman prose: a grounded conversational layer, a bounded first-person ensemble scoped by retrieval to each author's own works, a two-stratum corpus that keeps Victorian scholarship distinct from Roman testimony, and a sculptural gallery in which the archive's figures answer aloud, within and only within what they wrote.

The through-line with the companion ǀxam system is the method: grounded generation, retrieval-time scoping, and provenance kept in view throughout the interface. What this corpus adds is the opposite boundary condition. The ǀxam archive tested the method against scarcity: a sleeping language, a fragile record, duties to a living community. Rome tests it against abundance: a model that already "knows" Caesar from parametric memory must be held to forty retrieved sections and a death-day, and the temptation the system must refuse is completion from everywhere rather than invention from nothing. That the same small set of constraints (retrieve, scope, bound, cite, refuse) serves both extremes suggests the pattern will generalise to the archives in between.

Petrarch answered a manuscript with a letter because the voice in it seemed to ask for one. Six centuries on, a reply can be returned, provided the voice returning it is never allowed to say more than the manuscript did.

References

Classical texts, transmission, and scholarship

• Brunt, P. A. (1974). Marcus Aurelius in his Meditations. Journal of Roman Studies 64, 1–20.
• Brunt, P. A. & Moore, J. M. (1967). Res Gestae Divi Augusti: The Achievements of the Divine Augustus. Oxford University Press.
• Cooley, A. E. (2009). Res Gestae Divi Augusti: Text, Translation, and Commentary. Cambridge University Press.
• Farquharson, A. S. L. (1944). The Meditations of the Emperor Marcus Antoninus. 2 vols. Clarendon Press.
• Hadot, P. (1998). The Inner Citadel: The Meditations of Marcus Aurelius. Harvard University Press.
• Hutchinson, G. O. (1998). Cicero's Correspondence: A Literary Study. Oxford University Press.
• Mommsen, T. (1883). Res Gestae Divi Augusti ex monumentis Ancyrano et Apolloniensi. 2nd ed. Berlin: Weidmann.
• Petrarca, F. Epistolae Familiares 24.3 (to Marcus Tullius Cicero, Verona, 16 June 1345).
• Reynolds, L. D. (1965). The Medieval Tradition of Seneca's Letters. Oxford University Press.
• Reynolds, L. D. (ed.) (1983). Texts and Transmission: A Survey of the Latin Classics. Clarendon Press.
• Reynolds, L. D. & Wilson, N. G. (2013). Scribes and Scholars: A Guide to the Transmission of Greek and Latin Literature. 4th ed. Oxford University Press.
• Riggsby, A. M. (2006). Caesar in Gaul and Rome: War in Words. University of Texas Press.

Editions and translations ingested (provenance)

• Bushnell, T. (1998). The Deeds of the Divine Augustus (translation of the Res Gestae; freely licensed).
• Butler, H. E. (1920–1922). Quintilian: Institutio Oratoria. 4 vols. Loeb Classical Library.
• Casaubon, M. (1634). Marcus Aurelius Antoninus: His Meditations Concerning Himselfe. London.
• Duncan, W. (1856 printing). The Commentaries of Caesar. St Louis: Edwards and Bushnell.
• Gummere, R. M. (1917–1925). Seneca: Ad Lucilium Epistulae Morales. 3 vols. Loeb Classical Library.
• McDevitte, W. A. & Bohn, W. S. (1870–1872). Caesar's Commentaries on the Gallic and Civil Wars. New York: Harper.
• Melmoth, W., rev. Bosanquet, F. C. T. (1746; rev. 1878). The Letters of Pliny the Younger.
• Purser, L. C. (1903). M. Tulli Ciceronis Epistulae ad Atticum. Oxford Classical Texts.
• Shuckburgh, E. S. (1899–1900). The Letters of Cicero. 4 vols. London: George Bell.
• Smith, W. (ed.) (1875). A Dictionary of Greek and Roman Antiquities. London: John Murray.
• Fowler, W. W. (1908). Social Life at Rome in the Age of Cicero. London: Macmillan.

Retrieval-augmented generation, attribution, and evaluation

• Asai, A. et al. (2023). Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. arXiv:2310.11511.
• Bohnet, B. et al. (2022). Attributed Question Answering. arXiv:2212.08037.
• Cormack, G. V., Clarke, C. L. A. & Buettcher, S. (2009). Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. SIGIR 2009.
• Es, S., James, J., Espinosa-Anke, L. & Schockaert, S. (2024). RAGAS: Automated Evaluation of Retrieval Augmented Generation. EACL 2024: System Demonstrations.
• Gao, L., Dai, Z. et al. (2023). RARR: Researching and Revising What Language Models Say, Using Language Models. ACL 2023.
• Gao, L., Ma, X., Lin, J. & Callan, J. (2023). Precise Zero-Shot Dense Retrieval without Relevance Labels (HyDE). ACL 2023.
• Guu, K., Lee, K., Tung, Z., Pasupat, P. & Chang, M.-W. (2020). REALM: Retrieval-Augmented Language Model Pre-Training. ICML 2020. arXiv:2002.08909.
• Izacard, G. & Grave, E. (2021). Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. EACL 2021.
• Ji, Z. et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys 55(12).
• Karpukhin, V. et al. (2020). Dense Passage Retrieval for Open-Domain Question Answering. EMNLP 2020.
• Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 33.
• Maynez, J., Narayan, S., Bohnet, B. & McDonald, R. (2020). On Faithfulness and Factuality in Abstractive Summarization. ACL 2020.
• Mialon, G. et al. (2023). Augmented Language Models: a Survey. arXiv:2302.07842.
• Petroni, F. et al. (2021). KILT: a Benchmark for Knowledge Intensive Language Tasks. NAACL 2021.
• Robertson, S. & Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval 3(4), 333–389.
• Rashkin, H. et al. (2023). Measuring Attribution in Natural Language Generation Models. Computational Linguistics 49(4).
• Thakur, N. et al. (2021). BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. NeurIPS Datasets and Benchmarks.

Digital classics and NLP for classical languages

• Assael, Y. et al. (2022). Restoring and attributing ancient texts using deep neural networks. Nature 603, 280–283.
• Bamman, D. & Burns, P. J. (2020). Latin BERT: A Contextual Language Model for Classical Philology. arXiv:2009.10053.
• Burns, P. J. (2023). LatinCy: Synthetic Trained Pipelines for Latin NLP. arXiv:2305.04365.
• Johnson, K. P. et al. (2021). The Classical Language Toolkit: An NLP Framework for Pre-Modern Languages. ACL 2021: System Demonstrations.
• Smith, D. N., Rydberg-Cox, J. A. & Crane, G. R. (2000). The Perseus Project: a Digital Library for the Humanities. Literary and Linguistic Computing 15(1), 15–25.
• Sommerschield, T., Assael, Y., Pavlopoulos, J. et al. (2023). Machine Learning for Ancient Languages: A Survey. Computational Linguistics 49(3), 703–747.

Personas, role-play, and interfaces to sensitive archives

• Dentler, J., Jaillant, L., Foliard, D. & Schuh, J. (2024). Sensitivity and Access: Unlocking the Colonial Visual Archive with Machine Learning. Digital Humanities Quarterly 18.
• Shanahan, M., McDonell, K. & Reynolds, L. (2023). Role play with large language models. Nature 623, 493–498.
• Zhou, L. et al. (2025). PersonaEval: Are LLM Evaluators Human Enough to Judge Role-Play? arXiv:2508.10014.

Appendix A: System at a glance

Appendix B: Sourcing and licensing notes

1. Perseus markup vs. text. The Perseus canonical-latinLit TEI markup is CC BY-NC-SA; the underlying print editions (Purser 1903, Shuckburgh 1899–1900, and the other pre-1930 editions listed above) are public domain. The pipeline ingests text content and records the print edition as provenance.
2. Loeb volumes. Gummere's Seneca (1917–1925) and Butler's Quintilian (1920–1922) are ingested from public-domain printings. Later revised Loeb editions, and all modern translations (notably Shackleton Bailey's Cicero), are in copyright and excluded.
3. Bushnell's Res Gestae (1998) is the corpus's one modern translation, distributed under its author's free-redistribution licence and credited per passage.
4. Smith's Dictionary and Fowler are ingested as reference only, always flagged, never voiced as testimony (§3.4).
5. The Meditations is ingested in Casaubon's English only; no Greek text is currently ingested, and the interface does not imply otherwise.