Formulaic Language: The Premise of Prediction — bundled resource 公式化语言：预测的前提 — 总集资源

Epigraphy is, to an unusual degree, a predictive discipline. The stones reach us broken, abraded, re-used; the editor's task is to complete them. Completion is possible at all only because Roman inscriptions are formulaic — they repeat. A funerary altar, a votive plaque, a building dedication, an honorific base: each draws on a small, well-worn repertoire of phrases. Where the stone fails, the formula carries on.

The standard reference work states the method without hesitation:

Appendix II of the Oxford Handbook is a list of common abbreviations used in Latin inscriptions. It is indispensable — and it is candid about its own limits. Its opening paragraph is, in effect, the first challenge to predictive restoration, written by the predictors themselves:

To make this measurable we set the Handbook's menu beside the corpus. In EDCS the convention is the standard epigraphic one: the letters actually on the stone are printed plain, and the editor's completion follows in round brackets — D(ecimi), M(anibus), co(n)s(ul). Every such token in 537,262 inscriptions can therefore be harvested mechanically, and the engraved abbreviation tallied against each word the editors expanded it into.

The table below pairs, for each single-letter abbreviation, three figures: the size of its OHRE Appendix II menu (the curated teaching list); its total EDCS attestations; and the Shannon entropy of its expansion distribution, in bits. Entropy measures the ambiguity in that attested distribution: it is low when one expansion dominates (the abbreviation is near-deterministic) and high when probability is spread across many (a guess). Sort by any column; the entropy column is the spine of Challenge I.

The Handbook's method has a binary shape. Identify the formula; restore from it. On this picture an abbreviation either belongs to a recognised formula — and is therefore restorable — or it does not. But abbreviations do not divide so cleanly. Predictability is not a switch. It is a dial, and the dial has a number.

That number is entropy. For an abbreviation with attested expansions of probability p₁, p₂, …, the Shannon entropy H = −Σ pᵢ log₂ pᵢ measures, in bits, how much genuine uncertainty remains once you know only the letter. One bit is one yes/no question. Sorted by entropy, the single-letter abbreviations of Appendix II do not fall into two camps. They spread continuously across the whole range:

At the safe end sits E. In 21,148 attestations the bare letter E expands to est in 87% of cases — overwhelmingly the est of a funerary formula (h(ic) s(itus) e(st)). Its entropy, 1.26 bits, is barely more than a single yes/no question. Restoring a final [E] in h(ic) s(itus) [e(st)] is not a bet; it is a near-certainty. The predictive thesis is, here, simply correct.

At the other end sits P, entropy 5.32 bits. The corpus attests the single letter P expanding to pedes, Publius, pater, posuit, pius, pecunia, populus, plebs, pondo, pars, praetor, proconsul, provincia, perpetuus, parentes — and on. Appendix II's menu for P runs to nineteen words; the corpus runs to hundreds. A bare [P] in a lacuna is not restored from a formula; it is filled by whatever the surviving neighbours happen to license. Strip those away and "restoration" becomes indistinguishable from naming the most frequent option.

The lesson is not that restoration is unreliable. It is that reliability is a continuous quantity that the editorial notation does not record. The Leiden Convention gives us the square bracket: […] means "these letters were restored". It cannot say whether the restoration was a 1.26-bit near-certainty or a 5.32-bit guess. On the printed page [E] and [P…] look identical. The bracket records that a prediction was made; it hides the entropy of the prediction. Challenge I asks the editor to carry, in the mind if not on the page, the number the bracket omits.

Challenge I was statistical. Challenge II is textual, and sharper, because the witness against predictive restoration is the Oxford Handbook itself. On page 14 the Handbook offers its flagship example of a safe restoration — the funerary D M:

Now turn 773 pages, to the same book's Appendix II. Under D M the Handbook lists seven expansions; under D M S, two:

The contradiction is exact. Page 14 calls [D] M S a "safe bet" for a funerary epitaph; page 787 records that the very same three letters are also the standard opening of a Mithraic dedication — Deo Mithrae sacrum. A goddess (Dea Magna), a municipal councillor (decurio municipii), a legal clause warding off fraud (dolus malus), the cult of Mithras: the letters D M open all of them. The "safe bet" is safe only if you already know it is funerary — and what tells you that is not the letters but the context: the cippus, the cheap stone, the layout of an epitaph. Predictive restoration quietly imports that context as a free premise. A lacuna does not grant it for free. (This is the hinge into Challenge IV.)

Below are real inscriptions in which D M is not Dis Manibus. Decide each before revealing the answer.

Challenge II showed that restoring an abbreviation means choosing a genre. Challenge III asks where the editor gets the odds for that choice — and the answer is uncomfortable.

Database-based restoration has an obvious, principled rule: restore the expansion the corpus attests most often. It is the rule the frequency tables of this whole module are built to serve. But the rule hides a circularity. The corpus is not a record of inscriptions as they survive on stone. It is a record of inscriptions as editors have published them — and editors restore. Every […] in EDCS is a sequence of letters that is not on the stone: it is an earlier editor's prediction, printed as text.

How much of the corpus is prediction? In the 537,262-inscription export, the editorial bracket [ opens 1,086,124 times — an average of 2.02 restorations per inscription. The corpus is not a record of what the Romans wrote. It is a record of what the Romans wrote plus two centuries of what editors predicted they wrote — and the two are interleaved in the same strings, often indistinguishable to a frequency count.

The circularity is now visible. We restore a damaged D as Dis because D(is) is the corpus's commonest expansion of D. But a large share of those D(is) attestations are themselves [D(is)] — restored, by exactly the reasoning we are now applying, by earlier editors. A frequency search cannot easily separate the D(is) that is on the stone from the [D(is)] that an editor supplied. The frequency that licenses the restoration was produced by restorations. Predict the mode and you do not discover what D meant; you ratify what editors have always assumed it meant — and you ratify their errors with their successes, because a wrong restoration, once printed, is counted as evidence for the next.

This is not a marginal worry. It is the methodological floor of quantitative epigraphy. Kaše, Heřmánková and Sobotková's 2022 study counts occupational terms across Roman cities to measure the division of labour — a genuinely valuable enterprise, and one this course recommends. But it counts them in a dataset (LIRE) derived from EDCS. Every quantitative result drawn from such a corpus inherits its restorations: the numbers describe, in part, the editorial tradition that built the corpus, not only the society that cut the stones. The honest quantitative claim is always conditional — "given the corpus as restored."

Challenge II located the disambiguating work in context. Challenge IV measures what happens when context is removed — which is, precisely, what damage does.

An abbreviation is a compression. It can be decompressed because the reader supplies what the writer omitted — and the reader supplies it from context: the genre of the text, the formula in progress, the words on either side. The cruel structure of epigraphic damage is that it attacks the abbreviation and its context together. The same break that leaves you needing to restore M also removes the lines that would have told you whether M is Manibus, municipii, miles, mater, memoriae, or the praenomen Marcus.

Step through an inscription as the stone breaks. Watch the candidate set widen:

The shape of the result is the heart of Challenge IV. Predictive restoration is most confident when the text is nearly whole — when there is least to restore. It is least reliable when the text is badly broken — when there is most to restore. The method's accuracy is inversely correlated with the situation that calls for it. Where restoration is easy it is barely needed; where it is needed it is barely possible. The intact neighbours that make M obvious are a loan from the part of the inscription that did not need help — and a lacuna is, by definition, the place where that loan is unavailable.

A large language model is a prediction engine. Trained to continue text, it answers every prompt by emitting the most probable next tokens. Asked to complete a damaged inscription it does exactly what the predictive editor does — consult an internalised store of parallels and produce the likeliest continuation — but with two differences. It has read far more than any editor. And it has no brake.

A human editor completing [D] M S on a cippus carries — even if Challenge I showed the notation hides it — some sense of the entropy of the move. The editor can decline: can print [- - -], can mark a reading (?), can leave the lacuna open and say so. A language model, by default, does none of this. It is fluent by construction. It will expand D M with the same easy confidence whether the answer is a 1.26-bit near-certainty or a 5.32-bit guess. It almost never returns the most honest epigraphic answer — "the surviving letters do not determine this."

Use the tool below. Give DeepSeek a damaged text and read its restoration critically — against the entropy of Challenge I, the genres of Challenge II, the circularity of Challenge III, the collapsed context of Challenge IV.

The five Latin challenges turned on one engine of prediction: the frequency with which an abbreviation takes each expansion. Greek epigraphy supplies a second, quite different engine — and seeing it sharpens what the Latin one was.

Most Attic public inscriptions of the fifth and fourth centuries BC were cut stoichedon: every letter occupies one square of an invisible grid, aligned both across and down. Once the editor establishes how many letters fill one line, the letter-count of every lacuna on the stone is fixed — not estimated, counted. It is the strongest constraint anywhere in this module.

From 454/3 BC the Athenians inscribed, year by year, the aparche — the "first-fruits", one-sixtieth of the tribute (phoros) paid by each allied city, dedicated to Athena. The first stele, the lapis primus (IG I³ 259–272), is the most relentlessly formulaic Greek corpus there is: a column of entries, each [city] ⫶ [amount], repeated hundreds of times.

That formula, plus stoichedon counting, let the great edition — Meritt, Wade-Gery & McGregor's Athenian Tribute Lists (1939–53) — restore vast stretches of broken stone. Björn Paarmann's 2007 re-edition deliberately reverses the move. Toggle the entry below between the two editorial philosophies.

The Athens–Egesta alliance (IG I³ 11) carries its own date only in the archon's name, line 3 — of which just the final letters ]ΟΝ ΕΡΧΕ ("…on was archon") are securely read. In 1944 Raubitschek restored the rest as Habron (458/7 BC). The early date was then locked in place by a letter-form rule: the "three-bar sigma" was held to have gone out of use by about 446 — and this decree carries it thirteen times.

So a single restored archon-name, propped on a single letter-form, dated one step in the history of the Athenian empire. Harold Mattingly fought the rule for decades. In 1990 a laser beam, fired through the marble, re-read line 3 as Antiphon (418/7) — forty years later. Pick the archon and watch the history fork.

Three predictive engines have now been on the table: the frequency of a Latin abbreviation's expansions, the geometry of a stoichedon grid, and the formula of a genre. They are not the same kind of thing, and a restoration is only ever as strong as the engine beneath it.

Badian's two specimens are one-word bets that carry whole histories. In IG II² 399 the gap reads either πολ[εμί]ων ("enemies") or Moretti's [ληιστ]ῶν ("pirates") — and on that single word turns the reconstructed history of Athenian involvement in the war of Agis III. In the Philippi letter of Alexander, whether to restore the royal title [βασιλέα] would "settle" a long dispute over Macedonian kingship — and Badian's verdict is that it "does not seem a necessary and inevitable supplement". The bracket is doing the work the evidence cannot.

Challenge V's machine has a Greek-trained counterpart. Ithaca (Assael, Sommerschield et al., Nature 2022) is a deep neural network for the restoration, dating, and geographic attribution of Greek inscriptions — trained on the corpus itself, and built to assist rather than replace.

One workbench, three jobs. Generate a restoration; Compare a Greek and a Latin problem; or Critique a restoration — set the model, in Badian's role, against a published text or against its own output. Pick a mode; the prompt loads; edit it freely; run.

Read it for its method as much as its findings. As you read, hold Challenge III in mind: the study's evidence is occupational vocabulary counted in a Latin-epigraphic dataset (LIRE) derived from EDCS. Ask, at each result: which of these counts could be affected by editorial restoration, and by the survival bias of the "epigraphic habit"? The paper is a model of quantitative method; the challenges of this module are the conditions under which its numbers should be read.

Frequency and entropy figures were computed for this module from the EDCS text export (EDCS_text_cleaned_2022-09-12.json, 537,262 records; 1,086,124 restoration brackets) held in epidoc/big databases/public edcs. The abbreviation menus are transcribed from OHRE Appendix II. Example inscriptions are quoted from EDCS and cited by EDCS-ID. The wider data ecology referenced in the challenges — LIRE (the Latin Inscriptions of the Roman Empire dataset), the EDCS_ETL and Lat-Epig pipelines, and the kcl_tei EpiDoc corpora (Aphrodisias, Tripolitania, Cyrenaica) — is held in epidoc/inscription_databases.

The Greek track is built on the works below; quotations are reproduced from them, page-verified.

The four editions exist in full at the Governance Corpus with SCPP-standard line-keyed apparatus, hover-glossary, and EpiDoc TEI XML downloads. This section is the deep-dive layer between the §0 mosaic (one paragraph per case) and the full editions (the entire text + apparatus + commentary). It surfaces, for each case, the formula bank the editor consulted, a verbatim passage with restoration brackets visible, and the chain of links into the rest of the family.

The Laudatio's left column lines 30–34 carry the virtue-catalogue slot — a fixed eight-virtue inventory the genre uses (pudicitia, obsequium, comitas, facilitas, lanificium, religio sine superstitione, ornatus non conspiciendus, cultus modicus). Mantzilas 2017 finds the same eight-slot bank in five Latin laudationes mulierum; that is the parallel pool the editor consulted to restore the words Sirmond's transcript could no longer read.

The Segesta decree opens with the Athenian decree-prescript formula: ἐπὶ + ἄρχοντος-genitive. Only the letters ]ΟΝ ΕΡΧΕ are securely read; the archon's name is restored. For 46 years the restoration was Habr]on (Wade-Gery, on three-bar sigma palaeography → 458/7 BCE). In 1990, Chambers, Gallucci, and Spanos used laser-enhanced photography and read Antiph]on → 418/7 BCE. The 40-year shift cascades into the entire chronology of Athenian imperialism: every decree dated relative to this one moves with it.

40 years of annual aparche records (454/3–415/4 BCE). Each list opens with a fixed year-prescript (τάδε ἀπὸ τοῦ φόρου τῷ θεῷ ἀπὸ τοῦ τάλαντου), then lists cities in alphabet-blocks with their aparche figure to the right. The formula is so regular that the editors of the ATL (Meritt, Wade-Gery & McGregor 1939–53) routinely restored both the city-name and the quota figure from the surrounding entries. Paarmann 2007 has argued they restored too much — that the formula tells you a name preceded a figure but does not entitle the editor to write the name. Challenge VII above shows the contrast: ~46% restored (ATL) vs ~9% restored (Paarmann).

Persicus's edict on the Artemision (AD 44) survives in both Greek and Latin. The Greek face translates the Latin: Paullus Fabius Persicus, proconsul becomes Παῦλλος Φάβιος Πέρσικος ἀνθύπατος. The two faces let us watch a formula cross the language barrier — which Latin terms stay translated, which become Greek loanwords, which get expanded into glosses. This is the bridge between the Latin track (Challenges I–V) and the Greek track (Challenges VI–IX).

Each engine treats a lacuna differently. An LLM (DeepSeek, Claude, OpenAI) is trained on internet-scale text including epigraphic editions; it fills lacunae the way it fills any cloze — with the most likely token. Ithaca (Assael et al. 2022, Nature 603) is a transformer trained specifically on the I.PHI Greek epigraphic corpus, producing top-k character predictions plus chronological and geographic attribution. Aeneas / predictingthepast is the DeepMind successor work, structured around contextualised parallels rather than character-level prediction.

Pick a passage below. Each engine returns its top candidates and a tagged basis (formula · stoichedon-geometry · frequency · free-conjecture). The comparison strip beneath the three columns lines up their top-1 outputs side by side. The pedagogical question: when the engines agree, where does the agreement come from? When they disagree, which is the responsible editor's friend?

EDCS holds 537 262 Latin inscriptions; their provincial distribution is wildly uneven. The bars in the workshop's Challenge V show this numerically; the map below shows it geographically. Pins for the four Wk13 case studies are emphasised in their period colour; smaller dots show the broader corpus density.

What this layer adds. Challenge V plots provincial counts as bars; that's a histogram of where the corpus was produced. The map shows the same data as geography. The four Wk13 cases sit in: Rome (Latium/Italy), Segesta (Sicilia), Athens (Achaia), Ephesos (Asia). Three are coastal; one is the metropole. None is in Britannia, in Hispania, in Pannonia — the empire's epigraphic density does not mirror the empire's territorial extent. Each pin links to the canonical Pleiades record. The verified Pleiades URI table is at governance-corpus.html and was built under M-PLEIADES.1.