S U M M A

Summa is a weekly editorial digest of new scholarship in late Roman history, defined as the period from roughly 250 to 500 CE in the territories of the Roman Empire and its frontiers. It is written for scholars working in the field, graduate students orienting themselves in the literature, and serious readers who want to follow late Roman scholarship at the pace of the field itself.

Each Saturday issue surveys recent publications, reviews, and discussions across forty languages and sixteen connected academic fields — the languages and disciplines where the relevant scholarship actually appears. The result is presented as a single editorial document with bibliographic attribution, source links, and Anglophone bridges for non-English work.

Summa is centred on Rome. Adjacent fields — Sasanian studies, Syriac, papyrology, palaeogenetics, manuscript studies, and the others listed below — are surveyed for what they contribute to late Roman scholarship, not as ends in themselves. Languages beyond English are surveyed because that is where a substantial share of the field's working scholarship is published.

How issues are made

Every week the editor surveys forty languages and sixteen adjacent fields for new work bearing on the period. Candidate items are evaluated against published-this-week relevance, significance to live scholarly debates, and the editor's reading of the field. Each item gets a 2–3 sentence summary naming the scholar, the work, and the argument's significance. Non-English items get an explicit Anglophone bridge — what does this French paper contribute that English-language scholarship hasn't?

Summary synthesis is AI-assisted and runs against an automated editorial audit before send. The audit checks voice (banned phrases, hedging chains), item length, IP (no verbatim reproduction of copyrighted text), Anglophone bridge presence on non-English items, significance-flag balance, bibliographic plausibility (named scholars, source-registry match, optional URL liveness), section completeness, and cross-issue deduplication. Editorial judgement on what to include and how to frame it remains with the editor; the AI handles search breadth, language coverage, and routine summary drafting.

Where evidence is contested, Summa says so. Where translations are uncertain, Summa flags it. The discipline of accuracy is the point; the speed and breadth are the benefit.

Full methodology

What follows is the longer account: what the pipeline actually does, where the machine assists and where it defers, what is checked before an issue goes out, and where the limits are honest. If you are trying to decide whether Summa is legitimate scholarship-adjacent work or automated content, this is the section that should tell you.

The weekly cycle

Every Saturday at noon UTC, a job runs on GitHub Actions that does four things in sequence: it gathers candidate items from the past week's late Roman scholarship across sixteen disciplines and forty languages; it drafts the issue in Summa's editorial format; it runs that draft through fourteen audit checks; and, on a clean audit, it sends the issue by email and publishes it to the /issues/ archive. The whole cycle takes ten to twenty minutes. The editor reads every draft before send, though the automated send path can proceed if the editor is unreachable and the audit is clean.

The gather step is the slow one. It calls Anthropic's Claude Sonnet with the web-search tool and the following instruction set: survey seven buckets of source (Tier 1 academic journals; podcasts; newsletters and blogs; second-tier books and journals; non-English scholarship; hard science; adjacent disciplines) across the named languages in each bucket, respect a source registry of roughly one hundred and twenty-five accepted venues, and return a structured JSON payload with per-item metadata. That payload is the raw material the issue is built from.

The gather pipeline

Summa's gather is not a general web search. It is a targeted survey of a defined set of venues, per bucket, per language. The Tier 1 humanities journals are BMCR, JLA, JRA, JRS, and TMR; Tier 1 non-English sources include HAL for French, Antiquité Tardive and Edipuglia for Italian, De Gruyter and Mohr Siebeck for German, Dialnet and Emerita for Spanish, Repositório Aberto for Portuguese, Byzantina Symmeikta for modern Greek, Mnemosyne for Dutch, and the CCSL/CSEL series for critical Latin editions. The adjacent disciplines have their own registries: ZPE and BASP for papyrology, Numismatic Chronicle and Revue Numismatique for numismatics, Vigiliae Christianae and JECS for patristics, Byzantinische Zeitschrift and BMGS for Byzantine studies, and so on across the sixteen fields.

The full source registry is published above. It exists in the code as a Python list against which every gathered item's source string is checked. An item whose source does not match the registry is not discarded, but it does not accrue a NOTABLE top_journal criterion from source alone. The registry is extended when new venues come into view; three sources were added in June 2026 (Digital Numismatics blog, an arXiv digital-humanities block, and a small set of institutional repositories), and substantially more in July 2026 as part of a multilingual coverage push. Every extension is documented in the git history.

When an item's source or byline reads "surfaced by X" (a blog, an aggregator, a mailing list, a colleague), that reflects how the piece came into Summa's field of view. The gather searches known venues directly, but scholarship is also transmitted by proxy: BMCR's books-received listing, LT-ANTIQ posts, the AWOL feed, personal recommendations. Where the proxy is the reason an item was found, Summa says so.

The significance threshold

An item is eligible for a Summa issue if it addresses the period between roughly 250 and 500 AD (with soft edges at each end), if it was published or surfaced within the past seven days, and if its subject matter would be intelligible to a working academic in the field. Volume constraints apply: Summa aims for roughly fifteen to twenty-five items an issue, but there is no floor. Honest quiet is preferred to padding. Issue III (June 2026) shipped eleven items; Issue II shipped thirty. Both are within the normal range.

The threshold is a judgement call, not a formula. The gather step scores items against explicit criteria (venue prestige, cited scholars, novelty of evidence, disciplinary reach), but the final inclusion decision is editorial. A Byzantine coin hoard report from a specialist journal may pass the threshold in a thin week and fall below it in a busy one. Summa's discipline of accuracy is strict; its discipline of completeness is deliberately not.

The flags: NOTABLE and FIELD-SHIFTING

Two significance flags appear in Summa. Both are rare. The rarity is the point: a subscriber who sees a ◆ mark should know that something out of the ordinary run of weekly scholarship is being called out.

NOTABLE is applied when an item meets at least two of four criteria: publication in a top-tier journal (JLA, SLA, JRA, JRS, Antiquity, ZAC, Nature, Science, PNAS, Cell, or a premier monograph series such as Cambridge Studies in Late Antiquity or Oxford Late Antiquity); a direct challenge to a currently-accepted scholarly position; a genuinely new primary source, dataset, or material find; and authorship by a defining voice in the sub-field. The cap is two NOTABLE items per issue.

FIELD-SHIFTING is applied when a NOTABLE item plausibly reframes a named debate. The cap is one per issue and most issues have none. Confirmatory or extending papers from the same research consortium whose earlier work has already been flagged receive NOTABLE at most, not FIELD-SHIFTING; the reframing has already been done. The one time this rule has bitten, the HistoGenes consortium's Pannonian ancient-DNA paper in Science (Issue IV) was correctly downgraded to NOTABLE because the same group's Rhine paper had been flagged FIELD-SHIFTING in Issue II. The full calibration framework, including worked examples and a science-bias guard that requires a named humanities-side debate connection before a science item can be flagged, is documented in the repository's FIELD_SHIFTING_FRAMEWORK.md.

AI-assisted, editor-judged

This is the section academic readers should read most carefully.

The AI is Anthropic's Claude, in two roles. Claude Sonnet does the gather: search, source registration, first-pass summary drafting, structured field extraction, candidate significance assessment, and the drafting of the Anglophone bridge sentence that every non-English item carries. Claude Opus (or, in some editorial tasks, a comparable model) drafts the tonal pieces: the Correction each week when the pre-vetted bank is exhausted, the Number's prose when the bank is exhausted, the Translation Window's WHY IT MATTERS section, the signals frame that opens the Significance Summary, and the On This Day entries. Both models operate under the constraints written into Summa's VOICE.md, which the gather and drafting prompts embed by reference.

The editor does everything else. The final decision on which items enter which issue is editorial. The final significance calls (does this really merit NOTABLE? is this genuinely FIELD-SHIFTING or is the enthusiasm overrun?) are editorial. The choice of misconception for The Correction each week, when the editorial bank offers options, is editorial. The choice of translation window (which untranslated foreign-language work an English-language reader should know about) is editorial. The tonal review of every issue before send is editorial. Every correction to an AI-produced misattribution or fabrication is editorial. No item ever ships to subscribers without the editor having read the surrounding paragraph and, in cases of doubt, the source itself.

The pipeline is designed so that the machine handles breadth (forty languages, sixteen disciplines, a hundred and twenty-five sources) and the editor handles judgement (does this matter, is it stated accurately, does the framing serve the reader). No individual human being can survey forty languages of new scholarship every week; that is Summa's premise. But no automated system can reliably decide, this week, that the Codex Rossano article matters more than the Sasanian coin catalogue, or that the Pannonia aDNA paper is confirmatory rather than reframing. That decision requires a reader who is following the field; the machine gathers material for that reader to act on.

The stakes: if Summa lets the AI make the significance calls, or ships items without a human editor having looked at them, the project becomes automated content and academic readers will correctly disregard it. The commitment against that outcome is explicit and built into the send path: no issue is dispatched without the editorial review step being registered in the send log. Where that discipline lapses, subscribers should tell me, and I will correct it publicly in a subsequent issue.

The Anglophone bridge

Every non-English item in a Summa issue carries a sentence addressed to the reader who cannot read the original. The sentence names an Anglophone scholar or a named debate, and states what the non-English work specifically contributes. The reader who does not read Italian can still learn from a Zito coin article by knowing that it engages the Wiebe and Noreña debate on Julian's coinage with a Neoplatonic textual dimension previous Anglophone work has not developed. Without the bridge, the item is a foreign citation. With it, the item is usable.

The audit checks this. Bridges that name no Anglophone scholar or debate, or that use vague language ("adds to the literature") in place of specifics, are flagged before send. The check catches weak bridges reliably; it does not catch bridges whose scholarship is fictitious, which is why editorial review of the source itself is required for every non-English item that receives a bridge.

The editorial banks

Two of Summa's weekly features run against pre-vetted banks. The Number bank contains twenty entries covering the range of specific numbers in late Roman scholarship that are worth knowing about (the size of the Diocletianic army, the weight of Constantine's solidus, the surviving fragments of the Digest, Attila's death date). The Correction bank contains twenty entries covering popular misconceptions, each written to prd_13's four-section structure. Both banks are rotated: the state file records which entries have been used, and the picker returns the next unused one. When a bank is exhausted, the picker either begins the cycle again (as with the Number bank in current use) or falls back to fresh generation (as with the Correction bank when a specifically topical misconception is preferable to a rotation entry).

The banks exist because week-by-week generation of these features produced inconsistent quality. A pre-vetted entry, drafted with care and fact-checked against the scholarly literature, is more reliable than an ad-hoc weekly draft. The banks also let the editor extend Summa's coverage into misconceptions and numbers that would not naturally come up in a given week's news cycle but are worth naming.

The audit

Every issue draft is run through fourteen automated audit checks before it is sent. The checks cover voice (banned words and phrases per VOICE.md), item length (2–3 sentences with a hard 4-sentence cap), intellectual property (no verbatim reproduction of copyrighted source text longer than six words), the presence of an explicit Anglophone bridge on every non-English item, significance-flag balance (post-cap NOTABLE and FIELD-SHIFTING counts within limits), bibliographic plausibility (author-name format, publication-year sanity, source-registry membership, optional URL resolution), section completeness (all five required sections present and non-empty, with the Translation Window's variant handled correctly), cross-issue deduplication (no item whose URL or normalised title appeared in a prior issue), signals frame presence when flagged items are present, preheader format (three period-separated clauses, no ellipses, under 130 characters, no clauses ending in a stopword), TODO-marker scan (nothing labelled [NEEDED], [TODO], [PLACEHOLDER] survives to send), multilingual delivery (WARNING when fewer than four distinct languages appear in an issue), scholar attribution in bridges (WARNING when a non-English item summary names no scholar), and FIELD-SHIFTING calibration drift (WARNING when a flagged item lacks a machine-readable sig_reason field or when three consecutive issues carry zero flags across substantial item counts).

The audit produces a verdict of PASS, REVIEW, or BLOCK. PASS ships; REVIEW is inspected by the editor and, on the editor's call, ships; BLOCK does not ship until the underlying issue is fixed. The audit does not replace editorial review; it catches mechanical drift.

Corrections

Where Summa is wrong, Summa says so. Corrections come in via editor@summa.ac and are acknowledged in a Corrections block in the following issue. Substantive corrections are named; typos and minor errors are silently fixed in the online archive. Systematic patterns of error (a persistent misspelling of a scholar's name, a persistent misattribution of a debate to the wrong author) are addressed by editorial change to the underlying prompt or bank, so that the error is not repeated.

The archive at summa.ac/archive is the authoritative record. The email is a delivery mechanism; the archive is where subscribers, librarians, and future readers should turn. Nothing is silently overwritten in the archive except the mechanical corrections just described; substantive changes are logged.

Why the methodology is public

Working academics are the audience that most needs to know how the information they are being handed was assembled. Summa is AI-assisted current-awareness scholarship, and it earns its place in the reader's inbox only by being unambiguous about what the machine did and what the editor did, and about where each is strong and where each is fallible. The choice to publish the methodology, name the pipeline components, expose the source registry, describe the audit, and mark the banks explicitly is part of what it means for the project to be legitimate. If Summa concealed how it was made, the reader would be right to distrust it. Because it does not conceal, the reader can judge on the merits.

Languages

Forty languages, organised in three confidence tiers.

Sixteen connected fields

Eight humanistic, eight material. Surveyed when they bear on late Roman scholarship.

Source registry

The journals, reviews, podcasts, blogs, and field-report venues surveyed every week. The list grows when sources consistently produce period-relevant scholarship; retired sources drop entirely.

Infrastructure

Summa's editorial process draws on the broader digital humanities infrastructure for late antique scholarship. The following reference databases, research projects, and tools are part of the field's working apparatus; Summa does not redistribute their content, but credits the work that makes the field navigable.

Reference databases

• Pleiades — gazetteer of ancient places
• Pelagios — linked open data tooling for the ancient world
• DigilibLT — Digital Library of Late Antique Latin Texts
• Trismegistos — papyrological and epigraphic portal
• Papyri.info — papyrological aggregator
• Nomisma.org — numismatic linked open data
• Corpus Nummorum — numismatic database
• Syriaca.org — Syriac reference
• Open Greek and Latin — open digital texts

Research projects

• Connecting Late Antiquities — Exeter / Bonn / London prosopographical project for the third to seventh centuries
• Prosopography of Later Roman and Byzantine Worlds — Institute of Classical Studies, London

Tools

• Aeneas — Google DeepMind / Oxford AI model for restoring and contextualising Latin inscriptions

The editor

Summa is edited by a Roman History, product and tech enthusiast. The motivation is simple: the late Roman field publishes faster and more multilingually than any individual can track in parallel with their own work. Summa is the survey-and-presentation layer that makes the field's output legible at a weekly pace, for free. Correspondence and corrections go to editor@summa.ac.

Why Summa exists

The field of late Roman scholarship has expanded on every axis at once, and no single working scholar can now keep up with it. It includes the palaeoclimatology of Kyle Harper and Ulf Büntgen on the sixth-century plague and the LALIA hypothesis; the population genetics of the HistoGenes consortium at Mainz, whose 2026 Rhine and Pannonia papers reshaped what the discipline can say about migration and demographic change; systematic attention to Sasanian Iran as a peer polity of the late Roman state, following Richard Payne, Matthew Canepa, and Garth Fowden; and everything the field did before, from the Heather and Wickham exchange on collapse-or-transformation to the Christological controversies to the Justinianic legal corpus. The cost of not keeping up is not always visible. It shows up as citations to arguments already qualified by French, Italian, or German scholarship the reader has not seen.

Existing resources cover parts of this. LT-ANTIQ, the field's mailing list, is unmoderated by design. AWOL, Chuck Jones's Ancient World Online, is pan-antique and open-access-only. The Mary Jaharis Center runs a daily Byzantine digest. Studies in Late Antiquity's newsletter is annual. Journal table-of-contents alerts move at publication cadence and do not synthesise. Summa's specific niche is where none of these overlap: weekly cadence, multilingual coverage, cross-disciplinary reach, and consistent editorial synthesis in a working scholar's voice.

Summa is not a research output; it surveys and describes, and does not add new arguments. It is not a knowledge base; the archive is a chronicle of the week's increment. It is not a substitute for reading the scholarship, but a survey layer that helps a reader decide what to read carefully in full. It is not global late antiquity, though it borders on it: the scope is 250 to 500 AD with soft edges, centred on Rome and its neighbours. It does not push a particular hermeneutic. It is not a paid product and it will not become one.

The measure of success is not open rates but whether working scholars learn something on Saturday morning they would not have learned otherwise. Corrections and disagreements go to editor@summa.ac.

A 2025 De Gruyter / Brill survey of 1,800 researchers found that 75% prefer email newsletters over social media for staying current with their field.

Contact

Corrections, suggestions, and reader correspondence: editor@summa.ac.