Modeling decisions in changing context

A historian finds two Roman inscriptions with an unusual phrase in common. One was dedicated fourteen years before the other. They were found near one another in Mainz.

The resemblance opens several questions. Did the later writer copy an earlier dedication? Did both follow a local practice? Does their choice of deity tell us where the dedicators came from?

These questions need different comparisons. A nearby dedication can help explain a local practice without establishing a person’s birthplace. Two texts can resemble each other closely while leaving the historian’s present question unanswered.

Imagine a research instrument that lets the historian work through these possibilities together. Selecting a passage brings up related texts. Choosing an explanation changes which comparisons receive attention. Opening a source can reveal that a decisive word is an editor’s addition. The instrument then helps examine the alternatives that remain.

This article develops that instrument as a design proposal. The historical material and cited research are real; the continuing interaction described here has not been built or measured. The subject is how to train models for decisions whose meaning changes during the work.

Start with the comparison

The Mainz altar catalogued as CIL XIII 6665 names Lucius Maiorius Cogitatus. It records a dedication to the Aufanian goddesses and the protective deity of the place. The date is 15 July 211 CE. The Heidelberg edition preserves the transcription and its editorial expansions. Epigraphic Database Heidelberg, HD054789

In their Aeneas study, Assael and colleagues describe a retrieved parallel: a dedication by Iulius Bellator, dated 15 July 197 CE. The texts share an unusual expression concerning the safety and well-being of the dedicator and those close to him. The later archaeological finds also provide a local relationship between the monuments. This makes the pair useful for investigating a shared dedicatory practice. It does not, by itself, identify how that relationship arose. Contextualizing ancient texts with generative neural networks, Figure 3 and accompanying discussion

For the proposed instrument, the historian selects the shared expression and asks: “What could explain this wording here?”

Several searches can now serve the same inquiry. One looks for the expression elsewhere. Another looks for related dedications in the area. A third examines comparable uses by people with the same recorded office. Each can produce useful material, but usefulness depends on the relationship under investigation.

A model might find another close verbal match. A second model must decide what kind of comparison it offers. Does it repeat a general formula? Does it name the same people? Does its date permit the proposed direction of influence? Does it merely contain common dedicatory language?

The result should help the historian act. “Possible earlier example of this formula” invites a source comparison. “Same deity, no evidence of this person’s origin” keeps a cultural association from becoming a biographical conclusion. “The decisive name is supplied by an editor” directs attention to the edition and its argument.

These are learned judgments about the significance of evidence. They need more than a similarity score.

A question changes what a parallel can support

Keep both Mainz records open. Now change the question to “Can this dedication establish where Cogitatus came from?”

The wording and dates have not changed. Their role in the investigation has. A comparison that matters for local religious practice may offer only a weak clue about personal origin. The instrument should identify that limit and suggest evidence appropriate to the new question: an explicit origin statement, a securely identified person in another document, or an argued account of how this cult was distributed.

It must also distinguish those proposals. An explicit statement and an inference from religious practice are different forms of support. Two people with similar names need not be the same person. A local findspot can record where an object ended up, rather than where a person was born.

This gives the model a more precise task than “is this document relevant?” Its input includes a claim, the passages under comparison, and the evidence needed to interpret their relationship. Its answer describes both the relation and its limits.

For example, a customer could define these outputs:

• Formula parallel: a passage offers a comparable use of an expression.
• Possible person match: records may concern the same person, with the identifying evidence stated.
• Historical association: the comparison bears on a practice, place or group, without establishing identity.
• Unresolved relation: the available material does not distinguish the proposed explanations.

These categories can overlap. A text can supply a formula parallel and a historical association. The instrument should preserve both, with the relevant passages, instead of forcing every comparison into one universal ranking.

The next operation depends on that answer. A formula comparison opens aligned passages. A possible person match opens the identifying details and conflicting evidence. An unresolved relation opens the questions that further evidence could distinguish.

Preserve the difference between a text and a reading

Historical texts often reach a model through an edition. An editor may expand an abbreviation, identify an uncertain letter, or supply words missing from the object. Those operations have different meanings.

EpiDoc already provides ways to encode such distinctions, including uncertain readings, restorations and alternatives. A research tool should retain them when it prepares model input. Flattening all of them into ordinary text removes information the investigation needs. EpiDoc: restoration, uncertain restorations, alternative readings

Consider a separate, constructed test case. Two inscriptions seem to name the same office. In one version of the test, that title survives on both objects. In another, the title in the second text is a scholarly restoration based partly on the first text.

The visible words can be identical. The second version does not provide an independent surviving occurrence of the title. A model that sees only the words can miss this distinction entirely.

The useful response also depends on the historian’s question. For “How has this passage been reconstructed?”, that editorial dependence is directly relevant. For “How many surviving texts attest this office?”, it prevents the restored passage from being counted as another surviving attestation.

A source tool can expose the markup and the editor’s cited basis. The model then interprets what that dependence means for the current claim. Neither part can replace the other. Markup alone does not settle the historical argument, and a fluent judgment cannot recover evidence that the input discarded.

Let several explanations remain active

A continuing investigation needs more than a stream of replacement answers. The historian may want to pursue direct copying and a shared formula at the same time.

For a copying hypothesis, the instrument could seek distinctive shared details and evidence that permits contact or transmission. For a shared-formula hypothesis, it could seek independent occurrences beyond the pair. Those investigations can proceed in parallel once their questions are defined.

Their results can change one another. A third occurrence may weaken a claim that a phrase is unique to the pair, while strengthening a broader account of local practice. A revised date may rule out one proposed direction of copying while leaving the common-formula explanation open. The tool must preserve which claim each judgment addresses.

This suggests a division of work:

1. A retrieval model finds candidate passages for each active question.
2. Source tools open the records, align passages and expose editorial status.
3. Relation models assess what each candidate contributes to each explanation.
4. A further judgment identifies the distinction that the current evidence leaves unresolved.
5. The historian chooses which distinction to investigate and which claims to retain.

The fourth step deserves its own evaluation. “Find more examples” is often too vague to advance the work. “Look for this formula outside this locality” identifies a contrast. “Check whether this title survives or was supplied” identifies a source inspection. The proposed action should address an uncertainty that could change the interpretation.

These boundaries need not become five separate deployed models. Some tasks may share a model; some may require a larger research model or human work. The important boundary is the question each task answers and the evidence it receives. A small model cannot assess a historical relation from two isolated snippets if the relation depends on omitted dates or editorial arguments.

Connect existing research to the missing task

Ithaca jointly models restoration, geographical attribution and dating of ancient Greek inscriptions. Its restoration study uses deliberately hidden characters in known text. The paper also explains that historians bring material and historical knowledge beyond its textual input. This supports using several related predictions while keeping their evidence boundary explicit. It does not establish the continuing instrument proposed here. Restoring and attributing ancient texts using deep neural networks, methods and discussion

Aeneas also exposes a useful software boundary. Its retrieval interface receives text; attribution has a separate optional image input. A historian’s current claim and supporting argument are not separate inputs to that retrieval operation. A customer model that judges what a parallel supports would add a different task. Aeneas inference implementation

There is a further dependence to preserve. Restoration, dating and retrieval can share training material and representations. Agreement among them can be useful, but it is not equivalent to several independent historical witnesses. A proposed word can influence which parallels appear; those parallels cannot then count as wholly independent confirmation of the same proposed word.

The interface can make this relation understandable. Keep a conjecture marked as a conjecture. Let the historian compare searches with and without it. Show which comparisons survive that change. The purpose is to make alternatives easier to investigate while retaining an honest account of why each one seems plausible.

Train on changes in the investigation

Training only on isolated document pairs would omit much of the intended work. A model could learn that shared names or phrases usually mean “relevant,” then give the same answer after the question changes.

Build training examples around a specific investigation state: the current claim, admitted evidence, editorial alternatives and available source operations. Include the relation to assess and the action it would justify.

The data needs controlled contrasts:

Change in the example	What a qualified model should distinguish
Same records; local practice becomes personal origin	A useful cultural comparison does not establish the person’s birthplace.
Same displayed title; surviving text becomes an editorial restoration	Evidence about a reading differs from an independent surviving attestation.
Same wording; another independent occurrence appears	A claim of uniqueness can change while the value as a formula parallel remains.
Same two names; incompatible dates appear	Similar names do not preserve a proposed person match through contradictory evidence.
Same claim; an unrelated record is added	The existing interpretation need not change merely because the context grew.

These are proposed evaluation distinctions, not automatic historical truth labels. Specialists must assess the cases, the evidence and the permitted conclusions. Some examples should retain more than one defensible reading.

Synthetic data can vary questions, presentation and evidence availability around those reviewed cases. It should preserve the source relationships. Inventing a convenient inscription and treating it as a real witness would corrupt the task. Likewise, a generated rationale that mentions a source the model never receives would train an impossible decision.

Keep variants of the same inscription and dependent editions in the same evaluation partition. Otherwise, one version can appear in training and a near duplicate in the test. Test on the candidates that the actual retriever returns, including convincing but unhelpful matches. A model evaluated only on carefully selected pairs may fail when the retrieval stage supplies difficult neighbors.

Measure the relation and the resulting action separately. A correct “unresolved” answer with a fabricated reason is a failure. A good explanation followed by a search that cannot distinguish the alternatives is also a failure of the instrument.

Evaluate the work the person can now attempt

A general model with access to the same sources is a serious baseline. It may already conduct much of this investigation well. The proposed division should earn its complexity through better work, useful response times or lower cost for repeated comparisons.

Give each approach the same source access and ask historians to work through complete investigations. Include changes of question, editorial corrections and cases that remain unresolved. Compare which defensible interpretations they consider, which unsupported claims they adopt, and whether they can recover after a misleading parallel.

Measure time to a useful comparison separately from time to a finished argument. A fast local response cannot establish that the whole investigation is fast. A compelling interface cannot establish that its historical judgment is sound.

The intended change is in what the historian can afford to explore. Several explanations can remain visible while their supporting comparisons develop. A change of question can redirect the investigation without discarding the sources already understood. The person can follow an unexpected connection while still seeing which parts of the argument depend on it.

That is the design opportunity in changing context: give each consequential judgment the current question and the evidence it needs, then make its result useful for the next act of investigation.