方括号与概率

αικην επαρχηαν υπεξειρημενων των υποδικων κεφαλης υπερ ων ος αν την επαρχηαν διακατεχη αυτος διαγεινωσκειν κ?? ισταναι η συνβουλιον κριτων παρεχειν οφειλει υπερ δε των λοιπων πραγματων παντων ελληνας κριτας διδοσθαι αρεσκει ει μ

编者依据：No apparatus note at this point in the edition. The supplement completes καί at the end of l. 66; grounds are word-completion plus the required coordination of the two infinitives διαγεινώσκειν ... ἱστάναι.

读法：Function word in a fully formulaic legal coordination. Ithaca gives it 0.873 with a clean decay to the runner-up (0.032). The easiest class of bracket: the editor and the model are answering the same low-entropy question.

lepidi consulis collegae praesentis pedibus aduoluta humi non modo non adleuata sed tracta et seruil?? in modum rapsata liuoribus corporis repleta firmissimo animo eum admoneres edicti caesaris cum gratulatione restitu

编者依据：The edition prints the supplement without a dedicated apparatus note at l. 55 (the listApp entry at loc. 30 documents a different, virtue-catalogue slot, restored "on parallel-pool grounds from the other laudationes mulierum (Mantzilas 2017)"). Grounds here are idiomatic: seruilem in modum is a fixed adverbial phrase requiring the accusative singular.

读法：DIVERGENCE. Aeneas ranks the editor 5th at 0.050, preferring seruil[ia] (0.353) and seruil[is] (0.332). Every one of its top four is a morphologically possible Latin ending; what it does not command is the frozen idiom X in modum, which fixes the case. The editor is not out-guessing the model on letters but applying a phraseological constraint.

icorum lucili africani c larissimi v iri qui petunt ut ei permittatur in provincia afr ica regione beguensi territorio musulami??um ad casas nundinas iiii nonas novemb res et xii k alendas dec embres ex eo omnibus mensibus iiii non as et xii k alendas su

编者依据：No specific apparatus note; the edition supplies the genitive plural of the ethnic Musulamii, the tribe named as the territory-holder in this document.

读法：Proper-noun morphology. Aeneas returns or at 0.910. The name is common enough in African epigraphy that the model has the pattern; agreement here is cheap.

cormasa et conanam. neque tamen omnibu s huius rei ius erit sed procuratori principis optimi filioque eius usu dato ??que ad carra decem aut pro singulis carris mulorum trium aut pro singulis mulis asinorum binorum quibus eodem te

编者依据：No apparatus note at this point (the edition's listApp documents other lines, e.g. the engraver's omission at l. 5 and Frisch's correction at l. 7). Grounds: usque ad + accusative governs the following carra decem.

读法：Aeneas agrees at 0.477, but with a genuinely divided field: [ne] 0.233, [is] 0.117, [at] 0.104. The lowest-confidence agreement in the set, and a good illustration that a correct top-1 can still rest on a thin margin.

augusti caesaris patris patriae pontificis maximi tribuniciae potestatis xxv fili auguris consulis designati princip?? iuventutis patroni coloniae nostrae q. d. e. r. f. p. d. e. r. i. c. cum senatus populi romani inter c

编者依据：No dedicated apparatus note. The supplement completes the imperial title princeps iuventutis in the genitive, in agreement with the surrounding titulature of Lucius Caesar.

读法：Titulature is the most predictable material in Latin epigraphy. Aeneas returns is at 0.914 with the runner-up three orders of magnitude down (0.0025). Note the corpus status: the inscription is in the Aeneas training corpus but that corpus does NOT carry these restored letters, so the model is completing the title, not reciting it.

βασιλευς αντιοχος μητροφανει χαιρειν πεπ??καμεν λαοδικηι παννου κωμην και τημ βαριν και την προσουσαν χωραν τηι κωμηι ορος τηι τε ζελειτιδι χωραι και τηι κυζικηνηι και τηι

编者依据：The edition prints πεπ[ρά]καμεν, the perfect of πιπράσκω, "we have sold" - the operative verb of the whole dossier, which is a royal land-sale. The word is split across ll. 1/2 (lb break="no").

读法：DIVERGENCE, and the most instructive one. Ithaca prefers πεπ[οη]καμεν (0.884) over the editor's πεπ[ρά]καμεν (0.776) - a near-tie between two real Greek perfects. πεποιήκαμεν "we have made" is by far the commoner verb in the corpus; πεπράκαμεν is what the document is FOR. The editor is reading the legal act, the model is reading letter frequency. Note also that this text, restoration included, is in Ithaca's training corpus and the model still misses it.

ιλια ταλαντα λοιπον επτα ενενηκοντα λιτραι πεντε ικοσι διακοσια τρισχιλια μυρια ταλαντα ??μιαις εσοδος τεσσαρακοντα λιτραι οκτακισ χιλια δισμυρια ταλαντα εξοδος μια τρια κοντα λ

编者依据：No per-line apparatus note beyond the edition statement. The supplement yields ταμίαις, "to the treasurers", the dative plural that heads the account block.

读法：Ithaca returns τα at 0.912 with the runner-up at 0.0014. A financial document with heavy internal repetition gives the model exactly the redundancy it exploits best.

καρδαμαν χρηστος και αμεμπτος ετων ογδοηκοντα μηνας δε ??

编者依据：The EpiDoc encodes the whole sequence as <num value="10">: the editor reads δέ + [κα] as the single numeral δέκα, split across ll. 5/6 (lb break="no") with an ivy-leaf ornament standing INSIDE the word.

读法：DIVERGENCE, with a control that explains it. As fed, Ithaca ranks the editor 2nd (κα 0.066) behind ημ (0.127). The ivy-leaf ornament canonicalises to a space, so the model saw "μηνας δε ??" and read δέ as a separate particle. Re-fed WITHOUT that space ("μηνας δε??"), Ithaca returns κα at 0.712 - top-1, agreeing with the editor. The editor's advantage here is not linguistic but material: knowing that an ornament can sit inside a word, and that the layout does not mark a word boundary.

ανδρειας και καλοκαγαθιας και τας των προγονων αρετας δικαιον δε εστι τους αγαθους των ανδρων και ταν αυτων ε?????? ενδεικνυμενων τιμαν και προεδριαν τυνχανειν και αθανατον αυτων μναμαν παρα τοις ευ πα θοντ

编者依据：No per-line note; the edition rests on autopsy. The supplement gives εὔνοιαν, "goodwill", in the standard honorific collocation εὔνοιαν ἐνδείκνυσθαι.

读法：The longest supplement in the set (6 characters) and still a confident hit: Ithaca returns υνοιαν at 0.726, with υνοιας and υνοιαι further down - the model has the lexeme and is choosing among its endings. Decree formulae are the model's strongest ground.

ιος ουσα ευχα ριστουσα τω ειδιω αν δρι πολλας ευχαρισ τιας α ω ευομει?????

编者依据：No per-line note. The supplement completes the epithet εὐομείλητος, "affable", the last word of the epitaph; the preceding μεί is carried by a ligature (<hi rend="ligature">), which is why the editor can be sure of the stem.

读法：DIVERGENCE, and a total one: the editor's reading is absent from Ithaca's entire top-10, and the model's best guess is the non-word " και " at 0.009 - two orders of magnitude below its confidence on the other cases. The supplement sits at the very END of the text, so there is no right-hand context at all, and εὐομείλητος is a rare compound. When the model has nothing, its own confidence says so; this is the clearest case in the set of a bracket the editor fills from lexical knowledge the model does not have.