Pith. sign in

REVIEW 4 major objections 5 minor 90 references

An HTR-LLM Workflow for High-Accuracy Transcription and Analysis of Abbreviated Latin Court Hand

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A four-stage HTR-LLM workflow, grounded in a specialized handwriting-recognition baseline, transcribes abbreviated Latin court hand at 2-7% word error in four case studies.

desk verdict A useful, honest systems paper with a genuine Clean GT curation idea, but the 2-7% WER is not a firm number until the evaluation is made standard and reproducible. read the letter →

arxiv 2507.04132 v1 pith:HJCKMM4M submitted 2025-07-05 cs.DL cs.CLcs.CV

Joshua D. Isom This is my paper · ORCID
classification cs.DLcs.CLcs.CV
keywords HandwrittenTextRecognition(HTR)LargeLanguageModels(LLM)PaleographyDigitalHumanitiesTranscriptionWorkflowPost-CorrectionNamed-EntityCorrectionPleaRolls
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that neither a dedicated handwriting-recognition model nor a general-purpose multimodal language model alone is the right tool for transcribing the densely abbreviated Latin “court hand” of English plea rolls; the right tool is a fixed four-stage pipeline that uses each where it is strong. A specialized HTR model first produces a baseline transcription that keeps the abbreviations exactly as written on the page. Two large-language-model passes then correct that text against the original document image and expand the abbreviations into full scholarly Latin, and a final named-entity pass regularizes people and places while listing plausible alternatives for ambiguous readings. The paper validates this pipeline on four fifteenth-century cases, reporting word error rates of 2–7 percent against published scholarly editions. The intended payoff is that the mechanical labor of transcription is automated and a human reviewer is left with a short, targeted list of ambiguous and high-stakes readings.

What carries the argument

The load-bearing mechanism is the workflow as a whole, but its enabling invention is the “Clean Ground Truth” curation step that builds the Stage 1 HTR model. An older, noisier model transcribes a manuscript roll; the same annotated image chunk is sent twice to an LLM for multimodal correction; any line on which the two runs disagree, or on which the LLM departs substantially from the old HTR text, is discarded; the surviving high-confidence lines are used as reference transcriptions to train a new specialized HTR model. The remaining stages reuse the same image-plus-text grounding: Stage 2 corrects the HTR baseline against the image, Stage 3 expands abbreviations under explicit editorial rules, and Stage 4 performs named-entity correction and alternative generation. The named-entity report is what makes the system a human-in-the-loop tool rather than an unattended transcriber.

What would settle it

Recompute word error rate against the same printed editions with no forgiving rule: count every substituted, deleted, or inserted word, including names not chosen as the top reading, and include the full text in Case 4. If the strict percentages exceed the paper's reported 2–7 percent range on those same four cases, the advertised range is supported only under the relaxed scoring rule.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a fixed division of labor between a specialized handwriting-recognition model and a large language model (LLM) is the right design for the hardest historical documents. A dedicated HTR model first transcribes the abbreviated, abbreviation-preserving text; the LLM then corrects that text with the document image in front of it, expands the abbreviations into full scholarly Latin, and finally regularizes names while outputting a ranked list of plausible alternatives. The paper reports this pipeline reaching word error rates between roughly 2 and 7 percent across four fifteenth-century case studies when compared against printed scholarly editions under the paper's stated scoring rule, in which an entity error is forgiven when the correct reading appears among the generated alternatives.

Load-bearing premise

The load-bearing premise is that a word-error-rate score that forgives named-entity mistakes when the right name appears on the model's list of alternatives, and that compares only the overlapping portions of an abridged ground truth, is an honest measure of transcription accuracy.

Editorial extensions

If this is right

  • Researchers could process complete plea rolls rather than selected cases, enabling quantitative studies of litigation frequency, geography, and legal formulas.
  • Scholars without paleographic training could work with these sources, because the workflow converts transcription into verification of a short ambiguity report.
  • The expanded Latin output plus the named-entity alternatives gives editors a reproducible audit trail: every reading can be traced back to the HTR baseline and the LLM's alternatives.
  • The hybrid design should remain competitive as LLMs improve, since the specialized HTR model provides a stable, auditable baseline that reduces hallucination risk.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The “Clean Ground Truth” bootstrap is content-agnostic: the same two-run agreement filter could train specialized HTR models for other scripts and periods, since it only needs a noisy starting model and an LLM that can read the image.
  • One testable extension is to run the same four stages with the order of Stages 2 and 3 reversed, or with the named-entity report fed back into Stage 2; the paper does not test whether that would cut the residual semantic errors it lists.
  • The paper's own Case 4 observation—that an LLM expansion can be fuller than an abridged scholarly edition—suggests that future word-error benchmarks against editions should use full facsimile ground truths, otherwise a more faithful transcription is scored as an error.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a four-stage workflow for transcribing abbreviated Latin court-hand plea rolls: a specialized HTR model created through an LLM-driven 'Clean Ground Truth' curation process, multimodal LLM post-correction, scholarly expansion of abbreviations, and a final named-entity correction stage that also generates plausible alternative readings. The workflow is evaluated on four case studies against published scholarly ground truths, primarily J. H. Baker's edition of The Reports of John Caryll, with reported final Word Error Rates in the range of approximately 2-7%. The paper also argues that a hybrid HTR-LLM architecture is an enduring paradigm for historical document transcription and positions the system as a human-in-the-loop tool.

Significance. If the accuracy claims were fully substantiated, the workflow would be a practically useful contribution to digital paleography: it combines a publicly released Transkribus HTR model, detailed and reusable prompts, and a human-review report flagging ambiguous named entities. The paper is candid about failure modes such as semantic errors, hallucination from noisy input, and plausible-but-incorrect name substitutions. However, the headline 2-7% WER range is not yet supported by the evaluation as reported, because the metric is modified in ways that directly affect the numerator and denominator of the calculation, only two of four cases are presented with aligned error counts, and no Stage 1 baseline WER is given for the same test cases. The strength of the paper lies in the clarity of the pipeline description and the honest discussion of limitations, not in the quantitative validation as it currently stands.

major comments (4)
  1. [§4.2.1] The WER calculation in §4.2.1 uses a modified scoring rule under which a named-entity substitution is not penalized when the correct reading appears among the alternatives generated in the Stage 4 named-entity analysis. This changes the numerator of the reported 2.0% WER, and the effect on the reported number is not quantified. Because the abstract and conclusion present the 2-7% range without this qualification, the result is not comparable with standard WER benchmarks. The paper should report a strict WER alongside the modified metric, state the modification explicitly wherever the range is cited, and quantify how many errors were forgiven by this rule in each case.
  2. [§4.4.1] The Case 4 WER is computed 'only for the sections of text present in both' the model output and an abridged scholarly ground truth, yielding a 551-word reference. This excludes model-generated text that has no counterpart in the abridged GT from the denominator and counts insertions as errors even when the paper itself argues in §4.4.2 that the model's longer rendering may be a more faithful expansion of the manuscript. The resulting 6.2% is not a strict full-text WER against a complete reference, so it should not be pooled with Cases 1-3 to form the stated 2-7% range. The paper should either provide a complete aligned reference for Case 4 or report a strict WER over a defined, reproducible alignment.
  3. [§3.1, §4] No Stage 1 baseline WER is reported for the four validation cases; the only baseline performance given is a 5.25% CER on the model's validation set in §3.1. Without the Stage 1 WER on the same cases, the reader cannot determine whether Stages 2-4 actually reduce errors or merely reformulate them, and the contribution of each stage is not measurable. The paper should report per-stage WERs (Stages 1, 2, 3, and 4) on all four cases under the same metric and alignment protocol.
  4. [§4.1, §4.3] Cases 1 and 3 are reported only as approximate WERs (6.3% and 2.2%) with no aligned transcriptions, error counts, or specification of whether the modified or strict metric was used. These two cases constitute half of the evidence for the 2-7% range, so the numerical claims are not verifiable from the manuscript. Provide the full alignments and error counts for all four cases, or explicitly restrict the stated range to the cases with documented calculations.
minor comments (5)
  1. [Appendix A] The prompt text contains numerous development annotations such as '# ADDED Rule 5.1', '# CHANGED:', and 'CRITICAL REMINDER' that appear to be prompt-engineering notes; these should be removed or silently integrated into a clean prompt before publication.
  2. [Table 1] Table 1 is very wide and spans multiple pages, making it difficult to read in the current format; consider splitting it, rotating it, or moving it to supplementary material with a machine-readable version provided as a data file.
  3. [References [6]-[9]] The validation cases cite wiki pages authored by the author as the sources for the case-study ground truths; the paper should include the full alignments in an appendix or supplementary file so that the evaluation is self-contained and does not depend on retrievable but non-archival wiki pages.
  4. [Abstract and §1] The phrase 'ideal, four-stage workflow' is ambiguous; since the paper documents numerous failure modes and requires human review, the word 'ideal' should be clarified or replaced with a more neutral description such as 'proposed' or 'reference'.
  5. [§5] The argument in §5 that the hybrid paradigm will remain competitive 'for the foreseeable future' is speculative and not directly supported by the reported experiments; consider softening this to a discussion of design considerations supported by the evidence presented.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 2-7% WER range relies on a self-referential scoring rule: NE mismatches are forgiven when the Stage 4 LLM itself lists the ground-truth reading as an alternative.

  1. self definitional [Section 4.2.1, Quantitative Analysis (WER formula and modified scoring rule)]
    "For this analysis, a modified scoring rule is applied: an error in a named entity is not penalized if the correct reading was identified as a plausible alternative in the model's named-entity analysis step (e.g., 'Brome' was correctly chosen over the HTR's 'Grome'). This metric better reflects the utility of the workflow for a human reviewer, who would be alerted to the correct possibility."

    The WER numerator is defined relative to the Stage 4 model's own generated alternative list: a disagreement with the external scholarly ground truth is not counted as an error whenever the same model put the ground-truth reading in its candidate set. The reported accuracy therefore measures, in part, whether the model can include the target in its own output alternatives, not whether the final text matches the ground truth. The abstract's headline 'Word Error Rates (WER) in the range of 2-7%' inherits this rule; under a strict WER, those forgiven substitutions would increase the numerator and change the stated range.

full rationale

The validation is anchored to an external scholarly ground truth, primarily J. H. Baker's edition of The Reports of John Caryll ([5]), so the workflow itself is not trained and tested on the same LLM-generated labels; the Stage 0 'Clean GT' curation feeds the HTR model, but the end-to-end case studies are compared with a published edition. However, the central quantitative claim is not fully independent because Section 4.2.1 adopts a scoring rule that forgives named-entity errors when the Stage 4 model's own alternatives include the correct reading; that is a self-referential definition of 'error' and makes the WER range partly constructed rather than purely measured. Section 4.4.1 adds a second evaluation limitation: WER is computed 'only for the sections of text present in both' against an abridged ground truth, so model-generated text outside the abridged reference is excluded from the comparison; this weakens the 6.2% figure but is an incomplete-evaluation issue rather than a circular derivation. No load-bearing self-citation chain or imported uniqueness theorem was found; the author's own Wiki references ([6]-[9]) document case materials but the benchmark is Baker. A stricter, standard full-text WER with no NE forgiveness and no overlap-only denominator would be needed to substantiate the 2-7% claim independently.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The central results rest on unvalidated domain assumptions about LLM uncertainty proxies, transcription ground truth quality, and the effect of instructing the model to use standard legal formulas. The only explicit numeric free parameter is the >5% CER threshold in Stage 0; the prompt-specified expansion conventions shape the output but are not fitted to data.

free parameters (1)
  • CER uncertainty threshold for flagging 'unclear' lines = >5%
    Section 3.1 flags a line as unclear when the CER between two independent LLM runs exceeds approximately 5%; the threshold is hand-chosen and not justified by a calibration study.
assumptions (4)
  • domain assumption Baker's published scholarly edition is an accurate ground truth for the validation cases.
    Section 4 uses Baker's edition of The Reports of John Caryll as the reference for WER, assuming it is both accurate and appropriate for comparison with a diplomatic expansion of the manuscript.
  • domain assumption Output variance between two identical LLM calls is a reliable proxy for transcription uncertainty.
    Section 3.1 uses the CER between Run A and Run B to decide which lines are high-confidence enough to train the HTR model; this proxy is asserted without external validation.
  • domain assumption Training a Pylaia HTR model in Transkribus on the LLM-curated clean ground truth yields a specialized model that is better than the older general-purpose model.
    The paper reports CER 5.25% on a validation set for the new model but does not compare it against the older UCL-Toronto #7 model on the same validation set, so the improvement is not demonstrated.
  • domain assumption The multimodal LLM post-correction step is grounded by the image and does not systematically bias the text toward the prompt's expected formulas.
    The Stage 2 prompt instructs the LLM to prioritize the image, but the Stage 3 expansion prompt explicitly prescribes standard legal formulas, which could inflate agreement with editions that use those same formulas.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An HTR-LLM Workflow for High-Accuracy Transcription and Analysis of Abbreviated Latin Court Hand." pith.science (2026). https://pith.science/paper/HJCKMM4M

@misc{pith2026250704132,
  author       = {Pith},
  title        = {Pith review of: An HTR-LLM Workflow for High-Accuracy Transcription and Analysis of Abbreviated Latin Court Hand},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJCKMM4M}},
  note         = {Machine review of arXiv:2507.04132}
}
read the original abstract

This article presents and validates an ideal, four-stage workflow for the high-accuracy transcription and analysis of challenging medieval legal documents. The process begins with a specialized Handwritten Text Recognition (HTR) model, itself created using a novel "Clean Ground Truth" curation method where a Large Language Model (LLM) refines the training data. This HTR model provides a robust baseline transcription (Stage 1). In Stage 2, this baseline is fed, along with the original document image, to an LLM for multimodal post-correction, grounding the LLM's analysis and improving accuracy. The corrected, abbreviated text is then expanded into full, scholarly Latin using a prompt-guided LLM (Stage 3). A final LLM pass performs Named-Entity Correction (NEC), regularizing proper nouns and generating plausible alternatives for ambiguous readings (Stage 4). We validate this workflow through detailed case studies, achieving Word Error Rates (WER) in the range of 2-7% against scholarly ground truths. The results demonstrate that this hybrid, multi-stage approach effectively automates the most laborious aspects of transcription while producing a high-quality, analyzable output, representing a powerful and practical solution for the current technological landscape.

Figures

Figures reproduced from arXiv: 2507.04132 by the authors.

Figure 1
Figure 1. The Four-Stage HTR-LLM Workflow. The process begins with a one-time creation of a [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

90 extracted references · 77 canonical work pages

  1. [1]

    Unlocking the Archives: Using Large Language Models to Transcribe Handwritten Historical Documents

    M. Humphries, L. C. Leddy, Q. Downton, et al. Unlocking the archives: Large language models achievestate-of-the-artperformanceonthetranscriptionofhandwrittenhistoricaldocuments. arXiv preprint arXiv:2411.03340, 2024

    work page Pith review arXiv 2024

  2. [2]

    S. Kim, J. Baudru, W. Ryckbosch, et al. Early evidence of how llms outperform traditional systems on ocr/htr tasks for historical records.arXiv preprint arXiv:2501.11623, 2025

    arXiv 2025

  3. [3]

    Greif, N

    G. Greif, N. Griesshaber, and R. Greif. Multimodal llms for ocr, ocr post-correction, and named entity recognition in historical documents.arXiv preprint arXiv:2504.00414, 2025

    arXiv 2025

  4. [4]

    Crosilla, L

    G. Crosilla, L. Klic, and G. Colavizza. Benchmarking large language models for handwritten text recognition. arXiv preprint arXiv:2503.15195, 2025

    arXiv 2025

  5. [5]

    J. H. (ed.) Baker.The Reports of John Caryll (1485–1522), Part I, volume 115. Selden Society,

  6. [6]

    J. D. Isom. AI validation case: Cp40/922 m. 308. Wiki of the Anglo-American Legal Tradition, 2025.Retrieved5July2025,from https://waalt.uh.edu/index.php/AI_validation_case_ CP40/922_m._308

    2025

  7. [7]

    J. D. Isom. AI validation case: Cp40/906 m. 524. Wiki of the Anglo-American Legal Tradition, 2025.Retrieved5July2025,from https://waalt.uh.edu/index.php/AI_validation_case_ CP40/906_m._524

    2025

  8. [8]

    J. D. Isom. AI validation case: Kb27/915 m. 27. Wiki of the Anglo-American Legal Tradition, 2025.Retrieved5July2025,from https://waalt.uh.edu/index.php/AI_validation_case_ KB27/915_m._27

    2025

Show all 90 references
  1. [9]

    J. D. Isom. AI validation case: Kb27/925 m. 87d. Wiki of the Anglo-American Legal Tradition, 2025.Retrieved5July2025,from https://waalt.uh.edu/index.php/AI_validation_case_ KB27/925_m._87d. A System Prompt for LLM Post-Correction of HTR YOUR ROLE: Medieval Latin Paleography Ex...

    2025

  2. [10]

    Place it immediately after the last written letter before the omission/mark

    **Apostrophe Use:** Use a single straight apostrophe ( ') ONLY to represent a visible abbreviation mark (macron, hook, superscript, symbol) or clearly omitted letters *seen on the image*. Place it immediately after the last written letter before the omission/mark. Use universa...

  3. [11]

    Document Image (Absolute authority for form/content), 2

    **Source Priority:** Use sources in this order: 1. Document Image (Absolute authority for form/content), 2. HTR Transcription (Word ID aid ONLY - ignore its expansion status), 3. Named Entity List (Reference), 4. Your Knowledge (Lowest). **# ADDED: Always prioritize the Docume...

  4. [12]

    **HTR Use:** Consult HTR only for word identification clues; its expansion status is irrelevant for your 'abbreviated_latin_lines'. --- **B. CORE PRINCIPLE: IMAGE AUTHORITY & **ABSOLUTE LINE INTEGRITY** (Non-Negotiable)**

  5. [13]

    **Image Authority:** The Document Image dictates text content, word order, line breaks, and abbreviation forms

  6. [14]

    n"' matches the *first* word visually associated with the baseline marked by the red number 'n' on the image. Ensure the *last* word transcribed for JSON key '

    **Line Mapping & Structure (CRITICAL - Non-Negotiable):** * **Anchor to Red Numbers:** Your primary task is to map text directly to the **red numbers** visible on the image. Each red number (e.g., '10') is located to the left of and slightly above the start of the specific bas...

  7. [15]

    lines 10 to 19

    **Transcription Scope (CRITICAL):** * **Transcribe ONLY Requested Numbered Lines:** Transcribe ONLY the text visibly associated with (typically above) the specific **numbered baselines requested for this chunk**. The requested line numbers are specified in the User Prompt (e.g...

  8. [16]

    **Empty Lines:** If no text corresponds to a specific requested baseline 'n' (identified by its red number 'n'), output '""' for JSON key 'n'

  9. [17]

    **Line Count:** The number of lines (keys) in the output JSON MUST exactly match the number of lines requested for the chunk. --- **C. ABBREVIATION HANDLING ( 'abbreviated_latin_lines' - Visual Representation)**↩→

  10. [18]

    **Goal:** The 'abbreviated_latin_lines' MUST visually mirror the text on the corresponding image line (identified by its red number), including all abbreviations *exactly as written*

  11. [19]

    Transcribe fully spelled-out words as seen (e.g., 'quod', 'predictus', 'domino')

    **ABSOLUTELY NO EXPANSION (CRITICAL - Non-Negotiable):** Transcribe abbreviations *exactly* as seen on the image (e.g., 'q'd', 'p'd'co', 'd'no'). Transcribe fully spelled-out words as seen (e.g., 'quod', 'predictus', 'domino'). **NEVER** expand abbreviations shown on the image...

  12. [20]

    ** 'd'ni'/'d'no':** Transcribe 'dominus/i/o' abbreviation as 'd'ni' or 'd'no'

  13. [21]

    ** '&' (CRITICAL):** Transcribe Tironian 'et' as '&'

  14. [22]

    20 # ADDED: Explicit negative constraint for & / Et **CRITICAL REMINDER:** **NEVER** transcribe a spelled-out 'Et' or 'et' on the image as '&'

    ** 'et' / 'Et' (CRITICAL):** Transcribe spelled-out 'et' or 'Et' exactly as written (matching case). 20 # ADDED: Explicit negative constraint for & / Et **CRITICAL REMINDER:** **NEVER** transcribe a spelled-out 'Et' or 'et' on the image as '&'. **NEVER** transcribe a Tironian ...

  15. [23]

    ** '&c':** Transcribe 'etcetera' abbreviation (e.g., '&c', '&c.') as '&c'

  16. [24]

    ** 'q'd':** Transcribe 'quod' abbreviation (e.g., 'q' w/ macron or similar mark) as 'q'd'

  17. [25]

    ** 'quod':** Transcribe spelled-out 'quod' as 'quod'

  18. [26]

    ** 'p'':** Transcribe 'per'/'pro' abbreviation (crossed-p) as 'p''

  19. [27]

    ** 'per'/'pro':** Transcribe spelled-out 'per' or 'pro' as written

  20. [28]

    ** 'com'':** Transcribe 'com'/'con' prefix abbreviation (macron/hook) as 'com''

  21. [29]

    ** 'comitatus':** Transcribe spelled-out 'comitatus', 'comiti', etc., as written

  22. [30]

    ** 'Angl'':** Transcribe 'Anglie' abbreviation as 'Angl''

  23. [31]

    ** 'Dominus':** Transcribe spelled-out 'Dominus', 'domini', 'domino' as written (respect case)

  24. [32]

    ** 'Ioh'es'/'Ioh'em':** Transcribe 'Johannes/em' abbreviation as 'Ioh'es' or 'Ioh'em'

  25. [33]

    ** 'Iohannes':** Transcribe spelled-out 'Iohannes', 'Iohannem' as written

  26. [34]

    ** 'vic'':** Transcribe 'vicecomes/ti' abbreviation as 'vic''

  27. [35]

    ** 'vicecomes':** Transcribe spelled-out 'vicecomes', 'vicecomiti' as written

  28. [36]

    ** 'p'd'':** Transcribe 'predictus' forms abbreviation (e.g., 'p'dcus', 'p'dco') as 'p'd'cus', 'p'd'co', 'p'd'ca', etc

  29. [37]

    ** 'predictus':** Transcribe spelled-out 'predictus', 'predicto', etc., as written

  30. [38]

    ** 'Reg'':** Transcribe 'Regis/Rege/Regi' abbreviation as 'Reg''

  31. [39]

    ** 'Regis'/'Rege'/'Regi':** Transcribe spelled-out 'Regis', 'Rege', 'Regi' as written (capitalized)

  32. [40]

    ** 'sup'':** Transcribe 'super' abbreviation as 'sup''

  33. [41]

    ** 'Anglie':** Transcribe spelled-out 'Anglie' as written

  34. [42]

    ** 'Westm'':** Transcribe 'Westmonasterium' abbreviation as 'Westm''

  35. [43]

    ** 'Westmonasterium':** Transcribe spelled-out 'Westmonasterium' as written

  36. [44]

    ** 'attorn'':** Transcribe 'attornatus/um' abbreviation as 'attorn''

  37. [45]

    ** 'attornatus':** Transcribe spelled-out 'attornatus', 'attornatum' as written

  38. [46]

    ** 'nup'':** Transcribe 'nuper' abbreviation as 'nup''

  39. [47]

    ** 'nuper':** Transcribe spelled-out 'nuper' as written

  40. [48]

    ** 'ap'd':** Transcribe 'apud' abbreviation as 'ap'd'

  41. [49]

    ** 'apud':** Transcribe spelled-out 'apud' as written

  42. [50]

    ** 'scilicet':** Transcribe spelled-out 'scilicet' as 'scilicet'

  43. [51]

    ** 'super':** Transcribe spelled-out 'super' as written

  44. [52]

    ** 'saluo'':** Transcribe 'saluo' with an abbreviation mark as 'saluo''

  45. [53]

    ** 'saluo':** Transcribe spelled-out 'saluo' as 'saluo'

  46. [54]

    ** 'pl'ito':** Transcribe 'placito' abbreviation as 'pl'ito'

  47. [55]

    ** 'placito':** Transcribe spelled-out 'placito' as 'placito'

  48. [56]

    ** 'transgr'':** Transcribe 'transgressio...' abbreviation as 'transgr'' (or 't'nsgr'' if seen). 21

  49. [57]

    ** 'ib'm':** Transcribe 'ibidem' abbreviation as 'ib'm'

  50. [58]

    ** 'ibidem':** Transcribe spelled-out 'ibidem' as 'ibidem'

  51. [59]

    ** 'scil''/'sc'l't':** Transcribe 'scilicet' abbreviation as 'scil'' or 'sc'l't' (matching image)

  52. [60]

    **Long S:** Transcribe long 's' as standard 's'

  53. [61]

    ** 'Will'm':** Transcribe 'Willelmum/us' abbreviation as 'Will'm' (or specific form seen)

  54. [62]

    ** 'Willelmus':** Transcribe spelled-out 'Willelmus', 'Willelmum' as written

  55. [63]

    ** 'Ric'':** Transcribe 'Ricardus/um' abbreviation as 'Ric'' (or specific form seen)

  56. [64]

    ** 'Ricardus':** Transcribe spelled-out 'Ricardus', 'Ricardum' as written

  57. [65]

    ** 'Thom'':** Transcribe 'Thomas/am/e' abbreviation as 'Thom'' (or specific form seen)

  58. [66]

    ** 'Thomas':** Transcribe spelled-out 'Thomas', 'Thomam', 'Thome' as written

  59. [67]

    ** 'Rob't':** Transcribe 'Robertus/um' abbreviation as 'Rob't' (or specific form seen)

  60. [68]

    *(Note on Names: Pay close attention to whether names are abbreviated or fully spelled out on the image line and transcribe accordingly, following Rule C.9.)*↩→ --- **E

    ** 'Robertus':** Transcribe spelled-out 'Robertus', 'Robertum' as written. *(Note on Names: Pay close attention to whether names are abbreviated or fully spelled out on the image line and transcribe accordingly, following Rule C.9.)*↩→ --- **E. REFERENCE EXAMPLES (Transcribe O...

  61. [69]

    This applies to all words, including 'uersus', 'uilla', 'ualenciam', 'iudicium', 'iuratus', etc

    **CRITICAL LETTERFORMS (U/V & I/J):** **ALWAYS** use only 'u' or 'U' (never 'v' or 'V') and only 'i' or 'I' (never 'j' or 'J'). This applies to all words, including 'uersus', 'uilla', 'ualenciam', 'iudicium', 'iuratus', etc

  62. [70]

    **Structure Analysis:** Note paragraphs and line counts using the **red image numbers** and Named Entity List for context. Use the Named Entity List as a reference for potential names/places, but **# ADDED: always prioritize the spelling, capitalization, and form seen on the i...

  63. [71]

    **# ADDED: Be particularly careful distinguishing similar letter forms like 'a'/'o', 'e'/'o', 'D'/'S', 'G'/'B'

    **Differentiation:** Carefully distinguish minims (n/u, m/in/ni/iu/ui), c/t, and f/s based on the image. **# ADDED: Be particularly careful distinguishing similar letter forms like 'a'/'o', 'e'/'o', 'D'/'S', 'G'/'B'. Always verify against the image.**

  64. [72]

    **Standard Letters:** Transcribe other letters to standard modern equivalents based on image form

  65. [73]

    ** 'uersus'/'sicut':** Transcribe 'uersus'/'u'sus' and 'sicut'/'sic'' based on image form (abbreviated or full). --- **G. CAPITALIZATION (Based on Image)**

  66. [74]

    * *Example:* If the image shows 'predictus', transcribe 'predictus', NOT 'Predictus'

    **Strict Following (CRITICAL):** Follow manuscript capitalization **EXACTLY** for **ALL** words. * *Example:* If the image shows 'predictus', transcribe 'predictus', NOT 'Predictus'. If the image shows 'Comes', transcribe 'Comes', NOT 'comes'. If the image shows 'die', transcr...

  67. [75]

    ** 'Rex'/'Dominus'/'Regis':** Capitalize 'Rex', 'Dominus', 'Regis', 'Rege', 'Regi' if spelled out on image. 23

  68. [76]

    Retain abbreviation marks if present

    **Nouns/Titles:** Capitalize proper nouns (people, places like 'London'', 'Pasche'), titles/occupations ( 'Gentilman', 'yoman', 'clericus') *only if capitalized on the image*. Retain abbreviation marks if present. --- **H. NUMERALS & SYMBOLS**

  69. [77]

    **Roman Numerals:** Transcribe Roman numerals exactly as they appear (e.g., 'xij', 'xv', 'iiij'); do NOT convert to Arabic

  70. [78]

    **Paragraph Mark:** Transcribe the paragraph mark symbol as '¶' if present. --- **I. WORKFLOW & VERIFICATION**

  71. [79]

    **Image Focus:** Examine the image, identifying the **red numbers** marking the start of each requested baseline

  72. [81]

    corresponding to original document lines N, N+1, ...): * Identify the *last word* written in the HTR transcription for that line (use HTR only as a guide here)

    **PRELIMINARY STEP (Internal Grounding - DO NOT OUTPUT THIS): LAST WORD FOCUS** * For each **requested line number** (e.g., 10, 11, ... corresponding to original document lines N, N+1, ...): * Identify the *last word* written in the HTR transcription for that line (use HTR onl...

  73. [82]

    * Transcribe **ALL** text visually associated with the baseline marked by **red number 'n'** into JSON key '"n"'

    **Transcribe 'abbreviated_latin_lines' (FINAL OUTPUT):** * Go line-by-line according to the **red baseline numbers** requested for the chunk ( 'n'). * Transcribe **ALL** text visually associated with the baseline marked by **red number 'n'** into JSON key '"n"'. **Do not omit ...

  74. [83]

    n"' match the *first* word on the image baseline marked 'n'? Does the *last* word of JSON key '

    **Verify Output:** Cross-reference the generated 'abbreviated_latin_lines' JSON against the **numbered image lines**. Check: * **Line Content:** Does the text in JSON key '"n"' match the image text associated with the baseline marked by **red number 'n'**? * **Completeness (CR...

  75. [84]

    **Line Integrity:** Anchor transcription to **visible red baseline numbers** and ensure **NO words are moved between lines**

  76. [85]

    **DO NOT OMIT** any words or symbols

    **Completeness:** Transcribe **ALL** visible text associated with the numbered baseline. **DO NOT OMIT** any words or symbols

  77. [86]

    **No Expansion/Abbreviation:** Transcribe abbreviations and spelled-out words **exactly as they appear** on the image line (e.g., 'Et' stays 'Et', '&' stays '&', 'Will'm' stays 'Will'm', 'Willelmum' stays 'Willelmum')

  78. [87]

    **NEVER** convert spelled-out 'Et'/'et' to '&', or '&' to 'Et'/'et'

    **Exact '&' vs 'Et':** Follow the image precisely. **NEVER** convert spelled-out 'Et'/'et' to '&', or '&' to 'Et'/'et'

  79. [88]

    **Critical Abbreviations:** Follow the specific transcription rules for the critical forms listed in Section D

  80. [89]

    **Exact Capitalization:** Match manuscript capitalization **exactly** for 25 **ALL** words

  81. [90]

    aforesaid

    **Requested Lines Only:** Transcribe only the lines requested for the chunk. B LLM Prompts for Case Study Analysis B.1 LLM Expansion Prompt You are an expert in Latin paleography, specializing in the transcription of 15th-century English legal records. Your task is to produce ...

  82. [1999]

    Available at:https://archive.org/details/reportsofcasesby0000cary/. 17

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.