PaperlayerView exact revision on arXiv

Source-linked AI summary

OCR-Based Field Extraction for Archaeological Pottery Metadata: The CENTURIA Dataset

Gissu Valentina Naghavi, Dominik Hagmann, Martin Kampel, Irene Ballester

arXiv:2608.30616v1cs.CV

TL;DR

Handwritten archaeological pottery annotations carry information needed for comparison, dating, and provenance, but remain difficult to analyze computationally. This paper introduces CENTURIA and benchmarks OCR models, finding that fine-tuning on 57 samples brings transcription error below 1.5% and overall field accuracy above 87%.

  • Problem

    Transferability of text-recognition and extraction methods to handwritten archaeological records with scattered annotations and no canonical reading order remains unaddressed, despite their importance for cross-site comparison, chronology, and provenance.

    Source: §1 · p. 2, §2 · p. 4

  • Method

    The authors introduce CENTURIA, a benchmark of 507 expert-validated handwritten archaeological records, and systematically evaluate five OCR models for transcription and structured field extraction.

    Source: §1 · p. 2, §7 · p. 15

  • Results

    On CENTURIA, zero-shot models yield 15–32% SpACER-M, while LoRA fine-tuning on 57 samples lowers transcription error below 1.5% and raises overall field accuracy above 87%.

    Source: §1 · p. 2, §7 · p. 15

  • Takeaways & Limitations

    The benchmark and evaluation provide heritage practitioners evidence to inform decisions about deploying these tools on their own archives.

    Source: §1 · p. 2

    Agreement-based reliability estimates apply to selected subsets, and shared training data and rules mean pipeline agreement may overestimate reliability.

    Source: §6 · p. 14

Abstract

from arXiv · show

Pottery is a primary source for reconstructing the chronological and economic dimensions of past societies. Archaeologists often document ceramic finds through technical drawings and handwritten metadata. This metadata is critical for dating, provenance attribution, and cross-site comparison, but remains inaccessible to computational analysis, requiring manual transcription of every record. We investigate whether state-of-the-art document analysis models can address this task, and introduce CENTURIA, a dataset of 507 pottery records from the Roman site of Carnuntum, providing transcriptions, bounding boxes, and structured field-level labels across seven metadata categories. Benchmarking five OCR models reveals a substantial domain gap: zero-shot transcription error reaches 15-32% SpACER-M, far exceeding rates on printed archival documents, with domain-specific fields recovered in fewer than 3% of cases. LoRA fine-tuning on just 57 samples, reflecting a realistic archival annotation budget, closes this gap, reducing transcription error to below 1.5% and recovering overall field-level accuracy above 87%. Our results show that a small expert-validated fine-tuning set suffices to convert handwritten pottery documentation into structured, searchable metadata ready for archaeological databases.

1 Introduction

The paper introduces CENTURIA to evaluate machine extraction of structured metadata from handwritten archaeological pottery records, showing that limited LoRA fine-tuning substantially improves transcription and field extraction and that pipeline agreement can reduce manual review.

1 Introduction

  • Pottery drawings preserve archaeological records, but their handwritten context, typology, and inventory annotations remain inaccessible to computational analysis without labor-intensive transcription.Existing HTR benchmarks do not reflect scattered annotations, faded ink, or dense domain-specific abbreviations, while archive-scale analysis also requires mapping text to structured fields.Source: §1 · p. 2
  • Zero-shot HTR transcription errors of 15–32% fell below 1.5% after LoRA fine-tuning on 57 samples, with overall field-level accuracy above 87%.The 57-sample fine-tuning set represents a constrained archival annotation setting.Source: §1 · p. 2
  • CENTURIA is the first benchmark for HTR and field extraction of handwritten archaeological records with scattered annotations and technical terminology, comprising 507 expert-validated samples.Source: §1 · p. 2
  • Evaluation across five OCR models, three adaptation strategies, and two field extraction methods found transcription quality to be the primary determinant of field-level accuracy.The evaluation assesses transcription quality, per-field performance, and error modes.Source: §1 · p. 2
  • Agreement-based review auto-accepts 56.4% of scans at 95.5% field precision, rising to 98.7% on fields confirmed by all four pipelines.This heuristic concentrates manual review on unresolved conflicts.Source: §1 · p. 3

2 Related Work

Prior archaeological digitisation addresses artefact imagery and printed records, but not handwritten annotations; existing recognition and extraction methods and benchmarks likewise leave handwritten archaeological documentation unassessed. CENTURIA addresses this gap with a systematic benchmark for transcription and field-level extraction.

Document Analysis in Archaeological Research

  • PyPotteryScan requires manually annotated bounding boxes and lacks a published accuracy evaluation; CENTURIA provides a fully automatic pipeline and the first systematic evaluation for handwritten pottery documentation.The authors use PyPotteryScan for preliminary transcriptions, but manually correct all documents to create ground truth.Source: §2 · p. 3

Text Recognition Models and Field Extraction

  • HTR spans line-level systems to full-page models, and small domain-specific fine-tuning can reduce errors, but KIE transfer to scattered handwritten archaeological records remains unassessed.Existing KIE methods map text to schemas, but datasets assume printed business documents with standard layouts.Source: §2 · pp. 3–4

Benchmarks for HTR in Historical and Cultural Heritage Documents

  • CENTURIA is the first benchmark for archaeological documentation, evaluating transcription and field-level KIE across model comparison, adaptation strategies, and extraction methods.Existing heritage benchmarks cover marriage records, index cards, census tables, and medieval transcription, but not archaeological field documentation with domain-specific abbreviations and scattered annotations.Source: §2 · p. 4

3 CENTURIA Dataset

CENTURIA is a 507-document dataset of annotated pottery records from Carnuntum, created to support reproducible OCR fine-tuning and structured field extraction. Its records combine manual text and visual annotations with expert-checked, seven-category metadata.

3 CENTURIA Dataset

  • 507 Carnuntum pottery-record scans have bounding boxes, manually transcribed text, structured field values, and a proportional stratified train/test split of 57 and 450 documents.The split samples across six campaigns to support reproducible evaluation and fine-tuning.Source: §3 · pp. 4–5

3.1 Data Collection

  • CENTURIA samples 507 documents from an approximately 70,000-record Carnuntum archive across six campaigns from 1976–2017, capturing handwriting, scribal, condition, and annotation variation.The stratified sample also includes crossed-out text and overlapping annotations.Source: §3.1 · p. 5

Transcription Annotation.

  • All 507 transcriptions were manually corrected after PyPotteryScan, built on olmOCR2, produced preliminary text; correction averaged 70 seconds per record.Experts resolved readings that could not be determined from scans alone in 48 cases (9.47 %).Source: Transcription Annotation · p. 5

Field Annotation.

  • Transcriptions map to seven optional semantic categories, with regexes and controlled vocabularies assigning spans and a site specialist verifying annotations.Corrections were needed in 3.75 % of documents, involving edge cases such as crossed-out entries, field-boundary errors, and ambiguous plate references.Source: Field Annotation · pp. 5–6, Table 2 · p. 6

4 Method

The method converts a scanned pottery record into structured metadata through text recognition followed by field extraction. OCR models transcribe each page, and their output is normalized before extraction.

4 Method

  • The method uses two stages—text recognition and field extraction—to produce structured metadata from a scanned pottery record.The stages are shown in Fig. 3.Source: §4 · p. 6

4.1 Handwritten Text Recognition

  • Each OCR model transcribes a single-page scan into a character sequence, then post-processing normalizes formatting artefacts before field extraction.The post-processing steps are described in Appendix B.Source: §4.1 · p. 6

Baseline Models.

  • The baseline models span detection-recognition, full-page VLM, generalist multi-task, and reinforcement-learning-optimised paradigms, and are all open-source.The set comprises TrOCR, olmOCR, olmOCR2, Florence2, and LightOnOCR.Source: Baseline Models · p. 6

LoRA fine-tuning*

The method transcribes pottery records using multiple OCR architectures and adaptation strategies, then maps the resulting text into a seven-category metadata schema using rule-based or LLM-based extraction.

LoRA fine-tuning*

  • Recognition compares segmented TrOCR with full-page VLMs, including olmOCR, olmOCR2, Florence2, and lightweight LightOnOCR.TrOCR uses CRAFT line segmentation via EasyOCR, processing each detected region independently. LightOnOCR has 1B parameters, roughly an order of magnitude fewer than the strongest full-page VLMs on olmOCR-Bench.Source: Baseline Models · pp. 7–8

Domain Adaptation Strategies.

  • Adaptation tests zero-shot instructions, few-shot examples with k ∈ {1, 5, 8}, and LoRA fine-tuned on 57 CENTURIA training samples.LoRA uses low-rank weight decomposition applied to all linear projection layers; prompting conditions models without weight updates, while LoRA updates model weights.Source: Domain Adaptation Strategies · p. 8

4.2 Field Extraction

  • Field extraction maps transcriptions to seven semantic categories, comparing exact-pattern REGEX extraction with prompted Qwen3-8B [76] LLM extraction.REGEX shares the ground-truth rules and fails only on OCR noise, isolating noise propagation; the LLM uses an 8-shot prompt with corrected transcription examples and assigns null to absent or unrecognisable fields.Source: §4.2 · pp. 8–9

5 Evaluation

Evaluation shows that zero-shot OCR struggles with handwritten pottery records, while fine-tuning and cross-pipeline agreement enable accurate field recovery and selective automation. It benchmarks five models for transcription and examines adaptation and extraction using the two strongest baselines.

Transcription. / Field-Level Transcription.

  • SpACER accommodates scattered annotations rather than assuming linear reading order; field transcription uses EMR and ANLS, while extraction scores normalized field matches and macro-averages overall accuracy.Models with bounding boxes use IoU-matched SpACER-m, whereas models without spatial outputs use page-wide SpACER-M; EMR requires exact matches, and ANLS tolerates minor OCR errors.Source: Transcription · p. 9, Field-Level Transcription · p. 10

Baseline Performance.

  • Zero-shot models scored SpACER-M 15.58–32.35%; olmOCR2 achieved 30.55% EMR and 57.43% ANLS, while the fastest models ran under 1 s/doc.TrOCR and Florence2 were among the least accurate, and olmOCR was least efficient at 12.2 s/doc without a corresponding accuracy advantage; subsequent adaptation focused on olmOCR2 and LightOnOCR.Source: Table 3 · p. 10, Baseline Performance · p. 10

Domain Adaptation.

Field-level Performance.

  • Baseline EMR was 65–80% for measurements and 26–37% for pottery form, while Artefact Category and Publication Type were near zero; LoRA raised overall EMR above 89%.OCR noise corrupted abbreviations and domain tokens, whereas numeric measurements were visually robust.Source: Figure 5 · p. 11, Field-level Performance · p. 12

Domain Adaptation.

  • Zero-shot prompting improved olmOCR2 SpACER-M from 15.58% to 6.77%, but few-shot prompting drove LightOnOCR to 100% SpACER-M at 5 shots; LoRA outperformed prompting.At 5 shots, olmOCR2 reached 5.38% SpACER-M; LightOnOCR’s repetitive output loops are consistent with known repetition failure modes [38].Source: Table 4 · p. 11, Domain Adaptation · p. 11

Extraction Accuracy.

  • Fine-tuned olmOCR2 reached 91.73% overall extraction accuracy under REGEX versus 88.30% for LightOnOCR; baseline extraction was only 26–30% across models and methods.On ground-truth transcriptions, REGEX reached 99.53% and the LLM 96.76%, showing a 2.77-point gap independent of OCR noise; on OCR output, the extractors differed by less than one overall point.Source: Table 5 · p. 12, Extraction Accuracy · p. 12

Error Modes.

  • Publication Type had the highest incorrect rate, while short or domain-specific tokens were omitted more under REGEX and shifted toward incorrect predictions under LLM extraction.For LightOnOCR (+) Surface Treatment, omissions were about 41% under REGEX versus about 37% incorrect predictions under LLM; 60% of REGEX errors and 71% of LLM errors were shared across models.Source: Figure 6 · p. 13, Error Modes · p. 13

Confidence Estimation.

  • All-four-pipeline agreement was correct in 98.7% of cases; majority voting with field-specific rules auto-accepted 56.4% of documents at 95.5% overall precision.This approach concentrates manual review on scans where disagreement persists.Source: Confidence Estimation · pp. 13–14

6 Limitations and Future Work

The pipeline still has systematic transcription and extraction errors, while reliability estimates and cross-archive transfer remain conditional or untested. Future work will formalize human review, evaluate archival effort, and extend CENTURIA where access permits.

Remaining failure modes

  • Short domain-specific tokens, digit-level errors, unresolved percentage signs, and unfiltered struck-through text remain failure modes after fine-tuning.These errors affect REGEX extraction and both pipelines, and can introduce noise into downstream field extraction.Source: §6 · p. 14

Quality validation and human evaluation

  • 98.7 % of fields confirmed by all four pipelines are correct, while 56.4 % of documents auto-accepted reach 95.5 % precision.Both figures apply to agreement-selected subsets, and shared training data and rules mean pipeline agreement may overestimate reliability; future work will formalize human-in-the-loop review and evaluate time savings and correction effort with domain experts.Source: §6 · p. 14

Transferability and scope

  • CENTURIA spans six campaigns, but cross-archive generalisation depends on recording conventions, script, and language, and the minimum annotation effort for other archives remains unknown.The 57-record training split reflects a realistic archival annotation budget; cross-archive evaluation is constrained by collection access, and future work aims to extend CENTURIA and aggregate annotations toward a foundational OCR model.Source: §6 · pp. 14–15

7 Conclusion · Appendix

CENTURIA enables structured analysis of handwritten pottery metadata: LoRA fine-tuning on 57 samples sharply improves transcription and field recovery, while cross-pipeline agreement can reduce manual review. The appendix documents the data, experimental procedures, analyses, validation, and prompts supporting reproducibility.

7 Conclusion

  • Below 1.5% transcription error and above 87% overall field accuracy follow LoRA fine-tuning on 57 samples, versus 15–32% SpACER-M zero-shot transcription error.The conclusion reports that either REGEX or LLM-based extraction yields structured metadata after transcription quality improves.Source: §7 · p. 15
  • CENTURIA comprises 507 annotated handwritten pottery records from Carnuntum and benchmarks five OCR models on archaeological HTR and structured field extraction.The conclusion characterizes transcription quality, rather than extraction method, as governing end-to-end performance.Source: §7 · p. 15
  • LightOnOCR (+) achieves comparable overall accuracy to olmOCR2 (+) at substantially lower computational cost.Source: §7 · p. 15
  • 56.4% of documents can be auto-accepted at 95.5% field precision through cross-pipeline agreement, rising to 98.7% precision when all four pipelines agree.The procedure concentrates manual effort on genuine disagreements.Source: §7 · p. 15
  • A small expert-validated annotation effort makes archive-scale digitisation of Carnuntum’s handwritten pottery documentation feasible and its metadata searchable.Source: §7 · p. 15

Appendix

  • Sections A and B document CENTURIA sample selection, annotation, data format, inference setup, model checkpoints, post-processing, and LoRA fine-tuning hyperparameters.Source: Appendix · p. 16
  • Section C provides full field-level OCR results, the complete adaptation strategy comparison, qualitative OCR examples, and fine-tuned inference speed.Source: Appendix · p. 16
  • Sections D and E describe the Model Confidence and Smart Merge procedure with validation results and collect all prompts, respectively.Source: Appendix · p. 16

A Dataset Details … A.3 Data Format and Structure

The dataset samples varied handwritten pottery records from Carnuntum and defines consistent transcription, bounding-box, and semantic-field annotations. It is distributed in two JSON files linking scan text and regions with structured field labels.

A.1 Sample Selection Strategy

  • The sample spans six Carnuntum campaigns from 1976 to 2017 and captures variation in handwriting, character types, terminology, abbreviations, and annotation complexity.Eight sampling strata were selected across the campaigns.Source: §A.1 · p. 16

Excavation campaigns

  • The six excavation strata contain 43, 118, 121, 5, 5, and 66 entries, respectively, across Gstettenbreite, Palastruine, and Petroneller Burg campaigns.The cited campaign markers are Gstettenbreite,, Palastruine, and Petroneller Burg.Source: Excavation campaigns · p. 16

Field walking campaigns

  • The two field-walking strata contain 110 Gstettenbreite 2017 entries and 39 Käsemacherweide 2012 entries, while campaigns vary in identifier and measurement conventions.Conventions include find identifiers such as fn, quadrant, and kiste, and measurements recorded as RD/BD or radius and height.Source: Field walking campaigns · p. 16

A.2 Annotation Protocol

  • Bounding boxes follow left-to-right, top-to-bottom reading order; word or phrase boxes are used for irregular writing, and only text over drawings may overlap drawing boxes.Each scan also has one technical-drawing box; non-overlapping text boxes are preferred to avoid character duplication in spatial metrics such as SpACER [11].Source: §A.2 · p. 17
  • Transcriptions preserve capitalization, spacing, special characters, and abbreviations as written, with no inserted space in ambiguous slash or parenthesis spacing and one space between measurement values and units.For example, a measurement is transcribed as “12 cm.”Source: §A.2 · p. 17
  • Semantic-unit labels comprise seven categories: Excavation, Pottery Form, Measurement, Artefact Category, Publication Type, Surface Treatment, and Others.They cover contextual find information, vessel typology, dimensions, firing technique, typological references, surface finishing, and unmatched segments such as scale annotations, dates, and draughtsperson initials.Source: §A.2 · p. 17

A.3 Data Format and Structure

  • Transcription and field annotations are provided as two JSON files, with scan text, bounding-box coordinates and dimensions, reading-order and drawing indices, and a structured field dictionary.The field file stores the same text alongside the dictionary fields id, text, and fields, following the schema in Table 2.Source: §A.3 · p. 17

B Implementation Details … C Evaluation Details

The implementation combines five OCR checkpoints with model-specific inference and post-processing, LoRA fine-tuning, and prompt-based transcription and field extraction. The passages specify the hardware, training setup, and prompt configurations used.

B.1 Inference Setup

  • Experiments used RTX 3090 GPUs, model-specific precision and batching, and Flash Attention 2 for olmOCR and olmOCR2.Each GPU had 24 GB VRAM; VLMs used bfloat16, Florence2 float16, and TrOCR float32. VLM inference used batch size 1 per image, while TrOCR batched detected text regions within each image.Source: §B.1 · p. 18

B.2 Model Checkpoints

  • The five checkpoints were TrOCR, olmOCR, olmOCR2, Florence2, and LightOnOCR.The corresponding checkpoint identifiers were microsoft/trocr-large-handwritten, allenai/olmOCR-7B-0225-preview, allenai/olmOCR-2-7B-1025, florence-community/Florence-2-large, and lightonai/LightOnOCR-2-1B.Source: §B.2 · p. 18

B.3 HTR Post-processing

  • Post-processing removes markup, image references, URLs, and appended commentary after model-specific text extraction or reading-order concatenation.olmOCR text is extracted from the returned JSON natural_text field; TrOCR and Florence2 segments are concatenated in reading order before cleaning.Source: §B.3 · p. 18

B.4 LoRA Fine-tuning

  • LoRA fine-tuning trained olmOCR2 and LightOnOCR for two epochs on 57 CENTURIA samples, with results also reported at one epoch.LoRA ranks were 16 for olmOCR2 and 8 for LightOnOCR; both used α = 32, dropout 0.05, and AdamW at 2 × 10−4 learning rate.Source: §B.4 · p. 18

Zero-shot Instruction Prompt.

  • The zero-shot domain context prompt replaces the default instruction with vocabulary lists to guide disambiguation of recurring archaeological terms and tokens.The lists cover site abbreviations, excavation identifiers, vessel forms, ware types, measurement conventions, surface treatments, publication author abbreviations, and other recurring tokens.Source: Zero-shot Instruction Prompt · p. 19

Few-shot Prompt.

  • Few-shot prompting uses one, five, or eight training-split image-transcription examples selected from an ordered set stratified across all six campaigns.Examples were chosen to represent varied excavation contexts, field combinations, and label layouts.Source: Few-shot Prompt · p. 19

Instruction + Few-shot Prompt.

  • The combined strategy applies the domain context prompt as the task instruction in every turn, including all eight few-shot demonstration turns.It combines the context prompt with the eight demonstrations described for the few-shot setup.Source: Instruction + Few-shot Prompt · p. 19

LLM Field Extraction Prompt.

  • Field extraction uses greedy no-think Qwen3-8B decoding with a schema prompt containing hard rules, field definitions, a JSON template, and eight few-shot examples.The examples pair corrected transcriptions with expected field annotations; the study evaluates corrupted versus corrected examples in Section C.3.Source: LLM Field Extraction Prompt · p. 19

C.1 Per-field transcription analysis … C.5 Inference speed and output length

Across these analyses, LoRA fine-tuning outperforms prompting strategies, while many remaining extraction errors trace to OCR noise, including digit confusions and degraded scans. Fine-tuning also makes LightOnOCR output faster and shorter.

C.1 Per-field transcription analysis

  • olmOCR had 8.6% pottery form EMR but 64.4% measurement ANLS; after LoRA, both models reached 91.67% measurement EMR.olmOCR2 (+) led on Artefact Category (89.51% vs. 85.94%) and Publication Type (86.18% vs. 75.34%).Source: §C.1 · p. 19, Table 6 · p. 20

C.3 Field extraction error analysis

  • Table 8 details extraction-error patterns and root causes, while Fig. 6 shows the field-level error breakdown.The reported analysis covers errors under both REGEX and LLM extraction.Source: §C.3 · p. 20, Table 8 · p. 22

C.2 Impact of improvement strategies on top-performing models

  • LoRA fine-tuning dominates all prompting strategies for both models, while LightOnOCR reaches 48.71% SpACER-M at 1 shot and 100% at 5 shots.Few-shot prompting reduces olmOCR2 error to 5.38% SpACER-M at 5 shots; LightOnOCR’s multi-image few-shot context instead triggers repetitive output loops.Source: §C.2 · p. 20, Table 7 · p. 21

Model-specific errors.

  • LightOnOCR accounted for 77% of REGEX and 73% of LLM model-unique errors, especially through truncated Publication Type entries; olmOCR2’s main weakness was Pottery Form.LightOnOCR also omits some Artefact Category and Surface Treatment tokens under REGEX, while olmOCR2 truncates compound pottery forms or omits the field.Source: Model-specific errors · p. 21

Shared errors.

  • OCR errors were shared by both models in 60% of REGEX errors and 71% of LLM errors, with visually similar digit substitutions prominent across fields.Both models also omit rare domain tokens such as Eingegl. Ker. and Üz, and fail to preserve percentage signs in radius measurements.Source: Shared errors · p. 20, Table 8 · p. 22

Effect of few-shot example quality.

  • Corrected few-shot examples raised Publication Type accuracy from 74.53% to 82.11% for olmOCR2 (+) and from 62.87% to 68.83% for LightOnOCR (+).Differences were negligible for other fields; corrected examples were equal or better on every field for fine-tuned models, and the main-paper pipeline results use corrected examples.Source: Effect of few-shot example quality · p. 21, Table 9 · p. 23

C.4 Qualitative transcription examples

  • On the high-quality scan, both fine-tuned models transcribed all fields correctly; on the degraded scan, low contrast and faded ink caused shared OCR mistakes and one extraction-stage failure.The two degraded images are 0.4% of the dataset (2 of 507); the extraction-stage failure occurred for fn with LightOnOCR under both methods, where spacing broke the REGEX pattern and led to LLM misassignment.Source: §C.4 · pp. 22–23, Table 10 · p. 23, Figure 7 · p. 24

C.5 Inference speed and output length

  • LightOnOCR (+) was faster than its baseline (1.42 s vs. 2.95 s per document), alongside an 86% reduction in average output length (72 vs. 509 characters).olmOCR2 (+) showed no speed improvement over baseline (2.26 s vs. 2.19 s); its baseline output averaged 77 characters, leaving less room for length-based gains.Source: Table 11 · p. 25, §C.5 · p. 25

D Model Confidence and Smart Merge · E Prompts

The paper combines agreement-based field confidence with Smart Merge to auto-accept some documents while routing unresolved conflicts for review. Its prompts provide domain vocabulary and a training example for OCR, and define a schema and constraints for LLM field extraction.

D Model Confidence and Smart Merge

  • 254 of 450 images (56.4 %) are auto-accepted, and their auto-resolved fields achieve 95.5 % precision.Auto-acceptance requires every field to be CONFIRMED, HIGH, or INFERRED; the remaining 43.6 % of images require manual review.Source: D Model Confidence and Smart Merge · p. 26, Table 12 · p. 26
  • CONFIRMED fields total 2,524 at 98.7 % precision; HIGH and INFERRED cover 206 of 7,650 fields (2.7 %), at 55.9 % and 43.0 % precision, respectively.Within auto-accepted documents, HIGH and INFERRED fields account for most of the residual error.Source: D Model Confidence and Smart Merge · p. 26
  • Smart Merge assigns HIGH when three of four pipelines agree, uses field-specific rules for two-against-two splits to assign INFERRED, and sends cases without an applicable rule to REVIEW.Splits can occur between OCR models or extraction methods; the field-specific rule selects the more reliable source.Source: D Model Confidence and Smart Merge · p. 26, Table 13 · p. 26
  • Publication-type year errors and percentage-radius measurements (r_a/r_i) always go to review because no pipeline resolves them reliably.A document with any unresolved conflict requires manual review.Source: D Model Confidence and Smart Merge · p. 26

E Prompts

  • The zero-shot OCR prompt supplies domain-specific vocabulary lists for ambiguous handwriting, while the 1-shot prompt adds a CENTURIA image-transcription pair to demonstrate format and domain conventions.The vocabulary covers provenance tokens, excavation identifiers, vessel forms, ware types, measurements, surface treatments, and publication abbreviations.Source: Figure 10 · p. 27, Figure 11 · p. 27
  • The Qwen3-8B LLM extraction system prompt specifies the extraction schema, domain constraints, and output format.Source: Figure 12 · p. 28

Related papers

Loading 2608.30616v1…