Vesuvius Challenge awarded a $50K First Letters Prize for scroll PHerc. 306B, where the winning submission reached the 10-character target and the papyrology team agreed on a reading, tentatively a Greek phrase with parallels in Aristotle and Philodemus. Ink was detected by multiplying two models (TRACE v3B and a fine-tuned A-hn). Letters remain faint, the readable area is small, and the work relied on hundreds of hours of manual labeling.
Technology: ML ink-detection models (TRACE v3B, A-hn) with VC3D GrowPatch
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: The reading was checked by the papyrology team, which agreed on the letters with the winning submission, and was compared against 25 community submissions and a 10-character target; the page describes no independent replication.
- Still unknown: The reading is partial and uncertain, letters remain faint, training on machine-generated labels could compound errors (not assessed), and the Grand Prize is not yet awarded.
- AI role: Machine-learning ink-detection models (TRACE v3B and a fine-tuned A-hn) produced the ink maps; the readings and prize review were done by humans.
Evidence sources: https://scrollprize.substack.com/p/writing-seen-in-pherc-306b-for-the
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 2/5 · Validation 4/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Machine-learning ink detection on synchrotron scans let papyrologists read about 22 columns of a likely Stoic ethics text from scroll PHerc. 1667. The text is fragmentary because only the inner core survives, and humans transcribed and reviewed every reading.
Technology: Machine-learning ink detection with virtual unwrapping
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: Readings checked against explicit coverage and papyrological-review criteria; ink recovery validated independently on another scroll (PHerc. Paris 4).
- Still unknown: The text is fragmentary because only the inner core survives; systematic recovery across the library is described as future work.
- AI role: Machine learning used for ink detection on synchrotron scans, alongside improved computational unrolling.
Evidence sources: https://arxiv.org/abs/2606.29085
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: other. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Machine learning models trained on simulated landscape data flagged 9 of 12 known targets (model 1) and all 12 (model 2, with 686 false positives among 709 remaining predictions) in Kisatchie National Forest, Louisiana; a model trained on real South Carolina tar kilns found 11 of 12. Field visits to 11 targets and augering of 2 found none of the expected tar-kiln features, and the authors concluded the structures more likely relate to a WWII military training site. Evidence caveat: only 12 known targets, and field validation is qualitative.
Technology: Machine learning classifiers trained on simulated archaeological sites (with a real tar-kiln-trained comparison model)
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: Models trained on simulated objects flagged known tar-kiln-like targets, and 11 targets were visited in the field with 2 augered; no expected tar-kiln features were found, and the authors concluded the features are most likely associated with military training exercises.
- Still unknown: Access to all targets was limited by dense understory, and further historical research is needed to confirm the function of the objects.
- AI role: Mask R-CNN models trained on simulated lidar-embedded objects produced predictions, a rule-based script removed likely natural mounds, and human archaeologists reviewed and chose field targets.
Evidence sources: https://www.cambridge.org/core/product/identifier/S2326376825101307/type/journal_article
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The model restores texts with 73% top-20 accuracy for gaps up to ten characters (58% when length is unknown), attributes provenance with 72% accuracy across 62 Roman provinces, and dates inscriptions within 13 years; it was trained on over 176,000 inscriptions. In a study with 23 historians on 60 inscriptions, results were best when parallels and predictions were used together. Published in Nature.
Technology: Aeneas
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Corpus of 176,861 Latin inscriptions. Aeneas: 72% geographical accuracy over 62 provinces; dating within 13 years on average; restoration character error rate 40.5% with known length. 23 historians on 60 inscriptions, solo versus with Aeneas parallels and predictions.
- Still unknown: The paper notes a small dataset, survival bias, only about 5% of inscriptions with images, and weaker results where data are scarce. Training data include earlier editors' restorations. Other languages are suggested, not tested.
- AI role: Aeneas restored text, attributed provinces, estimated dates and retrieved parallels; historians evaluated outputs and combined them with their own knowledge.
Evidence sources: https://doi.org/10.1038/s41586-025-09292-5
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A fine-tuned deep-learning model trained with 1960s CORONA satellite images flagged eight previously unrecognized tell candidates in Iraq's Abu Ghraib district. Archaeologists investigated all eight; four contained surface ceramic sherds confirming archaeological remains and enabling dating. The AI suggestions led the field team to inspect locations that routine surveys would not have visited.
Technology: MANet deep convolutional segmentation network with attention, transfer learning and two-stage fine-tuning
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: direct test
- Tested: A fine-tuned deep-learning model trained with 1960s CORONA satellite images flagged eight previously unrecognized tell candidates in Iraq's Abu Ghraib district.
- Still unknown: This is an arXiv preprint. The results are limited to mounded tell sites in Abu Ghraib and depend on a small training set (208 images) and archaeologist-selected field follow-up. Four of eight AI-suggested sites were confirmed; the other four should not be described as archaeological finds. The source reports that the four confirmed sites were destroyed or no longer visible on the ground, with sherds establishing their presence and dates.
- AI role: Researchers retrained the existing neural model using CORONA historical imagery and generated probability heatmaps that archaeologists used to select previously unrecognized locations for field reconnaissance. Eight AI-suggested sites were investigated; four were confirmed through collected and studied ceramic sherds. The authors state the four locations would not have been investigated in routine field surveys without the model's suggestions.
Evidence sources: https://arxiv.org/pdf/2507.13420
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The Enoch tool dated the writing of 135 manuscripts with about 79% agreement with palaeographers and roughly 30-year precision, following new radiocarbon dating; many scrolls appear older than expected. The 79% measures agreement with palaeographers, not accuracy against known dates; the announcement gives no details on training data.
Technology: Enoch (palaeographic dating model)
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: The Enoch dates for 135 manuscripts, following new radiocarbon dating, were compared with palaeographers' judgments, with about 79% agreement.
- Still unknown: The source states the 79% measures agreement with palaeographers rather than accuracy against known dates, that about one in five datings disagreed, and that training-data details are not given.
- AI role: Enoch, a writing-style model trained on radiocarbon-dated samples, dated the manuscripts.
Evidence sources: https://www.sdu.dk/en/om-sdu/fakulteterne/naturvidenskab/nyheder-2025/dead-sea-scrolls · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1371/journal.pone.0323185&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 54 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A DeepLabV3+ segmentation model trained on about 3,600 high-quality samples from over 11,000 candidates mapped Angkor-period reservoirs in satellite imagery across the wider Angkorian realm. The best F1 score reached 45% in favourable regions, comparable to a trained human interpreter. The authors present the model as a second opinion for landscape archaeologists that requires human validation.
Technology: DeepLabV3+ semantic segmentation
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: A DeepLabV3+ model trained on about 3,600 samples mapped Angkor-period reservoirs in satellite imagery and was compared with a trained human interpreter; the best F1 score was 45% in favourable regions.
- Still unknown: The abstract states no limitations; the authors present the model as a second opinion that requires human validation.
- AI role: DeepLabV3+ semantic segmentation detected Angkor-period reservoirs in satellite imagery.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1371/journal.pone.0320452&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 9 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
An IBM model flagged 1,309 candidates; about a quarter were ground-checked over roughly six months, finding 303 new figurative geoglyphs among 430 already known and increasing the discovery rate 16-fold. Another 968 candidates remain to be checked; the study had been accepted by PNAS at the time of the announcement.
Technology: Neural network for aerial images
Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: direct test
- Tested: Field teams ground-checked about a quarter of the 1,309 AI-flagged candidates on foot over roughly six months, finding 303 new figurative geoglyphs.
- Still unknown: 968 candidates remain unsurveyed, training data was limited, and the study was accepted but not yet published in PNAS at the time of the announcement.
- AI role: An IBM model that works with few training samples analyzed aerial geospatial imagery to flag likely geoglyph areas.
Evidence sources: https://www.yamagata-u.ac.jp/en/information/info/20240924/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 24 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
The $700,000 prize went to a submission with 11 additional columns and more than 2,000 characters (about 5% of the first scroll), using models to detect ink and automated segmentation. Papyrologists evaluated 15 anonymized columns and recovered more text from the winning submission. Segmentation remains largely manual and costly.
Technology: Ink detection with vision models and automated segmentation
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: Papyrologists independently evaluated 15 anonymized columns, and the winning submission yielded the most text for every reviewer, meeting the stated criteria.
- Still unknown: The comprehensive study is not yet published, only rough draft transcriptions exist, segmentation is largely manual, and about 5% of the first scroll has been read.
- AI role: Models detected ink in CT scans (a TimeSformer-based model was strongest), and an automated tool produced the segmentation of the papyrus surfaces.
Evidence sources: https://scrollprize.org/grandprize
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
For the first time, a complete word (‘porphyras’, purple) was read from inside a never-opened carbonized scroll; a model trained on crack patterns yielded 13 letters independently and unanimously annotated by papyrologists. The letters are faint and lack context; other readings are not ruled out.
Technology: Machine-learning models for ink detection
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: A model trained on crack patterns produced 13 letters, which papyrologists independently annotated with varying confidence; the page links only to code and no paper.
- Still unknown: The letters are faint and lack context, word boundaries are unclear, and the page states other readings are not ruled out and the papyrological team was still investigating.
- AI role: Machine-learning models trained on ink crackle patterns (and a second model with pseudo-labeling) read letters from the scroll.
Evidence sources: https://scrollprize.org/firstletters
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A neural network reconstructs missing text in damaged Greek inscriptions, identifies their provenance and dates them: 62% accuracy in restoration, 71% in location, and dating within 30 years of reference ranges. With Ithaca, historians improved from 25% to 72% correct restorations. It is intended to assist historians.
Technology: Ithaca (deep neural network)
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: Restoration (62%), location (71%) and dating were checked on held-out dated inscriptions, and a historian study rose from 25% to 72% correct restorations with the tool.
- Still unknown: Accuracy is far from perfect, dating is a range within about 30 years, and the model was trained only on Greek inscriptions.
- AI role: Neural network restored missing text, identified provenance and dated Greek inscriptions.
Evidence sources: https://deepmind.google/discover/blog/predicting-the-past-with-ithaca/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-022-04448-z&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 268 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Independent analyses of character shapes distinguish two groups across the scroll’s 54 columns, consistent with two scribes alternating around the manuscript’s midpoint. This is statistical evidence from one artifact; the writers’ identities are unknown, and other explanations for the differences are not entirely ruled out.
Technology: Pattern recognition and Kohonen maps
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Independent analyses of character shapes across 54 columns separated two groups, treated as statistical evidence from one artifact.
- Still unknown: Not reported in the abstract read.
- AI role: Pattern recognition and AI techniques analysed handwriting features to identify writers.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1371/journal.pone.0249769&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 63 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A machine-learning model recognizes signs on tablets from the Persepolis Fortification Archive with about 80% accuracy on tablets excluded from training (over 6,000 annotated images and over 100,000 signs in the database). About 20% of signs are wrong; testing is internal only, and the source is a university announcement.
Technology: DeepScribe (machine learning on images)
Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: retrospective or benchmark
- Tested: Recognition of signs on tablets excluded from training reached about 80% accuracy; the source describes the testing as internal only.
- Still unknown: About 20% of signs are still wrong; the size of the held-out set and an error breakdown are not given; other uses are described as aspirations.
- AI role: Machine-learning model trained on over 100,000 annotated signs from Persepolis Fortification Archive tablet images to recognize cuneiform signs.
Evidence sources: https://news.uchicago.edu/story/how-ai-could-help-translate-written-language-ancient-civilizations
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.