inkbench
Copyright (c) 2026 KarLex AI, Inc. Licensed under the MIT License (LICENSE).

inkbench transcribes scoring rules and small functions from the following MIT-licensed projects, so that its numbers
are comparable to the published ones. Each source file that does so names its origin. Their copyright notices and the
MIT permission notice (reproduced once at the end of this file) apply to those portions.

1. DomRusso2/reader-v2 (https://github.com/DomRusso2/reader-v2), MIT, Copyright (c) 2026 Domenico Russo.
   - src/inkbench/metrics.py: auc_russo (scripts/v2_run.py, auc: 300 000 pixels per class, seed 0) and
     fa_at_recall (the blank-papyrus false-alarm rule of scripts/bkill.py).
   - src/inkbench/inputs.py: pick_crops (scripts/bkill.py, pick_crops) and the torch "area" resize of the suite.
   - src/inkbench/score.py, src/inkbench/refs.py: the evaluation regions of scripts/v2_run.py (evaluate: union of
     supervision and validation, the 0009B union taken on the fine canvas, map classes >= 128 and <= 40).
   - src/inkbench/predict.py: score_map, the 8-bit scoring of scripts/hecate_eval.py (round(255 p), resized back
     onto the native canvas with INTER_LINEAR).

2. DomRusso2/ir-truth-9um (https://github.com/DomRusso2/ir-truth-9um), code MIT, Domenico Russo.
   - src/inkbench/metrics.py: ir_quantise, ir_draws and ir_truth_score (scripts/score_map.py: percentile
     quantisation, 2000 random placements of the letter mask).

3. TAUIL-Abd-Elilah/9um-reader-letter-shape (https://github.com/TAUIL-Abd-Elilah/9um-reader-letter-shape), MIT,
   Copyright (c) 2026 TAUIL Abd Elilah.
   - src/inkbench/metrics.py: auc_tauil (src/eval9v2.py, auc) and highpass / hp_score (src/hp_eval.py).
   - src/inkbench/score.py: the near-2 mm human-label region (ink pixels and no-ink pixels within 214 px of ink).

4. Nieuwlaar/ink9um-dense-native (https://huggingface.co/Nieuwlaar/ink9um-dense-native), MIT (model card),
   Erwin Nieuwlaar.
   - src/inkbench/metrics.py: auc_nieuwlaar, native78_thresholds, native78_classes, nieuwlaar_pred,
     nieuwlaar_fp_control and misregistration_auc (native-eval/native_bench.py).

5. Jashann/vesuvius-sheet-chain (https://github.com/Jashann/vesuvius-sheet-chain), MIT,
   Copyright (c) 2026 Jashanjot Singh Gill and Nhat Nam Tran.
   - src/inkbench/score.py: the hf_mask_edge200 region of sheetchain/score_vs_labels.py (label mask, CT present,
     at least 200 px from the edge of the render).

6. slchicoineau-creator/tifxyz-canvas-map (https://github.com/slchicoineau-creator/tifxyz-canvas-map), MIT,
   Copyright (c) 2026 Sebastien Chicoineau and Chris Thomas.
   - src/inkbench/canvasmap.py: an independent implementation of their canvas-map rule (no code copied).
   - tests/test_canvasmap.py: three rows of results/pair_maps.tsv as test data.

Methods used without copying code (credited here and in README.md):

7. ChullMotors/ink-nullkit (https://github.com/ChullMotors/ink-nullkit), MIT, Copyright (c) 2026 Chris Hull.
   - src/inkbench/bootstrap.py follows the design of nullkit's paired_auroc_diff (256 px blocks with at least
     500 labelled pixels, the same draw for both readers, 2000 draws, MDE = 2.8 SD); the code is new.

8. ShribyrLabs/vesuvius-reports (https://github.com/ShribyrLabs/vesuvius-reports), MIT,
   Copyright (c) 2026 Chris Scheirer / Shribyr Labs.
   - The letter-scale score (report 02: Pearson r after a 48 um Gaussian high-pass), as item 3 applies it.

9. Nieuwlaar/pherc343-first-letters (https://github.com/Nieuwlaar/pherc343-first-letters), MIT,
   Copyright (c) 2026 Erwin Nieuwlaar.
   - src/inkbench/ensemble.py re-implements the ensemble rule of ink/scripts/ensemble_maps.py (per-model p1/p99.5
     stretch, z-score, mean of the z-scores) for single-depth maps.

At run time inkbench imports inksweep (MIT, KarLex AI, Inc.) for the readers, their pinned weights and the
inference engine. Model weights, scan data and anything derived from scan data (surface volumes, ink maps, labels,
images) are not part of inkbench: they are downloaded from their authors' repositories and the Vesuvius Challenge
open-data bucket under their own licences (scan data and derived data: CC BY-NC 4.0).

MIT License (applies to the portions listed above, with the respective copyright notices):

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
