diff --git a/CHANGELOG.md b/CHANGELOG.md index 87a43f8a02eda319de6a28acf84b40a426ec35ff..87267f225ad148dba2239624dab76483087a2c2a 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -3441,3 +3441,22 @@ Changes to the content of this index, oldest first. There is one entry for each - npm: refused - scotland-government-workbooks: changed pointer group file data/scotland-government-workbooks/2026-10-a-spreadsheets-with-answers-2/part-00000.parquet: 2 key(s) of the old version are in no data file of source scotland-government-workbooks after the commit (row 3 from batch 2026-10-a-spreadsheets-with-answers-2, row 6 from batch 2026-10-a-spreadsheets-with-answers-2); the removal record holds none of them, and none fails a publish check now - stackexchange-api-posts: refused + +## 2026-11-12: sources held back + +- Each source below keeps its newest version that passed every check; its newer batch waits until it passes them: + - borealis-dataverse-workbooks-cc-by: changed pointer group file data/borealis-dataverse-workbooks-cc-by/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 2 key(s) of the old version are in no data file of source borealis-dataverse-workbooks-cc-by after the commit (row 4 from batch 2026-10-a-spreadsheets-with-answers, row 5 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - borealis-dataverse-workbooks-cc0: changed pointer group file data/borealis-dataverse-workbooks-cc0/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 1 key(s) of the old version are in no data file of source borealis-dataverse-workbooks-cc0 after the commit (row 7 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - cfpb-complaints: refused + - data-gov-uk-workbooks: changed pointer group file data/data-gov-uk-workbooks/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 4 key(s) of the old version are in no data file of source data-gov-uk-workbooks after the commit (row 5 from batch 2026-10-a-spreadsheets-with-answers, row 6 from batch 2026-10-a-spreadsheets-with-answers, row 12 from batch 2026-10-a-spreadsheets-with-answers, 1 more); the removal record holds none of them, and none fails a publish check now + - data-gov-uk-workbooks-noassertion: changed pointer group file data/data-gov-uk-workbooks-noassertion/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 1 key(s) of the old version are in no data file of source data-gov-uk-workbooks-noassertion after the commit (row 6 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - dataversenl-workbooks-cc-by: changed pointer group file data/dataversenl-workbooks-cc-by/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 1 key(s) of the old version are in no data file of source dataversenl-workbooks-cc-by after the commit (row 17 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - dataversenl-workbooks-cc0: changed pointer group file data/dataversenl-workbooks-cc0/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 3 key(s) of the old version are in no data file of source dataversenl-workbooks-cc0 after the commit (row 7 from batch 2026-10-a-spreadsheets-with-answers, row 13 from batch 2026-10-a-spreadsheets-with-answers, row 16 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - dvn-fvhlgg-repro: changed pointer group file data/dvn-fvhlgg-repro/2026-10-c-repros/part-00000.parquet: 3 key(s) of the old version are in no data file of source dvn-fvhlgg-repro after the commit (row 5 from batch 2026-10-c-repros, row 7 from batch 2026-10-c-repros, row 16 from batch 2026-10-c-repros); the removal record holds none of them, and none fails a publish check now + - eu-institution-workbooks-cc-by: changed pointer group file data/eu-institution-workbooks-cc-by/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 3 key(s) of the old version are in no data file of source eu-institution-workbooks-cc-by after the commit (row 2 from batch 2026-10-a-spreadsheets-with-answers, row 4 from batch 2026-10-a-spreadsheets-with-answers, row 5 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - hdx-workbooks-cc0: changed pointer group file data/hdx-workbooks-cc0/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 2 key(s) of the old version are in no data file of source hdx-workbooks-cc0 after the commit (row 4 from batch 2026-10-a-spreadsheets-with-answers, row 5 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - hdx-workbooks-odbl: changed pointer group file data/hdx-workbooks-odbl/2026-10-a-spreadsheets-with-answers/part-00000.parquet: 1 key(s) of the old version are in no data file of source hdx-workbooks-odbl after the commit (row 4 from batch 2026-10-a-spreadsheets-with-answers); the removal record holds none of them, and none fails a publish check now + - hf-training-records: refused + - npm: refused + - scotland-government-workbooks: changed pointer group file data/scotland-government-workbooks/2026-10-a-spreadsheets-with-answers-2/part-00000.parquet: 2 key(s) of the old version are in no data file of source scotland-government-workbooks after the commit (row 3 from batch 2026-10-a-spreadsheets-with-answers-2, row 6 from batch 2026-10-a-spreadsheets-with-answers-2); the removal record holds none of them, and none fails a publish check now + - stackexchange-api-posts: refused diff --git a/README.md b/README.md index e9c0cc8e1cebd605ef89aeb038fd331076eefd1d..fc89ff77cef2ffa20d746ca983f623158c0cef83 100644 --- a/README.md +++ b/README.md @@ -35,18 +35,18 @@ this dataset is the foundation's public index. English | [简体中文](README.zh.md) · [Data page](https://huggingface.co/spaces/Battam/training-data-foundation) · [Usage video](https://huggingface.co/spaces/Battam/training-data-foundation#usage-video) - -What each training stage can use in this release. Pre-training (provisional): 30.6 TB deduplicated size, 149,118,607 rows, 29.9 TB allowed for training. Mid-training (provisional): 3.86 TB deduplicated size, 34,079,767 rows, 1.42 TB allowed for training. Post-training: 1,761 material sets, from 1,818 upstream sources, 1,409 allowed for training. + +What each training stage can use in this release. Pre-training (provisional): 30.6 TB deduplicated size, 149,118,607 rows, 29.9 TB allowed for training. Mid-training (provisional): 3.86 TB deduplicated size, 34,079,767 rows, 1.42 TB allowed for training. Post-training: 1,761 material sets, from 1,818 upstream sources, 1,409 allowed for training. - - + + ## Get data in one command ```sh -uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/release-2026-11-12.6/tools/fetch.py --form spreadsheets --sample 3 --out ./data +uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/release-2026-11-12.7/tools/fetch.py --form spreadsheets --sample 3 --out ./data ``` This fetches 3 units of every spreadsheet material set into `./data`, one folder per unit, and checks each file against the size and hashes its row records. @@ -56,7 +56,7 @@ the same command takes the same units of this release.
What the command takes, and in which order -The command and the queries below are pinned to the release tag `release-2026-11-12.6`. +The command and the queries below are pinned to the release tag `release-2026-11-12.7`. [uv](https://docs.astral.sh/uv/) reads the script's dependencies, at the versions published before the day of this release. The command returns units: 3 units from each material set whose form is `spreadsheets`, each in a folder of its own named after its `unit_id`. @@ -84,7 +84,7 @@ Choose material sets from the catalog with [DuckDB](https://duckdb.org/): ```sql SELECT source, collection, title, units, files, top_types, license.tier AS license_tier, needs_auth -FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.6/meta/catalog.parquet' +FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.7/meta/catalog.parquet' WHERE kind = 'material_set' AND needs_auth = 'no' AND license.tier IN ('permissive', 'share-alike') ORDER BY collection, units DESC; ``` @@ -95,7 +95,7 @@ with `kind = 'corpus'` it returns corpora instead, counted in rows (`files`). Choose files in one corpus: ```sql -SET VARIABLE files = (SELECT list('hf://datasets/Battam/training-data-foundation@release-2026-11-12.6/' || f) FROM (SELECT unnest(data_files) AS f FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.6/meta/catalog.parquet' WHERE source = 'if-archive')); +SET VARIABLE files = (SELECT list('hf://datasets/Battam/training-data-foundation@release-2026-11-12.7/' || f) FROM (SELECT unnest(data_files) AS f FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.7/meta/catalog.parquet' WHERE source = 'if-archive')); SELECT uri, size_bytes, lower(hex(sha256)) AS sha256 FROM read_parquet(getvariable('files')) WHERE source = 'if-archive' AND ext = 'pdf' AND size_bytes BETWEEN 100000 AND 5000000 @@ -111,12 +111,12 @@ A data file may hold the rows of several sources, so the query keeps this one's ## What's inside - -The training stages by top-level domain, largest first. Pre-training (provisional), deduplicated size, by domain: Data science 29.8 TB; Science 531 GB; Machine learning 82.8 GB; Media 48.9 GB; System administration 35.7 GB; Hardware 29.6 GB; Many domains 29.6 GB; Security 4.18 GB; Unclassified 1.43 GB; Misc 376 MB; Software engineering 175 MB; Operations 4.08 KB. Mid-training (provisional), deduplicated size, by domain: System administration 1.27 TB; Many domains 1.00 TB; Machine learning 503 GB; Science 461 GB; Data science 429 GB; Hardware 177 GB; Security 13.9 GB; Misc 3.45 GB; Software engineering 2.48 GB; Unclassified 2.21 GB; Media 701 MB; Operations 13.8 MB. Post-training, material sets, by domain: Data science 348; Software engineering 286; Media 273; Operations 206; Machine learning 182; Science 174; Security 103; Hardware 87; System administration 71; Misc 31. + +The training stages by top-level domain, largest first. Pre-training (provisional), deduplicated size, by domain: Data science 29.8 TB; Science 531 GB; Machine learning 82.8 GB; Media 48.9 GB; System administration 35.7 GB; Hardware 29.6 GB; Many domains 29.6 GB; Security 4.18 GB; Unclassified 1.43 GB; Misc 376 MB; Software engineering 175 MB; Operations 4.08 KB. Mid-training (provisional), deduplicated size, by domain: System administration 1.27 TB; Many domains 1.00 TB; Machine learning 503 GB; Science 461 GB; Data science 429 GB; Hardware 177 GB; Security 13.9 GB; Misc 3.45 GB; Software engineering 2.48 GB; Unclassified 2.21 GB; Media 701 MB; Operations 13.8 MB. Post-training, material sets, by domain: Data science 348; Software engineering 286; Media 273; Operations 206; Machine learning 182; Science 174; Security 103; Hardware 87; System administration 71; Misc 31. - - + +
@@ -148,12 +148,12 @@ Every material set and every corpus, with its counts, size, license tier and log ## How it is made - -How the index is made, in four steps. 1, collect: collectors read whole public upstreams into corpora, one row per file or record, the material of pre-training and mid-training. 2, admit: works the search found upstream and groups of corpus rows become material sets, the material of post-training. 3, check: licenses, personal data, overlap with benchmarks and sizes are checked before release. 4, release: one commit with its own tag, release-2026-11-12.6. + +How the index is made, in four steps. 1, collect: collectors read whole public upstreams into corpora, one row per file or record, the material of pre-training and mid-training. 2, admit: works the search found upstream and groups of corpus rows become material sets, the material of post-training. 3, check: licenses, personal data, overlap with benchmarks and sizes are checked before release. 4, release: one commit with its own tag, release-2026-11-12.7. - - + + Collectors read whole public upstreams into corpora, the material of pre-training and mid-training, and admission turns works found upstream and chosen corpus rows into material sets, the material of post-training. @@ -180,7 +180,7 @@ The full text, with the license details, personal information read from sample f @misc{training_data_foundation, title = {training-data-foundation: an index of public materials for building agent environments}, howpublished = {https://huggingface.co/datasets/Battam/training-data-foundation}, - note = {Release release-2026-11-12.6}, + note = {Release release-2026-11-12.7}, year = {2026} } ``` diff --git a/README.zh.md b/README.zh.md index fec36f78d34b29a9c957b342488e33e7c86c4ffd..390886ceb4bc6868bca9a5290cd28e584e2e1589 100644 --- a/README.zh.md +++ b/README.zh.md @@ -22,18 +22,18 @@ tags: [English](README.md) | 简体中文 · [数据页](https://huggingface.co/spaces/Battam/training-data-foundation) · [使用示例视频](https://huggingface.co/spaces/Battam/training-data-foundation#usage-video) - -本次发布中每个训练阶段能用的数据。预训练(暂定):去重后大小 30.6 TB,149,118,607 行,其中 29.9 TB 允许用于训练。中训练(暂定):去重后大小 3.86 TB,34,079,767 行,其中 1.42 TB 允许用于训练。后训练:材料集 1,761,来自 1,818 个上游来源,其中 1,409 允许用于训练。 + +本次发布中每个训练阶段能用的数据。预训练(暂定):去重后大小 30.6 TB,149,118,607 行,其中 29.9 TB 允许用于训练。中训练(暂定):去重后大小 3.86 TB,34,079,767 行,其中 1.42 TB 允许用于训练。后训练:材料集 1,761,来自 1,818 个上游来源,其中 1,409 允许用于训练。 - - + + ## 一条命令取数据 ```sh -uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/release-2026-11-12.6/tools/fetch.py --form spreadsheets --sample 3 --out ./data +uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/release-2026-11-12.7/tools/fetch.py --form spreadsheets --sample 3 --out ./data ``` 这条命令从每个电子表格材料集各取 3 个单元放进 `./data`,每个单元一个文件夹,并按每个文件所在行记录的大小和哈希值校验它。 @@ -43,7 +43,7 @@ uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/r
命令取什么、按什么顺序取 -这条命令和下面的查询都固定在发布标签 `release-2026-11-12.6` 上。 +这条命令和下面的查询都固定在发布标签 `release-2026-11-12.7` 上。 [uv](https://docs.astral.sh/uv/) 会读取脚本声明的依赖,取这个发布当天之前发布的版本。 命令返回单元:形式为 `spreadsheets` 的每个材料集各 3 个单元,每个单元放在以它的 `unit_id` 命名的单独文件夹里。 @@ -71,7 +71,7 @@ uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/r ```sql SELECT source, collection, title, units, files, top_types, license.tier AS license_tier, needs_auth -FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.6/meta/catalog.parquet' +FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.7/meta/catalog.parquet' WHERE kind = 'material_set' AND needs_auth = 'no' AND license.tier IN ('permissive', 'share-alike') ORDER BY collection, units DESC; ``` @@ -82,7 +82,7 @@ ORDER BY collection, units DESC; 在一个语料库里挑选文件: ```sql -SET VARIABLE files = (SELECT list('hf://datasets/Battam/training-data-foundation@release-2026-11-12.6/' || f) FROM (SELECT unnest(data_files) AS f FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.6/meta/catalog.parquet' WHERE source = 'if-archive')); +SET VARIABLE files = (SELECT list('hf://datasets/Battam/training-data-foundation@release-2026-11-12.7/' || f) FROM (SELECT unnest(data_files) AS f FROM 'hf://datasets/Battam/training-data-foundation@release-2026-11-12.7/meta/catalog.parquet' WHERE source = 'if-archive')); SELECT uri, size_bytes, lower(hex(sha256)) AS sha256 FROM read_parquet(getvariable('files')) WHERE source = 'if-archive' AND ext = 'pdf' AND size_bytes BETWEEN 100000 AND 5000000 @@ -98,12 +98,12 @@ ORDER BY size_bytes LIMIT 20; ## 内容概览 - -按顶层领域分的各训练阶段,由大到小。预训练(暂定),去重后大小,按领域:数据科学 29.8 TB;科学 531 GB;机器学习 82.8 GB;媒体 48.9 GB;系统管理 35.7 GB;硬件 29.6 GB;多领域 29.6 GB;安全 4.18 GB;未分类 1.43 GB;杂项 376 MB;软件工程 175 MB;运营 4.08 KB。中训练(暂定),去重后大小,按领域:系统管理 1.27 TB;多领域 1.00 TB;机器学习 503 GB;科学 461 GB;数据科学 429 GB;硬件 177 GB;安全 13.9 GB;杂项 3.45 GB;软件工程 2.48 GB;未分类 2.21 GB;媒体 701 MB;运营 13.8 MB。后训练,材料集,按领域:数据科学 348;软件工程 286;媒体 273;运营 206;机器学习 182;科学 174;安全 103;硬件 87;系统管理 71;杂项 31。 + +按顶层领域分的各训练阶段,由大到小。预训练(暂定),去重后大小,按领域:数据科学 29.8 TB;科学 531 GB;机器学习 82.8 GB;媒体 48.9 GB;系统管理 35.7 GB;硬件 29.6 GB;多领域 29.6 GB;安全 4.18 GB;未分类 1.43 GB;杂项 376 MB;软件工程 175 MB;运营 4.08 KB。中训练(暂定),去重后大小,按领域:系统管理 1.27 TB;多领域 1.00 TB;机器学习 503 GB;科学 461 GB;数据科学 429 GB;硬件 177 GB;安全 13.9 GB;杂项 3.45 GB;软件工程 2.48 GB;未分类 2.21 GB;媒体 701 MB;运营 13.8 MB。后训练,材料集,按领域:数据科学 348;软件工程 286;媒体 273;运营 206;机器学习 182;科学 174;安全 103;硬件 87;系统管理 71;杂项 31。 - - + +
@@ -135,12 +135,12 @@ ORDER BY size_bytes LIMIT 20; ## 怎样做成 - -索引分四步做成。1,采集:采集器整批读取完整的公开上游,每个文件或每条记录记一行,成为语料库,供预训练和中训练使用。2,收录:检索在上游找到的作品和成组的语料库行,收录为材料集,供后训练使用。3,检查:发布前核对许可证、个人信息、与基准测试的重叠和大小。4,发布:一个带有自己标签的提交,release-2026-11-12.6。 + +索引分四步做成。1,采集:采集器整批读取完整的公开上游,每个文件或每条记录记一行,成为语料库,供预训练和中训练使用。2,收录:检索在上游找到的作品和成组的语料库行,收录为材料集,供后训练使用。3,检查:发布前核对许可证、个人信息、与基准测试的重叠和大小。4,发布:一个带有自己标签的提交,release-2026-11-12.7。 - - + + 采集器整批读取完整的公开上游,记成语料库,供预训练和中训练使用,收录则把检索在上游找到的作品和选出的语料库行做成材料集,供后训练使用。 @@ -167,7 +167,7 @@ ORDER BY size_bytes LIMIT 20; @misc{training_data_foundation, title = {training-data-foundation: an index of public materials for building agent environments}, howpublished = {https://huggingface.co/datasets/Battam/training-data-foundation}, - note = {Release release-2026-11-12.6}, + note = {Release release-2026-11-12.7}, year = {2026} } ``` diff --git a/campaign_files/2026-10-a-spreadsheets-with-answers-2/swamdudi-pharma-inventory-model/README.md b/campaign_files/2026-10-a-spreadsheets-with-answers-2/swamdudi-pharma-inventory-model/README.md index b15538a58c9b58f2cdc9c67a9a714e75857f2a48..62adb4fdb27bc62a2f54520b9d59939b5455ab09 100644 --- a/campaign_files/2026-10-a-spreadsheets-with-answers-2/swamdudi-pharma-inventory-model/README.md +++ b/campaign_files/2026-10-a-spreadsheets-with-answers-2/swamdudi-pharma-inventory-model/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. -**Answers:** Each file with an answer has one answer file, at answer/.json (the unit path plus .json). Reference result kind: key_formula_values. - **Tools to open and check:** openpyxl (formulas and cached values); LibreOffice headless recalculation (soffice --headless --convert-to xlsx) **Licence (SPDX):** MIT diff --git a/campaign_files/2026-10-b-disk-images-3/tamuctf-2021-recovery/README.md b/campaign_files/2026-10-b-disk-images-3/tamuctf-2021-recovery/README.md index f661a62af1bee51d8de9e21b090109f9b9f815d5..d8fc26b92fbb7fffdf20e6f956e9172bfa9398c6 100644 --- a/campaign_files/2026-10-b-disk-images-3/tamuctf-2021-recovery/README.md +++ b/campaign_files/2026-10-b-disk-images-3/tamuctf-2021-recovery/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether the file needs other upstream files is not stated. Dependencies named upstream are listed per file in the notes. -**Answers:** Each file with an answer has one answer file, at answer/.json (the unit path plus .json). Reference result kind: deleted_files. - **Tools to open and check:** Open with tools for these file types: .zip. **Licence (SPDX):** NOASSERTION @@ -20,5 +18,4 @@ Files: - `forensics/recovery/floppy.img.zip`: ext2 floppy image with one deleted GIF /flag.gif (TAMUctf 2021 'Recovery'). Origin: CTF image made by the challenge author (problem.json author Matthew Barry). - Reference result: deleted_files Dependencies: unzip or Python zipfile to unpack; only needed to rerun the exercise, not to grade the answer diff --git a/campaign_files/2026-10-c-repros-2/staq-repro/README.md b/campaign_files/2026-10-c-repros-2/staq-repro/README.md index b341e99c3c82702979eae8f94a10ebf727ba0445..9729e8310d0756feff53f7dab19c8467e85824a8 100644 --- a/campaign_files/2026-10-c-repros-2/staq-repro/README.md +++ b/campaign_files/2026-10-c-repros-2/staq-repro/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether the unit needs other upstream files is not stated. Dependencies named upstream are listed per unit in the notes. -**Answers:** Each unit with an answer has one answer file, at answer//reported.json. - **Tools to open and check:** Open with tools for these file types: .sh. **Licence (SPDX):** MIT diff --git a/campaign_files/2026-10-c-repros-2/stochastic-dg-swe-repro/README.md b/campaign_files/2026-10-c-repros-2/stochastic-dg-swe-repro/README.md index 348757083b3db8016233c036cb118396719f948e..7035da2d7f7e42395caed36a3a45e5e39896c64a 100644 --- a/campaign_files/2026-10-c-repros-2/stochastic-dg-swe-repro/README.md +++ b/campaign_files/2026-10-c-repros-2/stochastic-dg-swe-repro/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether the unit needs other upstream files is not stated. 1 of 1 units name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per unit in the notes. Dependencies named upstream are listed per unit in the notes. -**Answers:** Each unit with an answer has one answer file, at answer//reported.json. - **Tools to open and check:** Open with tools for these file types: .jl; .toml. **Licence (SPDX):** MIT diff --git a/campaign_files/2026-10-g-solver-instances-2/steinlib/README.md b/campaign_files/2026-10-g-solver-instances-2/steinlib/README.md index 21f61f892de2c9ef5cbc5a907a3bb4180acf7a97..90cd59169b5c47066e34229ca1124ebb70f6f501 100644 --- a/campaign_files/2026-10-g-solver-instances-2/steinlib/README.md +++ b/campaign_files/2026-10-g-solver-instances-2/steinlib/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 15 of 15 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Published optimal objective value per instance. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Tools are named in the check written per file in the notes. **Licence (SPDX):** NOASSERTION @@ -21,120 +19,90 @@ Files: - `download/MC.tgz`: SteinLib set MC: 6 Steiner tree problem instances in graphs, randomly generated by Francois Margot (grid graphs and complete graphs). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; randomly generated by Francois Margot (grid graphs and complete graphs); set page: showset.php?MC line 22: "These problems were random generated by Francois Margot."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?MC (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 6 of 6 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 6 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?MC (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?MC Dependencies: no upstream code used - `download/X.tgz`: SteinLib set X: 3 Steiner tree problem instances in graphs, complete graphs with Euclidean distances converted by Ralf Borndoerfer and Carlos Ferreira from TSPLIB problems by selecting some nodes as terminals. Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; complete graphs with Euclidean distances converted by Ralf Borndoerfer and Carlos Ferreira from TSPLIB problems by selecting some nodes as terminals; set page: showset.php?X lines 22 to 25: "These instances are complete graphs with euclidian distances. They were converted by Ralf Bornörfer and Carlos Ferreira from TSPLIB problems, by selecting some nodes as Terminals."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?X (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 3 of 3 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 2 instances (world666, 221,445 edges, not run), consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?X (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?X Dependencies: no upstream code used - `download/SP.tgz`: SteinLib set SP: 8 Steiner tree problem instances in graphs, artificially constructed hard Steiner tree problems (Koch and Martin). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; artificially constructed hard Steiner tree problems (Koch and Martin); set page: showset.php?SP line 23: "Artificially constructed Steiner tree problems."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?SP (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 8 of 8 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 8 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?SP (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?SP Dependencies: no upstream code used - `download/I080.tgz`: SteinLib set I080: 100 Steiner tree problem instances in graphs, random sparse graphs with incidence edge weights chosen to defy preprocessing (Cees Duin), 80 nodes. Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; CRLF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; random sparse graphs with incidence edge weights chosen to defy preprocessing (Cees Duin), 80 nodes; set page: showset.php?I080 lines 22 to 23: "These instances were random generated sparse graphs with so called incidence edge weights, which are choosen to defy preprocessing."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?I080 (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 100 of 100 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 100 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?I080 (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?I080 Dependencies: no upstream code used - `download/P4E/P4E.tgz`: SteinLib set P4E: 11 Steiner tree problem instances in graphs, random complete graphs with Euclidean weights (Chopra, Gorres and Rao, CGR92). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; the archive also holds p4e100.grp, p4e200.grp (point coordinate files of the generator, not needed for judging); one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; random complete graphs with Euclidean weights (Chopra, Gorres and Rao, CGR92); set page: showset.php?P4E lines 22 to 23: "These instances are random generated complete graphs with euclidian weights."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?P4E (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 11 of 11 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 11 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?P4E (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?P4E Dependencies: no upstream code used - `download/P4Z.tgz`: SteinLib set P4Z: 10 Steiner tree problem instances in graphs, random complete graphs with random weights (Chopra, Gorres and Rao, CGR92). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; the archive also holds p4z100.grp (point coordinate files of the generator, not needed for judging); one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; random complete graphs with random weights (Chopra, Gorres and Rao, CGR92); set page: showset.php?P4Z lines 22 to 23: "These instances are random generated complete graphs with random weights."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?P4Z (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 10 of 10 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 10 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?P4Z (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?P4Z Dependencies: no upstream code used - `download/P6E.tgz`: SteinLib set P6E: 15 Steiner tree problem instances in graphs, random sparse graphs with Euclidean weights (Chopra, Gorres and Rao, CGR92). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; random sparse graphs with Euclidean weights (Chopra, Gorres and Rao, CGR92); set page: showset.php?P6E lines 22 to 23: "These instances are random generated sparse graphs with euclidian weights."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?P6E (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 15 of 15 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 15 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?P6E (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?P6E Dependencies: no upstream code used - `download/ALUE.tgz`: SteinLib set ALUE: 15 Steiner tree problem instances in graphs, real world VLSI layout: grid graphs with rectangular holes (Martin and Koch). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; real world VLSI layout: grid graphs with rectangular holes (Martin and Koch); set page: showset.php?ALUE lines 22 to 23: "These instances are grid graphs with rectangular holes. They come from a VLSI application"; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?ALUE (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 15 of 15 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 15 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?ALUE (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?ALUE Dependencies: no upstream code used - `download/ES10FST.tgz`: SteinLib set ES10FST: 15 Steiner tree problem instances in graphs, rectilinear Steiner problems: 10 random terminals on a 10,000,000 by 10,000,000 grid reduced to a graph by the GeoSteiner FST generator (Zachariasen and Warme). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; rectilinear Steiner problems: 10 random terminals on a 10,000,000 by 10,000,000 grid reduced to a graph by the GeoSteiner FST generator (Zachariasen and Warme); set page: showset.php?ES10FST lines 22 to 26: "First there were files with random generated points in the plane on a 10,000,000 by 10,000,000 grid. These points serve as terminals and were converted to rectalinear graphs with L1 edge weights"; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?ES10FST (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 15 of 15 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 15 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?ES10FST (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?ES10FST Dependencies: no upstream code used - `download/WRP3.tgz`: SteinLib set WRP3: 64 Steiner tree problem instances in graphs, industrial VLSI wire routing group Steiner problems, transformed to plain Steiner tree instances (Zachariasen and Rohe, ZR00). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; industrial VLSI wire routing group Steiner problems, transformed to plain Steiner tree instances (Zachariasen and Rohe, ZR00); set page: showset.php?WRP3 lines 22 to 23: "These instances are wire routing problems from industry. Only one terminal of a group needs to be connected to the tree."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?WRP3 (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 63 of 64 instances have a value; wrp3-95.stp is in the archive but has no row in the table, so it carries no answer and is left out of the task pool. the tree connects all terminals and its cost is never below the published optimum on any of 63 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?WRP3 (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?WRP3 Dependencies: no upstream code used - `download/DIW.tgz`: SteinLib set DIW: 21 Steiner tree problem instances in graphs, real world VLSI layout: grid graphs with rectangular holes (Martin and Koch). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; real world VLSI layout: grid graphs with rectangular holes (Martin and Koch); set page: showset.php?DIW lines 22 to 23: "These instances are grid graphs with rectangular holes. They come from a VLSI application"; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?DIW (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 21 of 21 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 21 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?DIW (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?DIW Dependencies: no upstream code used - `download/LIN.tgz`: SteinLib set LIN: 37 Steiner tree problem instances in graphs, sparse grid like graphs derived from a placement of rectangular blocks in the plane (C. Lin). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; sparse grid like graphs derived from a placement of rectangular blocks in the plane (C. Lin); set page: showset.php?LIN lines 24 to 26: "They are derived from a placement of rectangular blocks in a 2D plane, so all line segments are either horizontal or vertical."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?LIN (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 37 of 37 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 37 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?LIN (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?LIN Dependencies: no upstream code used - `download/TAQ.tgz`: SteinLib set TAQ: 14 Steiner tree problem instances in graphs, real world VLSI layout: grid graphs with rectangular holes (Martin and Koch). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; real world VLSI layout: grid graphs with rectangular holes (Martin and Koch); set page: showset.php?TAQ lines 22 to 23: "These instances are grid graphs with rectangular holes. They come from a VLSI application"; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?TAQ (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 14 of 14 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 14 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?TAQ (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?TAQ Dependencies: no upstream code used - `download/I160.tgz`: SteinLib set I160: 100 Steiner tree problem instances in graphs, random sparse graphs with incidence edge weights chosen to defy preprocessing (Cees Duin), 160 nodes. Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; CRLF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; random sparse graphs with incidence edge weights chosen to defy preprocessing (Cees Duin), 160 nodes; set page: showset.php?I160 lines 22 to 23: "These instances were random generated sparse graphs with so called incidence edge weights, which are choosen to defy preprocessing."; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?I160 (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 100 of 100 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 100 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?I160 (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?I160 Dependencies: no upstream code used - `download/ES20FST.tgz`: SteinLib set ES20FST: 15 Steiner tree problem instances in graphs, rectilinear Steiner problems: 20 random terminals on a 10,000,000 by 10,000,000 grid reduced to a graph by the GeoSteiner FST generator (Zachariasen and Warme). Format: gzip tar archive; each instance is one .stp text file (STP format version 1.00, https://steinlib.zib.de/format.php): Section Comment, Section Graph with "Nodes n", "Edges m" and m lines "E u v w" (undirected edge, integer weight), Section Terminals with "Terminals t" and t lines "T v", optional Section Coordinates; LF line endings; one case = one instance; the answer is a set of edges Origin: published benchmark instances collected in SteinLib; rectilinear Steiner problems: 20 random terminals on a 10,000,000 by 10,000,000 grid reduced to a graph by the GeoSteiner FST generator (Zachariasen and Warme); set page: showset.php?ES20FST lines 22 to 26: "First there were files with random generated points in the plane on a 10,000,000 by 10,000,000 grid. These points serve as terminals and were converted to rectalinear graphs with L1 edge weights"; recorded tool_fixture - Reference result: published optimal objective value per instance - Reference source: showset.php?ES20FST (set table: the Opt column gives the optimal value; values written in italics would be only best known upper bounds, and none in this table is in italics): 15 of 15 instances have a value. the tree connects all terminals and its cost is never below the published optimum on any of 15 instances, consistent with the table Check: Program: read the agent solution (a list of edges u v of the instance graph); feasible when every listed pair is an edge of Section Graph and the listed edges connect every terminal of Section Terminals into one component; objective = sum of the integer weights of the listed edges (for a pair listed twice in the file, the least weight); pass when feasible and the objective is at most the Opt value of the instance in showset.php?ES20FST (proven optimal, so a lower value means a checker error) Companions (not included): showset.php?ES20FST Dependencies: no upstream code used diff --git a/campaign_files/2026-10-g-solver-instances-2/sudoku-exchange-puzzle-bank/README.md b/campaign_files/2026-10-g-solver-instances-2/sudoku-exchange-puzzle-bank/README.md index 857104d70900e9ff8bf6fa3d99180467856ab907..76a32b848c28712bb07afd76caaf118fa5b34b21 100644 --- a/campaign_files/2026-10-g-solver-instances-2/sudoku-exchange-puzzle-bank/README.md +++ b/campaign_files/2026-10-g-solver-instances-2/sudoku-exchange-puzzle-bank/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Unique solution grid (plus difficulty rating). Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: The SE rating is a secondary difficulty label. No solver is needed at judge time. **Licence (SPDX):** LicenseRef-grantm-Unlicense-data-set @@ -21,7 +19,5 @@ Files: - `g023-01/diabolical.txt`: Sudoku Exchange puzzle bank, diabolical tier (119,681 graded 9x9 sudoku puzzles with unique solutions). Format: plain text, 100 bytes per record: 12-character SHA1 prefix, 81 puzzle digits (0 = blank), rating nn.n, LF; delivered whole as the upstream file (119,681 puzzles, one per line); a case is one line, split on LF. Origin: puzzle bank published as-is by Grant McLean (sudoku-exchange-puzzle-bank). - Reference result: unique solution grid (plus difficulty rating) - Reference source: No solution grid ships. QQWing ensures a unique solution (README). Check: The SE rating is a secondary difficulty label. No solver is needed at judge time. Dependencies: not packed; path and commit filled when the script is committed, before the run; not packed diff --git a/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling-benchmarks/README.md b/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling-benchmarks/README.md index 9dcbe2b6dcf2cad919b0e0671760b59eec47f077..acb3c2f7fb6d8b007ff8f283d601c8d0a04b8a83 100644 --- a/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling-benchmarks/README.md +++ b/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling-benchmarks/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 2 of 2 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Optimum or best known makespan per instance; best makespan per instance. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Program: the agent returns a job permutation; compute the permutation flow shop makespan and require it at most the best_lb_up.txt value for that instance; Program: check each job's operations do not overlap and each machine processes one operation at a time; makespan at most the value for that instance. **Licence (SPDX):** NOASSERTION @@ -21,14 +19,10 @@ Files: - `taillard/problemes.dir/ordonnancement.dir/flowshop.dir/tai20_5.txt`: Taillard permutation flow shop instances ta001 to ta010 (20 jobs, 5 machines) with best makespans. Format: plain text: 10 instances, each a header (jobs, machines, seed, upper bound, lower bound) and a 5 by 20 processing time matrix; 10 cases in one file Origin: Taillard (1993) generator, published instances - Reference result: optimum or best known makespan per instance - Reference source: best_lb_up.txt line 3 "ONLY PERMUTATION SCHEDULES ALLOWED"; line 5 ta001-005: 1278, 1359, 1081, 1293, 1235; line 7 ta006-010: 1195, 1234, 1206, 1230, 1108 Check: Program: the agent returns a job permutation; compute the permutation flow shop makespan and require it at most the best_lb_up.txt value for that instance Companions (not included): taillard/problemes.dir/ordonnancement.dir/flowshop.dir/best_lb_up.txt - `taillard/problemes.dir/ordonnancement.dir/openshop.dir/tai4_4.txt`: Taillard open shop instances tai4_4 (10 instances, 4 jobs, 4 machines) with optimal makespans. Format: plain text: 10 instances, each a header (jobs, machines, seeds, upper bound, lower bound), a processing time matrix and a machine matrix; 10 cases in one file Origin: Taillard (1993) open shop benchmark - Reference result: best makespan per instance - Reference source: best_lb_up.txt line 1 "Summary of lower and upper bounds (or optimum) for open shop problem instances"; lines 2 to 11 give the 10 tai4_4 values (193, 236, 271, 250, 295, 189, 201, 217, 261, 217) Check: Program: check each job's operations do not overlap and each machine processes one operation at a time; makespan at most the value for that instance Companions (not included): taillard/problemes.dir/ordonnancement.dir/openshop.dir/best_lb_up.txt diff --git a/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling/README.md b/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling/README.md index ba03775a6d1a1656683acd63b0a5e9659e88e703..ad532f0954d862c0971d9a8d3ad33df9dff51b33 100644 --- a/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling/README.md +++ b/campaign_files/2026-10-g-solver-instances-2/taillard-scheduling/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 3 of 3 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Optimum (single value in the bound table); optimum or best known value (the bound summary lists one value). Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Program: read the agent's job permutation, require a permutation of 50 jobs, compute the permutation flow shop makespan from the processing times of instance 1; pass when it is at most 2991; Program: read the agent's operation start times for ta01, check job precedence and that no machine runs two operations at once, compute the makespan; pass when it is at most 1231; Program: read the agent's operation start times for instance 1, check that no job and no machine runs two operations at once, compute the makespan; pass when it is at most 435. **Licence (SPDX):** NOASSERTION @@ -21,21 +19,15 @@ Files: - `taillard/problemes.dir/ordonnancement.dir/flowshop.dir/tai50_10.txt`: Permutation flow shop, 50 jobs x 10 machines, ta041 (first of ten); optimum makespan 2991. Format: plain text: header line with jobs, machines, seed, old upper and lower bound, then a machines by jobs matrix of processing times; ten instances one after another Origin: Taillard (1993) benchmarks for basic scheduling problems - Reference result: optimum (single value in the bound table) - Reference source: flowshop.dir/best_lb_up.txt line 25, row 'ta041-045', first column 2991; line 3: only permutation schedules allowed; a single value is the optimum, open instances are written as a range Check: Program: read the agent's job permutation, require a permutation of 50 jobs, compute the permutation flow shop makespan from the processing times of instance 1; pass when it is at most 2991 Companions (not included): taillard/problemes.dir/ordonnancement.dir/flowshop.dir/best_lb_up.txt - `taillard/problemes.dir/ordonnancement.dir/jobshop.dir/tai15_15.txt`: Job shop, 15 jobs x 15 machines, ta01 (first of ten); optimum makespan 1231. Format: plain text: header line with jobs, machines, seeds, bounds; then a Times matrix and a Machines matrix (jobs by operations); ten instances one after another Origin: Taillard (1993) benchmarks for basic scheduling problems - Reference result: optimum (single value in the bound table) - Reference source: jobshop.dir/best_lb_up.txt line 18, row 'ta01-05', first column 1231 (single value = optimum; updates up to 17.06.2013 on lines 3 to 16) Check: Program: read the agent's operation start times for ta01, check job precedence and that no machine runs two operations at once, compute the makespan; pass when it is at most 1231 Companions (not included): taillard/problemes.dir/ordonnancement.dir/jobshop.dir/best_lb_up.txt - `taillard/problemes.dir/ordonnancement.dir/openshop.dir/tai7_7.txt`: Open shop, 7 jobs x 7 machines, first tai7_7 instance; published value 435. Format: plain text: header line with jobs, machines, seeds, upper and lower bound, then processing times and machine matrices; ten instances one after another Origin: Taillard (1993) benchmarks for basic scheduling problems - Reference result: optimum or best known value (the bound summary lists one value) - Reference source: openshop.dir/best_lb_up.txt line 24: 'tai7_7 435' (first tai7_7 row); line 1: 'Summary of lower and upper bounds (or optimum)'; the instance header lower bound is also 435 (tai7_7.txt line 2), so 435 is optimal Check: Program: read the agent's operation start times for instance 1, check that no job and no machine runs two operations at once, compute the makespan; pass when it is at most 435 Companions (not included): taillard/problemes.dir/ordonnancement.dir/openshop.dir/best_lb_up.txt diff --git a/campaign_files/2026-10-g-solver-instances-2/textarena-rushhour/LICENSE b/campaign_files/2026-10-g-solver-instances-2/textarena-rushhour/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..f21af88946a98df9712facf4804df53927614750 --- /dev/null +++ b/campaign_files/2026-10-g-solver-instances-2/textarena-rushhour/LICENSE @@ -0,0 +1,3 @@ +Licence: MIT + +- MIT: https://github.com/LeonGuertler/TextArena/blob/1e3160dea659971f120bf42708fafa8ea3001b81/LICENSE (publisher: Leon Guertler and Bobby Cheng) diff --git a/campaign_files/2026-10-g-solver-instances-2/textarena-rushhour/README.md b/campaign_files/2026-10-g-solver-instances-2/textarena-rushhour/README.md new file mode 100644 index 0000000000000000000000000000000000000000..54dde0a61296e9e85bdf9d22a8d9516c955dbbcb --- /dev/null +++ b/campaign_files/2026-10-g-solver-instances-2/textarena-rushhour/README.md @@ -0,0 +1,22 @@ +# textarena-rushhour + +**What this is:** 1 unit (2 files) published by Leon Guertler and Bobby Cheng. Collection theme: Seeded puzzle generators (Simon Tatham puzzle collection class, with generation command, seed and solution). + +**What one unit is:** Each unit is a few named upstream files, delivered together, kept at its upstream path. Format: JSONL of generated boards with solutions, plus a generation record. + +**Where it comes from:** Upstream: LeonGuertler/TextArena. Kind: generated data. + +**Self-contained:** The unit is self-contained. Dependencies named upstream are listed per unit in the notes. + +**Tools to open and check:** Check: Judged by simulation from the stored board alone (the decision 6 of 2026-10-06 pattern used for LightsOut): replay the agent's moves on the stored layout with the movement rules of env.py step (line 317), rejecting any move into an occupied cell or off the board; correct iff the target car reaches the exit (win condition at line 341). TextArena is not needed to judge. Unique solution: not required, since any legal move list that solves the board is accepted. + +**Licence (SPDX):** MIT + +**Provenance:** https://github.com/LeonGuertler/TextArena + +Units: + +- `/` (2 files): TextArena RushHour: seeded sliding-car puzzle generated with a solvability check. + Format: JSONL of generated boards with solutions, plus a generation record + Check: Judged by simulation from the stored board alone (the decision 6 of 2026-10-06 pattern used for LightsOut): replay the agent's moves on the stored layout with the movement rules of env.py step (line 317), rejecting any move into an occupied cell or off the board; correct iff the target car reaches the exit (win condition at line 341). TextArena is not needed to judge. Unique solution: not required, since any legal move list that solves the board is accepted. + Dependencies: textarena at git commit 1e3160dea659971f120bf42708fafa8ea3001b81 (pyproject version 0.7.3; PyPI latest is 0.7.4 of 2025-10-16, so the commit is used, not PyPI), installed from the commit; generator only, not packed; openai, rich, nltk, chess, python-dotenv, requests, websockets (pyproject dependencies, unpinned there) resolved in the pinned image; full pip freeze in the generation record; not packed; Python 3.12; not packed; build-time tools (decision 8 of 2026-10-06): pip as shipped in the pinned base image, setuptools 80.9.0 and wheel 0.45.1 installed before the generator; their versions, with javac/gcc/cmake where present, go into the build_tools field of every generation record; not packed diff --git a/campaign_files/2026-10-gt-browser-tests-2/syncthing-config-tests/README.md b/campaign_files/2026-10-gt-browser-tests-2/syncthing-config-tests/README.md index ae692e147063618b66bf328c9dd0cc4ec5137cdf..d5a5af48f5d0fec3171d30b47ab7f97880d451e9 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/syncthing-config-tests/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/syncthing-config-tests/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether each file needs other upstream files is not stated. 4 of 4 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_settings_values. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** edit config.xml (~/.local/state/syncthing/config.xml or ~/.config/syncthing/config.xml) with a text editor or a script (Python xml.etree.ElementTree) while Syncthing is stopped, or change the running instance with the syncthing cli subcommand (`syncthing cli config options ...`, `syncthing cli config devices ...`, `syncthing cli config folders ...`) or the REST endpoints /rest/config/*, which make Syncthing write config.xml **Licence (SPDX):** MPL-2.0 @@ -21,8 +19,6 @@ Files: - `lib/config/testdata/overridenvalues.xml`: Syncthing options changed away from their defaults in the Settings dialog: listen address, announce servers and ports, bandwidth limits, NAT timings, minimum free disk 5.2 percent, usage and crash reporting URLs, audit log and connection priorities, plus a default folder path, stored in Syncthing's config.xml. Format: Syncthing config.xml (XML, with options, folder, device, defaults elements) Origin: config.xml fixture of Syncthing's config package tests (lib/config/testdata), loaded by config_test.go through copyAndLoad (tool_fixture) - Reference result: expected_settings_values - Reference source: TestOverriddenValues in lib/config/config_test.go lines 259-327: the OptionsConfiguration literal (about 45 fields) and the default folder path /media/syncthing, compared with the loaded configuration (messagediff.PrettyDiff or direct comparison); device IDs at lines 39-42 Check: Final configuration state, no screenshot needed. Parse the config.xml the agent leaves with an XML parser (or load it with Syncthing at the pinned commit) and compare the judged values with the expected values of the test, all exactly (upstream compares with messagediff.PrettyDiff and !=, no tolerance). URUniqueID is random and excluded, as the test does (line 318); minHomeDiskFree is compared as value 5.2 with unit % exactly. Tools: edit config.xml (~/.local/state/syncthing/config.xml or ~/.config/syncthing/config.xml) with a text editor or a script (Python xml.etree.ElementTree) while Syncthing is stopped, or change the running instance with the syncthing cli subcommand (`syncthing cli config options ...`, `syncthing cli config devices ...`, `syncthing cli config folders ...`) or the REST endpoints /rest/config/*, which make Syncthing write config.xml Companions (not included): lib/config/config_test.go; LICENSE @@ -30,8 +26,6 @@ Files: - `lib/config/testdata/deviceaddressesstatic.xml`: Three remote devices each given two static addresses (IPv4 and IPv6, with and without ports) in the Edit Device dialog, stored in Syncthing's config.xml. Format: Syncthing config.xml (XML, with options, folder, device, defaults elements) Origin: config.xml fixture of Syncthing's config package tests (lib/config/testdata), loaded by config_test.go through copyAndLoad (tool_fixture) - Reference result: expected_settings_values - Reference source: TestDeviceAddressesStatic in lib/config/config_test.go lines 418-459: the expected map of four DeviceConfiguration values: three devices with their tcp:// addresses, and the local device4 that Syncthing adds with address dynamic and compression metadata, compared with the loaded configuration (messagediff.PrettyDiff or direct comparison); device IDs at lines 39-42 Check: Final configuration state, no screenshot needed. Parse the config.xml the agent leaves with an XML parser (or load it with Syncthing at the pinned commit) and compare the judged values with the expected values of the test, all exactly (upstream compares with messagediff.PrettyDiff and !=, no tolerance). Syncthing adds the tcp:// scheme to bare addresses when it loads the file (the expected map in lines 418-459 lists tcp://192.0.2.1 for the stored 192.0.2.1), so a bare host or host:port in the agent's file counts as the tcp:// form; the local device4 entry and its Name (set from the host name at load) are added by Syncthing and are not judged. Tools: edit config.xml (~/.local/state/syncthing/config.xml or ~/.config/syncthing/config.xml) with a text editor or a script (Python xml.etree.ElementTree) while Syncthing is stopped, or change the running instance with the syncthing cli subcommand (`syncthing cli config options ...`, `syncthing cli config devices ...`, `syncthing cli config folders ...`) or the REST endpoints /rest/config/*, which make Syncthing write config.xml Companions (not included): lib/config/config_test.go; LICENSE @@ -39,8 +33,6 @@ Files: - `lib/config/testdata/versioningconfig.xml`: A folder with simple file versioning and two versioning parameters set in the Edit Folder dialog, stored in Syncthing's config.xml. Format: Syncthing config.xml (XML, with options, folder, device, defaults elements) Origin: config.xml fixture of Syncthing's config package tests (lib/config/testdata), loaded by config_test.go through copyAndLoad (tool_fixture) - Reference result: expected_settings_values - Reference source: TestVersioningConfig in lib/config/config_test.go lines 461-483: versioning type "simple" with exactly two params foo=bar and baz=quux on folder "test", compared with the loaded configuration (messagediff.PrettyDiff or direct comparison); device IDs at lines 39-42 Check: Final configuration state, no screenshot needed. Parse the config.xml the agent leaves with an XML parser (or load it with Syncthing at the pinned commit) and compare the judged values with the expected values of the test, all exactly (upstream compares with messagediff.PrettyDiff and !=, no tolerance). Tools: edit config.xml (~/.local/state/syncthing/config.xml or ~/.config/syncthing/config.xml) with a text editor or a script (Python xml.etree.ElementTree) while Syncthing is stopped, or change the running instance with the syncthing cli subcommand (`syncthing cli config options ...`, `syncthing cli config devices ...`, `syncthing cli config folders ...`) or the REST endpoints /rest/config/*, which make Syncthing write config.xml Companions (not included): lib/config/config_test.go; LICENSE @@ -48,8 +40,6 @@ Files: - `lib/config/testdata/pullorder.xml`: Eight folders each set to a different file pull order (random, alphabetic, smallest first, largest first, oldest first, newest first, and two left to the default), stored in Syncthing's config.xml. Format: Syncthing config.xml (XML, with options, folder, device, defaults elements) Origin: config.xml fixture of Syncthing's config package tests (lib/config/testdata), loaded by config_test.go through copyAndLoad (tool_fixture) - Reference result: expected_settings_values - Reference source: TestPullOrder in lib/config/config_test.go lines 697-748: the expected list of (folder, PullOrder) pairs for f1 to f8; f1 empty and f4 unknown value both fall back to random, compared with the loaded configuration (messagediff.PrettyDiff or direct comparison); device IDs at lines 39-42 Check: Final configuration state, no screenshot needed. Parse the config.xml the agent leaves with an XML parser (or load it with Syncthing at the pinned commit) and compare the judged values with the expected values of the test, all exactly (upstream compares with messagediff.PrettyDiff and !=, no tolerance). An empty or unknown order value counts as random, as the expected list says for f1 and f4. Tools: edit config.xml (~/.local/state/syncthing/config.xml or ~/.config/syncthing/config.xml) with a text editor or a script (Python xml.etree.ElementTree) while Syncthing is stopped, or change the running instance with the syncthing cli subcommand (`syncthing cli config options ...`, `syncthing cli config devices ...`, `syncthing cli config folders ...`) or the REST endpoints /rest/config/*, which make Syncthing write config.xml Companions (not included): lib/config/config_test.go; LICENSE diff --git a/campaign_files/2026-10-gt-browser-tests-2/talon-signature-stripped/README.md b/campaign_files/2026-10-gt-browser-tests-2/talon-signature-stripped/README.md index d43675c5b59ea1c0d4b03e8f1f99e19b67fe8aad..9b8dd72665f964033a7043863bf4120b077caf59 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/talon-signature-stripped/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/talon-signature-stripped/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_files. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** talon.signature.extract (Python, ML classifier) or a hand-written signature heuristic in Python; a script reads the body and sender files and writes the stripped text and the signature to two files **Licence (SPDX):** Apache-2.0 @@ -21,8 +19,6 @@ Files: - `tests/fixtures/signature/emails/stripped/long_body`: Strip the signature block from a long plain-text message given its sender. Format: plain-text message body plus a sender file; expected signature as a text file Origin: talon signature extraction test fixture (tool_fixture) - Reference result: expected_output_files - Reference source: tests/fixtures/signature/emails/stripped/long_signature is the expected signature; the expected message text is the body with that signature cut off (extraction_test.py lines 47-50) Check: Final file state: the agent writes the extracted signature and the remaining text to files; the signature file stripped of outer whitespace must equal long_signature stripped, and the text file stripped must equal the body stripped minus the last len(long_signature) characters, stripped. No screenshot Tools: talon.signature.extract (Python, ML classifier) or a hand-written signature heuristic in Python; a script reads the body and sender files and writes the stripped text and the signature to two files Companions (not included): tests/fixtures/signature/emails/stripped/long_sender; tests/fixtures/signature/emails/stripped/long_signature; tests/signature/extraction_test.py; tests/signature/bruteforce_test.py diff --git a/campaign_files/2026-10-gt-browser-tests-2/talon-standard-replies/README.md b/campaign_files/2026-10-gt-browser-tests-2/talon-standard-replies/README.md index 9fb5944242d68a912ff2da8b38ba5ec2d89f4db0..e4d99549336b28487e3682fc3a00cbf219a253ea 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/talon-standard-replies/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/talon-standard-replies/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_file. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Python email.message_from_file plus email.iterators.typed_subpart_iterator to take the text/plain part, then talon.quotations.extract_from_plain or a hand-written quote stripper; the script writes the reply text to a file **Licence (SPDX):** Apache-2.0 @@ -21,8 +19,6 @@ Files: - `tests/fixtures/standard_replies/sparrow.eml`: Extract only the new reply from a Sparrow client message. Format: RFC 5322 multipart/alternative message (.eml); expected reply as a text file Origin: talon quotation test fixture (tool_fixture) - Reference result: expected_output_file - Reference source: tests/fixtures/standard_replies/sparrow_reply_text, read by test_standard_replies (text_quotations_test.py lines 763-783) Check: Final file state: take the text/plain part of sparrow.eml, strip quotations and the signature, write the reply to a file; the file must equal sparrow_reply_text after .strip() exactly. No screenshot Tools: Python email.message_from_file plus email.iterators.typed_subpart_iterator to take the text/plain part, then talon.quotations.extract_from_plain or a hand-written quote stripper; the script writes the reply text to a file Companions (not included): tests/fixtures/standard_replies/sparrow_reply_text; tests/text_quotations_test.py; tests/quotations_test.py diff --git a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-enterprisepolicies-tests/README.md b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-enterprisepolicies-tests/README.md index 762044d7ce55d21a2f4b70b131e5f2f67cfe0760..16978b4c15435d286681f7b1671f45af4065ed09 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-enterprisepolicies-tests/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-enterprisepolicies-tests/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_preference_values. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** write distribution/policies.json next to the Thunderbird binary, or into the system policies directory, with a text editor or script; start Thunderbird headless (`thunderbird --headless -profile DIR`) so the policy engine applies it, then read the resulting prefs back; the policies.json itself can also be checked with any JSON parser **Licence (SPDX):** CC0-1.0 AND MPL-2.0 @@ -21,8 +19,6 @@ Files: - `mail/components/enterprisepolicies/tests/xpcshell/test_simple_pref_policies.js`: Thunderbird enterprise policies: write policies.json so that 21 policies (34 test entries) set and lock the expected mail and network preferences. Format: JavaScript array POLICIES_TESTS of {policies, lockedPrefs, unlockedPrefs}: the policies.json object as input, the expected pref values (and whether they are locked) as answer Origin: Thunderbird enterprise-policy xpcshell test written by Mozilla developers (tool_fixture) - Reference result: expected_preference_values - Reference source: mail/components/enterprisepolicies/tests/xpcshell/test_simple_pref_policies.js lines 17-431: 34 test entries over 21 policies (for example DisableChat, DisableFileLink, DisableMessageForwardingFilters, DisableQRExport, Authentication, DNSOverHTTPS, SSLVersionMin, DisabledCiphers), each with policies, lockedPrefs and unlockedPrefs giving the exact pref values; lines 434-453 apply them with checkLockedPref and checkUnlockedPref (head.js lines 82-88, which call EnterprisePolicyTesting.checkPolicyPref, lines 83-97) Check: Final configuration state, no screenshot needed: the agent writes policies.json for the wanted settings; start Thunderbird headless with it and read each listed pref: a lockedPrefs entry must be locked and equal, an unlockedPrefs entry equal and unlocked (EnterprisePolicyTesting.sys.mjs lines 83-97, Assert.equal). Exact value equality, no tolerance. Static judge, no screenshot needed: parse the policies.json the agent wrote and check that its "policies" object equals the policies object of the chosen test task (JSON equality, key order ignored). Tools: write distribution/policies.json next to the Thunderbird binary, or into the system policies directory, with a text editor or script; start Thunderbird headless (`thunderbird --headless -profile DIR`) so the policy engine applies it, then read the resulting prefs back; the policies.json itself can also be checked with any JSON parser Companions (not included): mail/components/enterprisepolicies/tests/xpcshell/head.js; toolkit/components/enterprisepolicies/tests/EnterprisePolicyTesting.sys.mjs; LICENSE diff --git a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ics-import-latin1/README.md b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ics-import-latin1/README.md index 4189f92213e853dc35e485cc66f2ae7ca7fcbe2d..6243776f0f9f45253a225052aa932f498354aa3a 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ics-import-latin1/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ics-import-latin1/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_values_in_test_code. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Python icalendar after decoding the bytes as ISO-8859-1 (or iconv to UTF-8 first): a script writes the titles of the events in the range 2019-01-01 to 2019-01-02 to a UTF-8 file **Licence (SPDX):** MPL-2.0 @@ -21,8 +19,6 @@ Files: - `calendar/test/unit/data/importLatin1.ics`: Import an ISO-8859-1 calendar file and keep non-ASCII titles intact. Format: iCalendar (.ics) in ISO-8859-1 plus its UTF-8 twin; expected values in an xpcshell test Origin: Thunderbird calendar test data (tool_fixture) - Reference result: expected_values_in_test_code - Reference source: calendar/test/unit/test_CalendarFileImporter.js lines 49-77: importLatin1.ics parses to 4 items, all 4 fall in 20190101T000000 to 20190102T000000, and one has the title "Event Två"; import.ics is the UTF-8 twin with the same expectations (lines 16-44) Check: Final file state: the UTF-8 output lists exactly 4 event titles and contains "Event Två" with the å decoded correctly (no mojibake). No screenshot Tools: Python icalendar after decoding the bytes as ISO-8859-1 (or iconv to UTF-8 first): a script writes the titles of the events in the range 2019-01-01 to 2019-01-02 to a UTF-8 file Companions (not included): calendar/test/unit/data/import.ics; calendar/test/unit/test_CalendarFileImporter.js diff --git a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ispdb/README.md b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ispdb/README.md index 787cfafa4aabd4ea5d48a7c3feaaeeee245b0b66..26233bba6328fee9e3ae3080c6f81826450b85ce 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ispdb/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-ispdb/README.md @@ -8,8 +8,6 @@ **Self-contained:** Each file is self-contained. 2 of 2 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_account_server_settings. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** write the account into the Thunderbird profile's prefs.js (mail.server.serverN.* and mail.smtpserver.smtpN.* user_pref lines) with a text editor or script, or run Thunderbird headless with an autoconfig XML in the profile's isp/ directory; the XML itself is read with any XML parser **Licence (SPDX):** MPL-2.0 @@ -21,8 +19,6 @@ Files: - `ispdb/posteo.de.xml`: Mail account setup for a Posteo address: IMAP, POP3 and SMTP servers with SSL or STARTTLS and normal password. Format: ISPDB clientConfig 1.1 XML (one provider: domains, incoming IMAP and POP3 servers, outgoing SMTP servers with hostname, port, socketType, username template, authentication) Origin: real mail provider settings collected for Thunderbird account autoconfiguration (real); the ISPDB entry cites the provider help page at line 98 - Reference result: expected_account_server_settings - Reference source: ispdb/posteo.de.xml lines 56-97 (the server blocks): imap posteo.de 993 SSL or 143 STARTTLS, pop3 posteo.de 995 SSL or 110 STARTTLS, smtp posteo.de 465 SSL or 587 STARTTLS, username %EMAILADDRESS%, authentication password-cleartext; domains at lines 4-53 Check: Final configuration state: read the agent-created account from the Thunderbird profile prefs.js (mail.server.serverN.hostname, port, socketType, authMethod, userName, type and the mail.smtpserver.smtpN.* lines) and check that it equals one of the listed server alternatives for the chosen protocol, with userName the full address; exact equality of strings and ports, socketType and authentication mapped to the Thunderbird pref codes. No screenshot needed Tools: write the account into the Thunderbird profile's prefs.js (mail.server.serverN.* and mail.smtpserver.smtpN.* user_pref lines) with a text editor or script, or run Thunderbird headless with an autoconfig XML in the profile's isp/ directory; the XML itself is read with any XML parser Companions (not included): LICENSE @@ -30,8 +26,6 @@ Files: - `ispdb/yahoo.com.xml`: Mail account setup for a Yahoo address: IMAP, POP3 and SMTP over SSL with OAuth2 or password authentication. Format: ISPDB clientConfig 1.1 XML (one provider: domains, incoming IMAP and POP3 servers, outgoing SMTP servers with hostname, port, socketType, username template, authentication) Origin: real mail provider settings collected for Thunderbird account autoconfiguration (real) - Reference result: expected_account_server_settings - Reference source: ispdb/yahoo.com.xml lines 31-62: imap imap.mail.yahoo.com 993 SSL (or export.imap.mail.yahoo.com), pop3 pop.mail.yahoo.com 995 SSL, smtp smtp.mail.yahoo.com 465 SSL, username %EMAILADDRESS%, authentication OAuth2 or password-cleartext; domains (yahoo.com, ymail.com, rocketmail.com and others) at lines 4-26 Check: Final configuration state: read the agent-created account from the Thunderbird profile prefs.js (mail.server.serverN.* and mail.smtpserver.smtpN.* lines) and check hostname, port, socketType, authMethod and userName against the listed alternatives; exact equality of strings and ports, socketType and authentication mapped to the Thunderbird pref codes. No screenshot needed Tools: write the account into the Thunderbird profile's prefs.js (mail.server.serverN.* and mail.smtpserver.smtpN.* user_pref lines) with a text editor or script, or run Thunderbird headless with an autoconfig XML in the profile's isp/ directory; the XML itself is read with any XML parser Companions (not included): LICENSE diff --git a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-mbox-to-maildir/README.md b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-mbox-to-maildir/README.md index faf0a35a5d790d8291d3c7c33b042d669aa6eee4..4a5c2414a183b3caa47796389c7cf0d5f737243e 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-mbox-to-maildir/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-mbox-to-maildir/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_values_in_test_code. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Python mailbox (stdlib): a script reads the mbox file (three variants: Thunderbird modern with space-escaped From lines, mboxrd with >From escaping, old unquoted) and writes each message to a Maildir; the Maildir cur/ directory is counted **Licence (SPDX):** MPL-2.0 @@ -21,8 +19,6 @@ Files: - `mailnews/test/data/mbox_unquoted`: Convert mbox mailboxes in three From-line conventions to Maildir without losing or splitting messages. Format: mbox files (three variants); expected counts in an xpcshell test Origin: Thunderbird mailnews test data (tool_fixture) - Reference result: expected_values_in_test_code - Reference source: mailnews/base/test/unit/test_mailstoreConverter.js lines 57-59 and 124-147: each of mbox_modern, mbox_mboxrd and mbox_unquoted converts to a Maildir whose Inbox/cur directory holds exactly 2 messages; readme.txt lines 66-103 describe the three separator conventions Check: Final file state: after conversion the Maildir directory has a cur/ subdirectory with exactly 2 message files per input mbox (a body line beginning "From " must not split a message). No screenshot Tools: Python mailbox (stdlib): a script reads the mbox file (three variants: Thunderbird modern with space-escaped From lines, mboxrd with >From escaping, old unquoted) and writes each message to a Maildir; the Maildir cur/ directory is counted Companions (not included): mailnews/test/data/mbox_modern; mailnews/test/data/mbox_mboxrd; mailnews/base/test/unit/test_mailstoreConverter.js; mailnews/test/data/readme.txt diff --git a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex1/README.md b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex1/README.md index 69fd59e70cc01a457895cd7e8932ea0766bd8037..97d5bec7ca84f50d21374ce7b9f60e351673af2b 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex1/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex1/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_file. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Python email (stdlib) plus an HTML to text step (html.parser): a script picks the text/html part the client would show, strips the tags and writes the text to a file **Licence (SPDX):** MPL-2.0 @@ -21,8 +19,6 @@ Files: - `mailnews/test/data/multipart-complex1`: Extract the displayed body text of a nested multipart message with HTML stripped. Format: RFC 5322 multipart message (no extension); expected output in a .out file Origin: Thunderbird mailnews test data (tool_fixture) - Reference result: expected_output_file - Reference source: mailnews/test/data/multipart-complex1.out ("This part should be returned."), the expected output of GetMsgTextFromStream for test case line 51 of mailnews/base/test/unit/test_getMsgTextFromStream.js Check: Final file state: the agent writes the extracted body text to a file; after trimming surrounding white space it must equal the contents of multipart-complex1.out exactly. No screenshot Tools: Python email (stdlib) plus an HTML to text step (html.parser): a script picks the text/html part the client would show, strips the tags and writes the text to a file Companions (not included): mailnews/test/data/multipart-complex1.out; mailnews/base/test/unit/test_getMsgTextFromStream.js; mailnews/test/data/readme.txt diff --git a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex2/README.md b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex2/README.md index 6d214c40f1cf81a4bda1b55b358aadac65b6e3eb..47fb53fb549e77fe380352151bbec6ffa74eb8dc 100644 --- a/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex2/README.md +++ b/campaign_files/2026-10-gt-browser-tests-2/thunderbird-msgtext-complex2/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_file. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Python email (stdlib): a script walks the MIME tree, picks the text/plain part the client would show and writes it to a file **Licence (SPDX):** MPL-2.0 @@ -21,8 +19,6 @@ Files: - `mailnews/test/data/multipart-complex2`: Find the plain text body the client shows in a nested multipart message. Format: RFC 5322 multipart message (no extension); expected output in a .out file Origin: Thunderbird mailnews test data (tool_fixture) - Reference result: expected_output_file - Reference source: mailnews/test/data/multipart-complex2.out ("This is the correct answer."), the expected output for test cases lines 52-53 of mailnews/base/test/unit/test_getMsgTextFromStream.js Check: Final file state: the extracted text, trimmed of surrounding white space, must equal multipart-complex2.out exactly (the same with and without HTML stripping). No screenshot Tools: Python email (stdlib): a script walks the MIME tree, picks the text/plain part the client would show and writes it to a file Companions (not included): mailnews/test/data/multipart-complex2.out; mailnews/base/test/unit/test_getMsgTextFromStream.js; mailnews/test/data/readme.txt diff --git a/campaign_files/2026-10-hw-eda-tests-2/taxi/README.md b/campaign_files/2026-10-hw-eda-tests-2/taxi/README.md index 9581f9eaba9349174e31ef4812b0df5853f3023f..1001f7282548882fef2527c85e9f59145f5de94b 100644 --- a/campaign_files/2026-10-hw-eda-tests-2/taxi/README.md +++ b/campaign_files/2026-10-hw-eda-tests-2/taxi/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each unit is used together with other files from the same upstream. 7 of 7 units name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per unit in the notes. Dependencies named upstream are listed per unit in the notes. -**Answers:** No answer files are included. Reference result: self_checking_testbench_pass_fail. Where the reference result comes from and how a solution is checked is written per unit in the notes. - **Tools to open and check:** Tools are named in the check written per unit in the notes. **Licence (SPDX):** CERN-OHL-S-2.0 AND MIT AND CERN-OHL-S-2.0 @@ -21,56 +19,42 @@ Units: - `src/axis/tb/taxi_axis_demux/` (3 files): Taxi axis_demux: AXI4-Stream demultiplexer (tdest routed) with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module with a SystemVerilog wrapper top; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_axis_demux.py sends frames of lengths 1 to 4*byte_lanes, 512 and 64 single-byte frames through the cocotbext-axi AxiStreamSource to a chosen output port and asserts on each received frame that tdata, tid and the masked tdest match the sent frame and tuser is clear (lines 128-131), then that the sink is empty (line 133); cocotb.parametrize at lines 82-89 crosses idle and backpressure inserters with the output port Check: Build in a sandbox with the versions listed in dependencies and run make inside src/axis/tb/taxi_axis_demux/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/axis/rtl/taxi_axis_demux.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/axis/rtl/taxi_axis_demux.sv; src/axis/rtl/taxi_axis_if.sv Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; cocotbext-axi==0.1.28 (pinned upstream in tox.ini line 16 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed - `src/axis/tb/taxi_axis_cobs_decode/` (3 files): Taxi axis_cobs_decode: AXI4-Stream COBS decoder with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module with a SystemVerilog wrapper top; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_axis_cobs_decode.py COBS-encodes payloads with its own Python reference encoder, sends them to the DUT and asserts that the decoded rx_frame.tdata equals the original payload and tuser is clear (lines 180-181), then that the sink is empty (line 183); cocotb.parametrize at lines 154-160 crosses four payload generators (all zero, nonzero incrementing, zero framed, PRBS) with idle and backpressure inserters Check: Build in a sandbox with the versions listed in dependencies and run make inside src/axis/tb/taxi_axis_cobs_decode/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/axis/rtl/taxi_axis_cobs_decode.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/axis/rtl/taxi_axis_cobs_decode.sv; src/axis/rtl/taxi_axis_if.sv Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; cocotbext-axi==0.1.28 (pinned upstream in tox.ini line 16 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed - `src/apb/tb/taxi_apb_ram/` (3 files): Taxi apb_ram: APB RAM with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module with a SystemVerilog wrapper top; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_apb_ram.py writes and reads byte strings of every length 1 to 2*byte_lanes-1 at every offset through the cocotbext-axi ApbMaster and asserts the read-back data (with guard bytes for the write test) equals what was written (lines 84, 114 and 145: write test, read test, stress workers) Check: Build in a sandbox with the versions listed in dependencies and run make inside src/apb/tb/taxi_apb_ram/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/apb/rtl/taxi_apb_ram.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/apb/rtl/taxi_apb_ram.sv; src/apb/rtl/taxi_apb_if.sv Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; cocotbext-axi==0.1.28 (pinned upstream in tox.ini line 16 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed - `src/lss/tb/taxi_i2c_master/` (3 files): Taxi i2c_master: I2C master with AXI-Stream command and data interfaces with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module with a SystemVerilog wrapper top; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_i2c_master.py drives commands and data through cocotbext-axi stream sources against cocotbext-i2c I2cMemory models on the bus; run_test_write asserts the memory contents equal the written bytes (line 116), run_test_read asserts the read data equal the stored bytes (line 142), run_test_nack addresses a missing device Check: Build in a sandbox with the versions listed in dependencies and run make inside src/lss/tb/taxi_i2c_master/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/lss/rtl/taxi_i2c_master.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/lss/rtl/taxi_i2c_master.sv; src/axis/rtl/taxi_axis_if.sv Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; cocotbext-axi==0.1.28 (pinned upstream in tox.ini line 16 at commit cc70b270); open source, not packed; cocotbext-i2c==0.1.2 (pinned upstream in tox.ini line 20 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed - `src/lfsr/tb/taxi_lfsr_scramble/` (2 files): Taxi lfsr_scramble: LFSR self-synchronizing scrambler (64b/66b polynomial) with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_lfsr_scramble.py picks the Python reference scrambler by parameter (scramble_64b66b for LFSR_W 58, lines 95-102), scrambles a 1024-byte incrementing block (line 118) with it and asserts each DUT output word equals the reference (line 136) Check: Build in a sandbox with the versions listed in dependencies and run make inside src/lfsr/tb/taxi_lfsr_scramble/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/lfsr/rtl/taxi_lfsr_scramble.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/lfsr/rtl/taxi_lfsr_scramble.sv; src/lfsr/rtl/taxi_lfsr.sv Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed - `src/lfsr/tb/taxi_lfsr_prbs_gen/` (2 files): Taxi lfsr_prbs_gen: LFSR PRBS generator (PRBS31) with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_lfsr_prbs_gen.py picks the Python reference generator by polynomial (prbs31 for 31'h10000001, lines 101-109) and asserts each of 512 DUT output words equals the reference (loop at line 130, assert at line 136); the scrambler test (lines 141-145) is skipped for this polynomial because no scrambler reference matches it Check: Build in a sandbox with the versions listed in dependencies and run make inside src/lfsr/tb/taxi_lfsr_prbs_gen/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/lfsr/rtl/taxi_lfsr_prbs_gen.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/lfsr/rtl/taxi_lfsr_prbs_gen.sv; src/lfsr/rtl/taxi_lfsr.sv Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed - `src/eth/tb/taxi_gmii_basex_enc/` (2 files): Taxi gmii_basex_enc: GMII to 1000BASE-X (8b/10b) encoder with autonegotiation with its cocotb testbench. Format: cocotb testbench directory: Makefile (cocotb makefile flow, COCOTB_TOPLEVEL, VERILOG_SOURCES, PARAM_* overrides) and a cocotb 2.0 Python test module; DUT in synthesizable SystemVerilog Origin: self-checking cocotb testbench written by the IP author for his own RTL module (tool_fixture); taxi is the maintained successor of the alexforencich verilog-* repositories - Reference result: self_checking_testbench_pass_fail - Reference source: test_taxi_gmii_basex_enc.py sends Ethernet frames through a cocotbext-eth GMII source, decodes the DUT's 8b/10b output with the BASE-X sink model of basex.py and asserts payload, FCS and no error on every frame (lines 114-116); run_test_an sends autonegotiation configs and asserts the decoded ability, ack and idle matches (lines 140-143 and 150-152) Check: Build in a sandbox with the versions listed in dependencies and run make inside src/eth/tb/taxi_gmii_basex_enc/ (the Makefile's own parameter set; the pytest matrix at the bottom of the test file is not part of the judge); the unit passes when every cocotb test reports PASS and the results.xml has no failure. For a write-the-RTL task, hide or replace the DUT file src/eth/rtl/taxi_gmii_basex_enc.sv and keep the testbench, the interface or model companions and the Makefile unchanged. Companions (not included): LICENSE; src/eth/rtl/taxi_gmii_basex_enc.sv; src/eth/tb/basex.py Dependencies: Python 3 (CI .github/workflows/regression-tests.yml runs on ubuntu-24.04; the Python version is not pinned upstream); open source, not packed; pytest==8.3.4 (pinned upstream in tox.ini line 10 at commit cc70b270); open source, not packed; pytest-xdist==3.6.1 (pinned upstream in tox.ini line 11 at commit cc70b270); open source, not packed; pytest-split==0.10.0 (pinned upstream in tox.ini line 12 at commit cc70b270); open source, not packed; cocotb==2.0.1 (pinned upstream in tox.ini line 13 at commit cc70b270); open source, not packed; cocotb-bus==0.3.0 (pinned upstream in tox.ini line 14 at commit cc70b270); open source, not packed; cocotb-test==0.2.6 (pinned upstream in tox.ini line 15 at commit cc70b270); open source, not packed; cocotbext-eth==0.1.28 (pinned upstream in tox.ini line 18 at commit cc70b270); open source, not packed; Verilator v5.050 (pinned upstream in .github/workflows/regression-tests.yml lines 22-24, install-verilator-action with `version: v5.050`, at commit cc70b270); open source, not packed diff --git a/campaign_files/2026-10-hw-eda-tests-2/tgingold-cern-cheby-testfiles-features/README.md b/campaign_files/2026-10-hw-eda-tests-2/tgingold-cern-cheby-testfiles-features/README.md index 268994e3db10b85d1c7a1706eee552b5efb1efb3..93555453ede9bc77bceb48b54c23790c02e7f96f 100644 --- a/campaign_files/2026-10-hw-eda-tests-2/tgingold-cern-cheby-testfiles-features/README.md +++ b/campaign_files/2026-10-hw-eda-tests-2/tgingold-cern-cheby-testfiles-features/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether each file needs other upstream files is not stated. 2 of 2 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: register_map_header_values. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Task: from blkprefix5.cheby produce both C headers. Judge: lay out the memory map (register widths, blocks aligned to a power of two, block-prefix False emitting no block define) and compare every address and size define in both headers by exact integer equality. Task: from memwide_ua.cheby produce both C headers. Judge: lay out the memory (12 byte element padded to 16, memsize 128 so 8 elements, placed at 0x80, block size 256) and compare every address and size define in both headers by exact integer equality. **Licence (SPDX):** GPL-3.0-or-later @@ -21,16 +19,12 @@ Files: - `testfiles/features/blkprefix5.cheby`: Cheby memory map with nested blocks and block-prefix false: generate the C headers with matching addresses, sizes and power of two block alignment. Format: cheby YAML memory map; C headers Origin: feature test of the cheby generator with its committed expected outputs - Reference result: register_map_header_values - Reference source: testfiles/features/blkprefix5.h (committed expected output of the cheby test suite); e.g. line 4 `#define BLKPREFIX3_SIZE 48 /* 0x30 */` and line 46 `#define BLKPREFIX3_B2_R2 0x28UL` Check: Task: from blkprefix5.cheby produce both C headers. Judge: lay out the memory map (register widths, blocks aligned to a power of two, block-prefix False emitting no block define) and compare every address and size define in both headers by exact integer equality. Companions (not included): LICENSE; testfiles/features/blkprefix5.h; testfiles/features/blkprefix5_union.h Dependencies: cheby: only needed to regenerate; Ruby 3.x standard library Psych to read the cheby YAML in the judge; Python 3.9 or later standard library for the comparison judge - `testfiles/features/memwide_ua.cheby`: Cheby memory map with a wide memory of 12 byte elements: generate the C headers with matching element padding, memory address and block size. Format: cheby YAML memory map; C headers Origin: feature test of the cheby generator with its committed expected outputs - Reference result: register_map_header_values - Reference source: testfiles/features/memwide_ua.h (committed expected output of the cheby test suite); e.g. line 3 `#define MEMWIDE_UA_SIZE 256 /* 0x100 */`, line 10 `#define MEMWIDE_UA_TS 0x80UL`, line 11 `#define MEMWIDE_UA_TS_SIZE 16 /* 0x10 */` Check: Task: from memwide_ua.cheby produce both C headers. Judge: lay out the memory (12 byte element padded to 16, memsize 128 so 8 elements, placed at 0x80, block size 256) and compare every address and size define in both headers by exact integer equality. Companions (not included): LICENSE; testfiles/features/memwide_ua.h; testfiles/features/memwide_ua-consts.h Dependencies: cheby: only needed to regenerate; Ruby 3.x standard library Psych to read the cheby YAML in the judge; Python 3.9 or later standard library for the comparison judge diff --git a/campaign_files/2026-10-m-issue-trackers-2/squid-users/README.md b/campaign_files/2026-10-m-issue-trackers-2/squid-users/README.md index 534e5fb4a7013341ec69ac4f0685fe76b3259192..f8405fee1df48835badef6911894441805e0ce71 100644 --- a/campaign_files/2026-10-m-issue-trackers-2/squid-users/README.md +++ b/campaign_files/2026-10-m-issue-trackers-2/squid-users/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. -**Answers:** No answer files are included. Reference result: Facts computed from the thread headers. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Read the archive as mbox with Python's mailbox and email modules (or mutt -f / notmuch on the decompressed file); rebuild the thread from Message-ID, In-Reply-To and References headers. **Licence (SPDX):** NOASSERTION @@ -21,7 +19,5 @@ Files: - `m089-14/2019-March.txt.gz`: [squid-users] attempting to disable (or mute) logs (12 messages, 2019-03-13 to 2019-03-24). Format: Pipermail month archive, mbox text gzip-compressed: full headers (From with obfuscated address and name, Date, Subject, Message-ID, In-Reply-To, References) and plain-text bodies. Cut the thread by root Message-ID through the In-Reply-To closure (last References entry when In-Reply-To is missing). Origin: Real list traffic in a closed past month of the list's Pipermail archive. The unit is one thread, cut out of the month file by its root Message-ID; the month file is the pinned pointer and is delivered whole. - Reference result: facts computed from the thread headers - Reference source: Message-ID, In-Reply-To, References, From and Date headers of the messages in the thread, read from the pinned month file. Check: Exact match against values computed from the pinned month file: parse it as mbox, take the root Message-ID, collect every message whose In-Reply-To (or last References entry when In-Reply-To is missing) chains back to it, then count, take dates in UTC and compare sender names. Message-IDs are compared after removing double quotes (one archive writes a quoted local part that replies cite unquoted; found in the hand check). People are counted by person: From entries with the same address, or the same name after removing accents and letter case, are one person (the names in scale are the first spelling of each). Deterministic. The program was checked by hand on 2 threads of this source before giving answers: the chosen thread (read message by message) and 'Why squid on CentOS is faster than Debian ?' (20 messages by headers, one subject throughout); it agreed on both. Tools: Read the archive as mbox with Python's mailbox and email modules (or mutt -f / notmuch on the decompressed file); rebuild the thread from Message-ID, In-Reply-To and References headers. diff --git a/campaign_files/2026-10-m-issue-trackers-2/strongswan-users/README.md b/campaign_files/2026-10-m-issue-trackers-2/strongswan-users/README.md index c8368f25006eaafeec8de444aaf3e669e76158f8..952c7c6636b8e7bfd754a07c4e2a895d5fab4aa7 100644 --- a/campaign_files/2026-10-m-issue-trackers-2/strongswan-users/README.md +++ b/campaign_files/2026-10-m-issue-trackers-2/strongswan-users/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. -**Answers:** No answer files are included. Reference result: Facts computed from the thread headers. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Read the archive as mbox with Python's mailbox and email modules (or mutt -f / notmuch on the decompressed file); rebuild the thread from Message-ID, In-Reply-To and References headers. **Licence (SPDX):** NOASSERTION @@ -21,7 +19,5 @@ Files: - `m089-15/2019-March.txt.gz`: [strongSwan] Error connecting from Fortigate VPN to Strongswan (5 messages, 2019-03-14 to 2019-03-15). Format: Pipermail month archive, mbox text gzip-compressed: full headers (From with obfuscated address and name, Date, Subject, Message-ID, In-Reply-To, References) and plain-text bodies. Cut the thread by root Message-ID through the In-Reply-To closure (last References entry when In-Reply-To is missing). Origin: Real list traffic in a closed past month of the list's Pipermail archive. The unit is one thread, cut out of the month file by its root Message-ID; the month file is the pinned pointer and is delivered whole. - Reference result: facts computed from the thread headers - Reference source: Message-ID, In-Reply-To, References, From and Date headers of the messages in the thread, read from the pinned month file. Check: Exact match against values computed from the pinned month file: parse it as mbox, take the root Message-ID, collect every message whose In-Reply-To (or last References entry when In-Reply-To is missing) chains back to it, then count, take dates in UTC and compare sender names. Message-IDs are compared after removing double quotes (one archive writes a quoted local part that replies cite unquoted; found in the hand check). People are counted by person: From entries with the same address, or the same name after removing accents and letter case, are one person (the names in scale are the first spelling of each). Deterministic. The program was checked by hand on 2 threads of this source before giving answers: the chosen thread (read message by message) and 'VPN connection to remote Fortigate client' (12 messages by headers, one subject throughout); it agreed on both. Tools: Read the archive as mbox with Python's mailbox and email modules (or mutt -f / notmuch on the decompressed file); rebuild the thread from Message-ID, In-Reply-To and References headers. diff --git a/campaign_files/2026-10-m-issue-trackers-2/swift-users/README.md b/campaign_files/2026-10-m-issue-trackers-2/swift-users/README.md index 9ba677dacbea48659e63dc5b308d2e307ddf3a79..a92ee00772a148d3c29359f0060fa89513e0893d 100644 --- a/campaign_files/2026-10-m-issue-trackers-2/swift-users/README.md +++ b/campaign_files/2026-10-m-issue-trackers-2/swift-users/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. -**Answers:** No answer files are included. Reference result: Facts computed from the thread headers. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Read the archive as mbox with Python's mailbox and email modules (or mutt -f / notmuch on the decompressed file); rebuild the thread from Message-ID, In-Reply-To and References headers. **Licence (SPDX):** NOASSERTION @@ -21,7 +19,5 @@ Files: - `m259-05/Week-of-Mon-20170306.txt.gz`: [swift-users] What is the intended response to Array change notification for dependents? (10 messages, 2017-03-07 to 2017-03-07). Format: Pipermail weekly archive, mbox text gzip-compressed: full headers (From with obfuscated address and name, Date, Subject, Message-ID, In-Reply-To, References) and plain-text bodies. Cut the thread by root Message-ID through the In-Reply-To closure (last References entry when In-Reply-To is missing). Origin: Real list traffic in a closed past week of the list's archive. The unit is one thread, cut out of the week file by its root Message-ID; the week file is the pinned pointer and is delivered whole. - Reference result: facts computed from the thread headers - Reference source: Message-ID, In-Reply-To, References, From and Date headers of the messages in the thread, read from the pinned week file. Check: Exact match against values computed from the pinned week file: parse it as mbox, drop repeated copies of a Message-ID already seen (first copy kept), take the root Message-ID, collect every message whose In-Reply-To (or last References entry when In-Reply-To is missing) chains back to it, then count, take dates in UTC (a Date header without zone is read as UTC) and compare sender names. Message-IDs are compared after removing angle brackets and double quotes. People are counted by person: From entries with the same address, or the same name after removing accents and letter case, are one person (the names in scale are the first spelling of each). The first-reply interval is the exact difference in whole seconds between the Date headers of the root and the earliest reply, converted to UTC; Date headers have one-second resolution, so no rounding is involved. Deterministic. The program was checked on 2 threads of this source by a second, independent parser (splits the raw text on mbox From lines and reads headers with regular expressions, no mailbox or email module): the chosen thread (also read message by message) and '[swift-users] How to optionally link swift framework into another framework' (8 messages, root <5B8476E3-CD12-46F0-BE41-71F456498723@avast.com>); message count, last message time and first-reply seconds agreed on both. Tools: Read the archive as mbox with Python's mailbox and email modules (or mutt -f / notmuch on the decompressed file); rebuild the thread from Message-ID, In-Reply-To and References headers. diff --git a/campaign_files/2026-10-m-issue-trackers-2/swiftshader-gerrit-changes/README.md b/campaign_files/2026-10-m-issue-trackers-2/swiftshader-gerrit-changes/README.md index 31768bc21497ac2448483e47f7c10b6b06c46a75..6aaa99a9d1555c576058bdf48ae1c7c0f2ef2bd5 100644 --- a/campaign_files/2026-10-m-issue-trackers-2/swiftshader-gerrit-changes/README.md +++ b/campaign_files/2026-10-m-issue-trackers-2/swiftshader-gerrit-changes/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. -**Answers:** No answer files are included. Reference result: Facts computed from the change's own review timeline (status, submit time and submitter, patch sets, votes, reviewers added or removed, review comments). Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** curl the Gerrit REST API (GET https://swiftshader-review.googlesource.com/changes/77788/detail), strip the first line ")]}'" and read the JSON with jq or Python json; the git-review CLI and Gerrit's ssh 'gerrit query --format=JSON --all-approvals --comments' reach the same data where ssh access is open **Licence (SPDX):** NOASSERTION @@ -21,7 +19,5 @@ Files: - `changes/77788/detail.json`: SwiftShader change 77788: Backport Vulkan-Hpp include fix. Format: Gerrit REST API ChangeInfo JSON from the /detail endpoint, preceded by the anti-XSSI line ")]}'": status, created, updated, submitted, submitter, owner, labels with every vote, messages (author, real_author, date, tag, _revision_number, message text) and reviewer_updates (updated, updated_by, reviewer, state REVIEWER/CC/REMOVED). Patch-set uploads are messages tagged autogenerated:gerrit:newPatchSet; votes appear in message text as 'Patch Set N: Label+V' and in labels.*.all. The endpoint returns all messages and reviewer updates of the change, so the record is complete. Origin: Live Gerrit REST interface of a closed (merged) code review change, collected as a NOASSERTION pointer row under spec general rule 4: the detail JSON was fetched twice apart and was byte-identical. Unit file kept only outside the repository for checking and then deleted. - Reference result: facts computed from the change's own review timeline (status, submit time and submitter, patch sets, votes, reviewers added or removed, review comments) - Reference source: The pinned response: changes/77788/detail.json. Check: Exact match against values read from the pinned JSON: status, submitted and submitter; reviewer_updates entries (reviewer, state, updated_by, updated); messages tagged autogenerated:gerrit:newPatchSet (patch-set count, _revision_number); votes from labels.*.all (value, _account_id) and from message first lines 'Patch Set N: Label+V'. A human review comment is a message whose author is not a bot or service account, whose tag does not start with 'autogenerated:', and whose text has content beyond the 'Patch Set N: votes' first line or names inline comments '(N comments)'. Bot and service accounts excluded from the comment count: SwiftShader RegressionBot, swiftshader-scoped@luci-project-accounts.iam.gserviceaccount.com, accounts tagged SERVICE_USER, and any account whose name, username or email matches the pattern (bot|zuul|jenkins|ci|luci|gerrit code review|service user|automation|buildbot|lint|sonar|tryjob|cq|autosubmit|copybara|tricium|gerritforge|verifier|checker|scheduler|maloo|janitor|autotest|gerrit). Deterministic, no model judgement. Before answers are issued, the extraction program must pass on 2 hand-checked units of this source (spec general rule 5); hand-checking is pending for the delivery step. Tools: curl the Gerrit REST API (GET https://swiftshader-review.googlesource.com/changes/77788/detail), strip the first line ")]}'" and read the JSON with jq or Python json; the git-review CLI and Gerrit's ssh 'gerrit query --format=JSON --all-approvals --comments' reach the same data where ssh access is open diff --git a/campaign_files/2026-10-m-issue-trackers-2/swindonbus-passenger-gtfs/README.md b/campaign_files/2026-10-m-issue-trackers-2/swindonbus-passenger-gtfs/README.md index 7aee37eb3a36855ef66973b3744ac681d550e017..b387a11d19e9ec025d561f4546c0778bec246f92 100644 --- a/campaign_files/2026-10-m-issue-trackers-2/swindonbus-passenger-gtfs/README.md +++ b/campaign_files/2026-10-m-issue-trackers-2/swindonbus-passenger-gtfs/README.md @@ -8,8 +8,6 @@ **Self-contained:** The unit is self-contained. 1 of 1 units name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per unit in the notes. -**Answers:** No answer files are included. Reference result: Facts computed from the service calendar itself (active services on a date, removed or added dates, day counts). Where the reference result comes from and how a solution is checked is written per unit in the notes. - **Tools to open and check:** Read the CSV files with Python csv or pandas (or sqlite3 .import) and expand service days per GTFS rules; gtfs-kit or partridge compute active service_ids per date headless; dates are local to agency_timezone (Python zoneinfo). **Licence (SPDX):** OGL-UK-3.0 @@ -22,8 +20,6 @@ Units: Read out of the archive: https://s3-eu-west-1.amazonaws.com/passenger-sources/swindonbus/gtfs/swindonbus_1790768069.zip Format: text/csv (GTFS Schedule calendar.txt; companions calendar_dates.txt, agency.txt) Origin: Live static GTFS feed published by Swindon's Bus Company (published through Passenger, data.discoverpassenger.com) at https://s3-eu-west-1.amazonaws.com/passenger-sources/swindonbus/gtfs/swindonbus_1790768069.zip. calendar.txt and its companions were read by member out of the zip with HTTP Range requests (spec general rule 4: by member); kept as a snapshot pinned by sha256. - Reference result: facts computed from the service calendar itself (active services on a date, removed or added dates, day counts) - Reference source: calendar.txt weekday flags with start_date and end_date, calendar_dates.txt exception rows, agency_timezone of agency.txt, all from the pinned members. Check: deterministic, no model judgement. Hand-checking 2 samples per source is pending for the delivery step (spec general rule 5). Tools: Read the CSV files with Python csv or pandas (or sqlite3 .import) and expand service days per GTFS rules; gtfs-kit or partridge compute active service_ids per date headless; dates are local to agency_timezone (Python zoneinfo). Companions (not included): calendar_dates.txt; agency.txt diff --git a/campaign_files/2026-10-md-media-and-documents/spoken-wikipedia-en-island-beach-new-jersey/README.md b/campaign_files/2026-10-md-media-and-documents/spoken-wikipedia-en-island-beach-new-jersey/README.md index 1a4dc0c7030b6fdaf5a762ee464be69f2a3c2023..f4cb638b5168548f94a1716e08c42bf40d6587b9 100644 --- a/campaign_files/2026-10-md-media-and-documents/spoken-wikipedia-en-island-beach-new-jersey/README.md +++ b/campaign_files/2026-10-md-media-and-documents/spoken-wikipedia-en-island-beach-new-jersey/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Reference text the reader read (article revision). Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Give the agent the ogg only; it writes the transcript; score word error rate (WER) against the reference words, reported as a number with no pass mark; reference = the prose of the oldid wikitext with markup, infobox, references, tables and image captions removed; the spoken intro and outro are cut from the hypothesis; the reading is not word for word (readers expand numbers and skip parentheticals), and the revision is inferred from timestamps, so the score is reported as a soft reference score. **Licence (SPDX):** CC-BY-SA-4.0 @@ -21,7 +19,5 @@ Files: - `wikipedia/commons/a/ad/Spoken_Wikipedia_en_Island_Beach_New_Jersey.ogg`: Spoken Wikipedia recording of the English article "Island Beach, New Jersey" with the article revision it reads (oldid 659659189). Format: Ogg Vorbis, 2,186,837 bytes, about 157 s, one read encyclopedia article Origin: volunteer human reading of a fixed article revision (uploaded 2016-01-01T17:57:23Z, template date=2016-01-01) - Reference result: reference text the reader read (article revision) - Reference source: w/index.php?oldid=659659189&action=raw: the article prose of revision 659659189; how the revision was chosen: the file page template gives no oldid, only date=2016-01-01; the revision list (companion w/api.php.revisions_Island_Beach_New_Jersey_2016-01-01.json) shows revision 659659189 (2015-04-28, Ricky81682, 3,330 bytes) was current when the ogg was uploaded at 2016-01-01T17:57:23Z; the reader's own edits 697754509 (18:14:05, same size) and 697754569 (18:14:37, adds 80 bytes, the spoken-article template) came after the upload, so 659659189 is the text read; this is an inference from timestamps, not a value the file page states Check: give the agent the ogg only; it writes the transcript; score word error rate (WER) against the reference words, reported as a number with no pass mark; reference = the prose of the oldid wikitext with markup, infobox, references, tables and image captions removed; the spoken intro and outro are cut from the hypothesis; the reading is not word for word (readers expand numbers and skip parentheticals), and the revision is inferred from timestamps, so the score is reported as a soft reference score Companions (not included): w/index.php?oldid=659659189&action=raw; w/api.php.File_Spoken_Wikipedia_en_Island_Beach_New_Jersey.ogg.json; w/api.php.revisions_Island_Beach_New_Jersey_2016-01-01.json diff --git a/campaign_files/2026-10-md-media-and-documents/striprtf/README.md b/campaign_files/2026-10-md-media-and-documents/striprtf/README.md index 7be27ffc8562c2112e07492d47a2b40d7faa53d9..c8745846f43a00506e31df2ae6e1d9ab01c06f8f 100644 --- a/campaign_files/2026-10-md-media-and-documents/striprtf/README.md +++ b/campaign_files/2026-10-md-media-and-documents/striprtf/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected plain text of an RTF document. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** striprtf.rtf_to_text() from Python (or the striprtf command line script) on the RTF file; output compared byte for byte with the expected text **Licence (SPDX):** BSD-3-Clause @@ -21,8 +19,6 @@ Files: - `tests/rtf/french.rtf`: Striprtf test: French RTF with typographic apostrophes converted to plain text. Format: RTF input (french.rtf) and one expected UTF-8 text file (french.txt) Origin: a short French RTF document with typographic apostrophes and its expected plain text written by the striprtf maintainer; tests/test_french.py is the judge - Reference result: expected plain text of an RTF document - Reference source: tests/text/french.txt Check: give the agent french.rtf and ask for its plain text; read the expected french.txt as UTF-8 and require exact string equality with the produced text (no normalisation) Tools: striprtf.rtf_to_text() from Python (or the striprtf command line script) on the RTF file; output compared byte for byte with the expected text Companions (not included): tests/text/french.txt; tests/test_french.py; LICENSE diff --git a/campaign_files/2026-10-md-media-and-documents/tabula-java/README.md b/campaign_files/2026-10-md-media-and-documents/tabula-java/README.md index b66d2a707865d26fb5fc3d88da7f4a0942279d3b..eae448877d3f80e0439b3f82299c64b21e4809fb 100644 --- a/campaign_files/2026-10-md-media-and-documents/tabula-java/README.md +++ b/campaign_files/2026-10-md-media-and-documents/tabula-java/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected CSV of two lattice tables on a PDF page. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** tabula-java command line (java -jar tabula.jar --lattice --pages 1 --format CSV spanning_cells.pdf) or the SpreadsheetExtractionAlgorithm Java API; CSV compared with the expected file **Licence (SPDX):** MIT @@ -21,8 +19,6 @@ Files: - `src/test/resources/technology/tabula/spanning_cells.pdf`: Tabula-java test: PDF page with spanning cells extracted to CSV. Format: PDF input (1 page) and one expected CSV file Origin: a one-page PDF with two lattice tables and the expected CSV committed by the tabula maintainers; TestSpreadsheetExtractor.testSpanningCellsToCsv is the judge - Reference result: expected CSV of two lattice tables on a PDF page - Reference source: src/test/resources/technology/tabula/csv/spanning_cells.csv Check: give the agent spanning_cells.pdf and ask for the tables on page 1 as CSV; the result must hold exactly 2 tables and, after the expected CSV is re-printed in Excel CSV format with CR row ends, equal the produced CSV exactly Tools: tabula-java command line (java -jar tabula.jar --lattice --pages 1 --format CSV spanning_cells.pdf) or the SpreadsheetExtractionAlgorithm Java API; CSV compared with the expected file Companions (not included): src/test/resources/technology/tabula/csv/spanning_cells.csv; src/test/java/technology/tabula/TestSpreadsheetExtractor.java; src/test/java/technology/tabula/UtilsForTesting.java; LICENSE diff --git a/campaign_files/2026-10-md-media-and-documents/tavern/LICENSE b/campaign_files/2026-10-md-media-and-documents/tavern/LICENSE index b9550257773feb3d799bfe727a789c80b810e5fe..18a4930770f17f247fb908208fe97caefb8595b0 100644 --- a/campaign_files/2026-10-md-media-and-documents/tavern/LICENSE +++ b/campaign_files/2026-10-md-media-and-documents/tavern/LICENSE @@ -1,3 +1,4 @@ Licence: CC-BY-SA-4.0 - CC-BY-SA-4.0: https://creativecommons.org/licenses/by-sa/4.0/ (publisher: Johanna Devaney et al., TAVERN (GitHub jcdevaney/TAVERN)) +- CC-BY-SA-4.0: https://creativecommons.org/licenses/by-sa/4.0/ (publisher: Johanna Devaney et al., TAVERN) diff --git a/campaign_files/2026-10-md-media-and-documents/tavern/README.md b/campaign_files/2026-10-md-media-and-documents/tavern/README.md index 6bde01cba670905fa3fb182829248f1cfc1a3b4e..629a13bd0d9177bc4fa7327ef69baddcbd00160e 100644 --- a/campaign_files/2026-10-md-media-and-documents/tavern/README.md +++ b/campaign_files/2026-10-md-media-and-documents/tavern/README.md @@ -8,8 +8,6 @@ **Self-contained:** Each unit is self-contained. 4 of 4 units name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per unit in the notes. Dependencies named upstream are listed per unit in the notes. -**Answers:** No answer files are included. Reference result: roman_numeral_and_function_spines. Where the reference result comes from and how a solution is checked is written per unit in the notes. - **Tools to open and check:** music21 (Python) parses Humdrum **kern and the **harm / **function spines; humlib or the Humdrum toolkit (extract) pulls spines on the command line **Licence (SPDX):** CC-BY-SA-4.0 @@ -21,8 +19,6 @@ Units: - `Mozart/K025/` (2 files): Roman numeral and phrase-function analysis of Mozart, 7 Variations in D major on "Willem van Nassau", K.25, by two independent annotators. Format: directory of 101 files, 182,891 bytes: Joined/ 48 .krn files (**function, **harm, two **kern spines; suffix _a or _b names the annotator); Encodings/ holds only Encoder_B; the Joined/ files carry both annotators as _a and _b copies (README line 18); Krn/ per-phrase *score.krn plus whole-work files (K025.krn, K025.mid, K025.sib, interp.krn, kernsplit_debugged) Origin: human expert analyses with double annotation, encoded from public-domain piano works - Reference result: roman_numeral_and_function_spines - Reference source: **harm (Roman numerals) and **function (T/PD/D phrase-model function, after Laitz) spines written by human annotators A and B Check: give the agent the score phrases (Krn/*score.krn, or the **kern spines of Joined/ with **harm and **function removed); it writes **harm and **function spines; score per event against annotator A and annotator B separately, aligned on the Humdrum data lines (music21 or humlib); report the A versus B agreement on the same phrases beside the agent scores as the noise band; no pass line is set before that band is measured Tools: music21 (Python) parses Humdrum **kern and the **harm / **function spines; humlib or the Humdrum toolkit (extract) pulls spines on the command line Companions (not included): README.md (file naming key) @@ -30,8 +26,6 @@ Units: - `Mozart/K179/` (2 files): Roman numeral and phrase-function analysis of Mozart, 12 Variations in C major on a Minuet by Fischer, K.179 (189a), by two independent annotators. Format: directory of 283 files, 3,361,215 bytes: Joined/ 140 .krn files (**function, **harm, two **kern spines; suffix _a or _b names the annotator); Encodings/ holds only Encoder_B; the Joined/ files carry both annotators as _a and _b copies (README line 18); Krn/ per-phrase *score.krn plus whole-work files (K179.krn, K179.sib, K179.xml) Origin: human expert analyses with double annotation, encoded from public-domain piano works - Reference result: roman_numeral_and_function_spines - Reference source: **harm (Roman numerals) and **function (T/PD/D phrase-model function, after Laitz) spines written by human annotators A and B Check: give the agent the score phrases (Krn/*score.krn, or the **kern spines of Joined/ with **harm and **function removed); it writes **harm and **function spines; score per event against annotator A and annotator B separately, aligned on the Humdrum data lines (music21 or humlib); report the A versus B agreement on the same phrases beside the agent scores as the noise band; no pass line is set before that band is measured Tools: music21 (Python) parses Humdrum **kern and the **harm / **function spines; humlib or the Humdrum toolkit (extract) pulls spines on the command line Companions (not included): README.md (file naming key) @@ -39,8 +33,6 @@ Units: - `Beethoven/B064/` (2 files): Roman numeral and phrase-function analysis of Beethoven, 6 Variations on a Swiss Song, WoO 64, by two independent annotators. Format: directory of 138 files, 832,077 bytes: Joined/ 42 .krn files (**function, **harm, two **kern spines; suffix _a or _b names the annotator); Encodings/ holds Encoder_A, Encoder_B, Encoder_C; Krn/ per-phrase *score.krn plus whole-work files (Wo064.krn, Wo064.mid, Wo064.sib, Wo064.xml, working064) Origin: human expert analyses with double annotation, encoded from public-domain piano works - Reference result: roman_numeral_and_function_spines - Reference source: **harm (Roman numerals) and **function (T/PD/D phrase-model function, after Laitz) spines written by human annotators A and B Check: give the agent the score phrases (Krn/*score.krn, or the **kern spines of Joined/ with **harm and **function removed); it writes **harm and **function spines; score per event against annotator A and annotator B separately, aligned on the Humdrum data lines (music21 or humlib); report the A versus B agreement on the same phrases beside the agent scores as the noise band; no pass line is set before that band is measured Tools: music21 (Python) parses Humdrum **kern and the **harm / **function spines; humlib or the Humdrum toolkit (extract) pulls spines on the command line Companions (not included): README.md (file naming key) @@ -48,8 +40,6 @@ Units: - `Beethoven/B080/` (2 files): Roman numeral and phrase-function analysis of Beethoven, 32 Variations in C minor, WoO 80, by two independent annotators. Format: directory of 175 files, 5,898,229 bytes: Joined/ 68 .krn files (**function, **harm, two **kern spines; suffix _a or _b names the annotator); Encodings/ holds Encoder_A, Encoder_B; Krn/ per-phrase *score.krn plus whole-work files (Wo080.krn, Wo080.mid, Wo080.sib, Wo080.xml, working080) Origin: human expert analyses with double annotation, encoded from public-domain piano works - Reference result: roman_numeral_and_function_spines - Reference source: **harm (Roman numerals) and **function (T/PD/D phrase-model function, after Laitz) spines written by human annotators A and B Check: give the agent the score phrases (Krn/*score.krn, or the **kern spines of Joined/ with **harm and **function removed); it writes **harm and **function spines; score per event against annotator A and annotator B separately, aligned on the Humdrum data lines (music21 or humlib); report the A versus B agreement on the same phrases beside the agent scores as the noise band; no pass line is set before that band is measured Tools: music21 (Python) parses Humdrum **kern and the **harm / **function spines; humlib or the Humdrum toolkit (extract) pulls spines on the command line Companions (not included): README.md (file naming key) diff --git a/campaign_files/2026-10-md-media-and-documents/tectonic-tex-outputs/README.md b/campaign_files/2026-10-md-media-and-documents/tectonic-tex-outputs/README.md index fb7b22d55a9e1ab43f8746b391da86d3d484dfbf..2eef1b55b9d61f11584c430a0a2bdeeba66bb2cb 100644 --- a/campaign_files/2026-10-md-media-and-documents/tectonic-tex-outputs/README.md +++ b/campaign_files/2026-10-md-media-and-documents/tectonic-tex-outputs/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 2 of 2 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected XDV output and TeX log. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** tectonic -X compile --outfmt xdv otf_basic.tex (headless CLI); upstream cargo test --test tex-outputs otf_basic; tectonic -X compile --outfmt xdv tex_logo.tex (headless CLI); upstream cargo test --test tex-outputs tex_logo **Licence (SPDX):** MIT @@ -21,8 +19,6 @@ Files: - `tests/tex-outputs/otf_basic.tex`: Tectonic test otf_basic: paragraph in Latin Modern OpenType with hyphenation and a missing DESERET character, with expected XDV and log. Format: plain TeX source (.tex, 11 lines) plus expected XDV (6304 bytes) and log Origin: Tectonic (XeTeX-derived engine) output test: an OpenType font paragraph with hyphenation and a character outside the BMP, written by the Tectonic developers - Reference result: expected XDV output and TeX log - Reference source: tests/tex-outputs/otf_basic.xdv and otf_basic.log; the log records "Missing character: There is no 𐐷 (U+10437) in font [lmroman12-regular]!" and the test expects the outcome Warnings (tests/tex-outputs.rs lines 244 to 248) Check: hide the .xdv and .log; the agent predicts the warnings and line breaks or fixes the source; the run must end with warnings, not errors Tools: tectonic -X compile --outfmt xdv otf_basic.tex (headless CLI); upstream cargo test --test tex-outputs otf_basic Companions (not included): tests/tex-outputs/otf_basic.xdv; tests/tex-outputs/otf_basic.log @@ -30,8 +26,6 @@ Files: - `tests/tex-outputs/tex_logo.tex`: Tectonic test tex_logo: the TeX logo macro (kern -.1667em, lowered E) with expected XDV and log. Format: plain TeX source (.tex, 6 lines) plus expected XDV (228 bytes) and log Origin: Tectonic output test: the TeX logo built from kerns and a lowered box in plain TeX, written by the Tectonic developers - Reference result: expected XDV output and TeX log - Reference source: tests/tex-outputs/tex_logo.xdv (glyph positions with the kerns and the lowered E) and tex_logo.log ("Output written on tex_logo.xdv (1 page, 228 bytes).") Check: hide the .xdv; the agent writes the plain TeX macro for the logo. Tools: tectonic -X compile --outfmt xdv tex_logo.tex (headless CLI); upstream cargo test --test tex-outputs tex_logo Companions (not included): tests/tex-outputs/tex_logo.xdv; tests/tex-outputs/tex_logo.log diff --git a/campaign_files/2026-10-md-media-and-documents/tesseract-ocr-test/README.md b/campaign_files/2026-10-md-media-and-documents/tesseract-ocr-test/README.md index 578d5e3b58489433a4d181f16fcfc793ecee7459..e15af4101943f63bec49c9d90b401c792e7b3d2d 100644 --- a/campaign_files/2026-10-md-media-and-documents/tesseract-ocr-test/README.md +++ b/campaign_files/2026-10-md-media-and-documents/tesseract-ocr-test/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 2 of 2 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Ground-truth text of the page image. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Hide phototest.txt; the agent OCRs phototest.tif; pass when the text equals phototest.txt exactly, as OCRTester does with EXPECT_EQ(gtText, outText); a lenient variant can score character error rate, which the upstream test does not define; hide eurotext.txt; the agent OCRs eurotext.tif; pass when the text equals eurotext.txt exactly, as OCRTester does with EXPECT_EQ(gtText, outText); a lenient variant can score character error rate, which the upstream test does not define. **Licence (SPDX):** Apache-2.0 @@ -21,16 +19,12 @@ Files: - `testing/phototest.tif`: Tesseract test page phototest.tif: OCR to the exact ground-truth text. Format: bilevel TIFF page (38668 bytes) plus ground-truth text (286 bytes) Origin: Tesseract OCR test image with ground-truth text, used by the tesseract unit test apiexample_test that requires the OCR output to equal the text exactly - Reference result: ground-truth text of the page image - Reference source: testing/phototest.txt (the text the tesseract unit test requires) Check: hide phototest.txt; the agent OCRs phototest.tif; pass when the text equals phototest.txt exactly, as OCRTester does with EXPECT_EQ(gtText, outText); a lenient variant can score character error rate, which the upstream test does not define Companions (not included): testing/phototest.txt; unittest/apiexample_test.cc (tesseract-ocr/tesseract @49ac41de31d3ad3826272e2b2158552cb6b6ed5d); testing/README.md Dependencies: Tesseract with the eng model, since exact equality holds only for that model - `testing/eurotext.tif`: Tesseract test page eurotext.tif: multilingual Latin-script OCR to the exact ground-truth text. Format: bilevel TIFF page 1024x800 (102598 bytes) plus ground-truth text in English, German, French, Italian, Spanish and Portuguese Origin: Tesseract OCR test image with ground-truth text, used by the tesseract unit test apiexample_test that requires the OCR output to equal the text exactly; the page mixes English with German, French, Italian, Spanish and Portuguese sentences - Reference result: ground-truth text of the page image - Reference source: testing/eurotext.txt (the text the tesseract unit test requires) Check: hide eurotext.txt; the agent OCRs eurotext.tif; pass when the text equals eurotext.txt exactly, as OCRTester does with EXPECT_EQ(gtText, outText); a lenient variant can score character error rate, which the upstream test does not define Companions (not included): testing/eurotext.txt; unittest/apiexample_test.cc (tesseract-ocr/tesseract @49ac41de31d3ad3826272e2b2158552cb6b6ed5d) Dependencies: Tesseract with the script/Latin model, since exact equality holds only for that model diff --git a/campaign_files/2026-10-md-media-and-documents/tex-live-texlive-source-texk-web2c-triptrap/README.md b/campaign_files/2026-10-md-media-and-documents/tex-live-texlive-source-texk-web2c-triptrap/README.md index 0fecf9aac7a7315ff51a7adedd407957611703e5..71f37786b3d3bf9818b4cc2313e32d0a37f6d1a6 100644 --- a/campaign_files/2026-10-md-media-and-documents/tex-live-texlive-source-texk-web2c-triptrap/README.md +++ b/campaign_files/2026-10-md-media-and-documents/tex-live-texlive-source-texk-web2c-triptrap/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether the file needs other upstream files is not stated. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected TeX log, terminal output and typeset box summary of the trip run. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** build tex with the trip constants (make trip in texk/web2c, which runs tex on trip1.in and trip2.in) and diff the produced trip.log, trip.fot and trip.typ against the stored ones **Licence (SPDX):** NOASSERTION @@ -21,8 +19,6 @@ Files: - `texk/web2c/triptrap/trip.tex`: Knuth's trip test for TeX (texlive-source triptrap): torture test of TeX with the expected log and typeset output. Format: TeX source plus property list font plus expected logs Origin: Donald Knuth's trip test, the acceptance test every TeX implementation must pass, kept in the TeX Live source tree - Reference result: expected TeX log, terminal output and typeset box summary of the trip run - Reference source: texk/web2c/triptrap/trip.log, trip.fot, trip.typ and tripin.log, written by Knuth Check: run tex in trip mode on trip.tex with trip1.in and trip2.in, then diff each output against the stored file; only the kinds of differences that tripman.tex and README allow are accepted; hide the expected files when posing the task Tools: build tex with the trip constants (make trip in texk/web2c, which runs tex on trip1.in and trip2.in) and diff the produced trip.log, trip.fot and trip.typ against the stored ones Companions (not included): texk/web2c/triptrap/trip.pl; texk/web2c/triptrap/trip1.in; texk/web2c/triptrap/trip2.in; texk/web2c/triptrap/tripin.log; texk/web2c/triptrap/trip.log; texk/web2c/triptrap/trip.fot; texk/web2c/triptrap/trip.typ; texk/web2c/triptrap/tripman.tex diff --git a/campaign_files/2026-10-md-media-and-documents/texhackse-komascript-tagging-lab-koma-testfiles/README.md b/campaign_files/2026-10-md-media-and-documents/texhackse-komascript-tagging-lab-koma-testfiles/README.md index 6778c2a6a0cbc0e6966f9cddf8825cbb6c796092..6826ab84e8ed4d70ceb40ba4498c8e0b8d6feabc 100644 --- a/campaign_files/2026-10-md-media-and-documents/texhackse-komascript-tagging-lab-koma-testfiles/README.md +++ b/campaign_files/2026-10-md-media-and-documents/texhackse-komascript-tagging-lab-koma-testfiles/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether the file needs other upstream files is not stated. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected normalised TeX log (.tlg) with the paper and media sizes. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** l3build check typearea-pagesize with KOMA-build.lua (KOMA-Script's own build script) as the build file and the test in its default testfiles/ directory (headless LaTeX run, log normalised and diffed against the .tlg) **Licence (SPDX):** LPPL-1.3c @@ -21,8 +19,6 @@ Files: - `KOMA-testfiles/typearea-pagesize.lvt`: KOMA-Script typearea test pagesize: with a5paper and a standard class, paper and PDF media sizes must both be 421.10083pt by 597.50793pt. Format: LaTeX test file (.lvt) plus expected log (.tlg) Origin: regression test of the KOMA-Script maintainer (issue 23 of the KOMA-Script tracker, named in the .lvt header); the .tlg is the accepted normalised log - Reference result: expected normalised TeX log (.tlg) with the paper and media sizes - Reference source: KOMA-testfiles/typearea-pagesize.tlg line 6 "Paper: width=421.10083pt, height=597.50793pt" and the media line after it Check: run the .lvt with l3build on each engine and compare the normalised .log with the .tlg line by line; hide the .tlg when posing the task Tools: l3build check typearea-pagesize with KOMA-build.lua (KOMA-Script's own build script) as the build file and the test in its default testfiles/ directory (headless LaTeX run, log normalised and diffed against the .tlg) Companions (not included): KOMA-build.lua; build.lua; KOMA-README.md; README.md diff --git a/campaign_files/2026-10-md-media-and-documents/texmath-regression-187/README.md b/campaign_files/2026-10-md-media-and-documents/texmath-regression-187/README.md index ee7444e79a159fa9b0ba3f8dca4718522f783d5f..910be348cccd1383b087f87d76f8e6036ac9661c 100644 --- a/campaign_files/2026-10-md-media-and-documents/texmath-regression-187/README.md +++ b/campaign_files/2026-10-md-media-and-documents/texmath-regression-187/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected LaTeX math for one OMML equation. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** texmath command line program (texmath -f omml -t tex < input.xml) or pandoc on a .docx that holds the equation; Haskell API readOMML and writeTeX **Licence (SPDX):** GPL-2.0-only @@ -21,8 +19,6 @@ Files: - `test/regression/187.test`: Texmath regression test 187: an Office Math Markup (OMML) equation converted to TeX. Format: texmath golden test file: a line "<<< omml", the OMML XML input, a line ">>> tex", then the expected TeX output Origin: regression test written by the texmath maintainers for issue 187 (OMML input that previously converted wrongly); not copied from another suite - Reference result: expected LaTeX math for one OMML equation - Reference source: the lines after ">>> tex" in 187.test Check: split the file at the "<<< omml" and ">>> tex" lines as readGoldenTest in test/test-texmath.hs does; show the agent only the OMML input (hide everything after ">>> tex"); the agent converts it to TeX as texmath's writeTeX does; rebuild the file from the input and the agent's output with a final newline added (ensureFinalNewline) and pass when it is byte-identical to 187.test (the upstream runs diff -u and fails on any difference) Tools: texmath command line program (texmath -f omml -t tex < input.xml) or pandoc on a .docx that holds the equation; Haskell API readOMML and writeTeX Companions (not included): test/test-texmath.hs; texmath.cabal diff --git a/campaign_files/2026-10-md-media-and-documents/textile-php-textile-fixtures/README.md b/campaign_files/2026-10-md-media-and-documents/textile-php-textile-fixtures/README.md index 006b9ed9709d42f24edfb56db466e24772750272..910156cbe84eca4ff8282bca380ed17dabc5b41b 100644 --- a/campaign_files/2026-10-md-media-and-documents/textile-php-textile-fixtures/README.md +++ b/campaign_files/2026-10-md-media-and-documents/textile-php-textile-fixtures/README.md @@ -8,8 +8,6 @@ **Self-contained:** The file is self-contained. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected HTML per case. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Hide the expect fields; the agent converts each input with the case's setup (setRawBlocks true); pass per case when the output equals expect after trimming trailing whitespace and removing footnote ids, as test/Netcarver/Textile/Test/ParserFixtureTest.php does (lines 57 to 78). **Licence (SPDX):** BSD-3-Clause @@ -21,7 +19,5 @@ Files: - `test/fixtures/rawblocks.yaml`: PHP-Textile raw block fixtures: 12 Textile inputs converted to HTML. Format: YAML fixture file: 12 named cases, each with setup, input (Textile) and expect (HTML) Origin: PHP-Textile fixture file for raw blocks: 12 Textile inputs with the HTML the parser must produce - Reference result: expected HTML per case - Reference source: the expect field of each case Check: hide the expect fields; the agent converts each input with the case's setup (setRawBlocks true); pass per case when the output equals expect after trimming trailing whitespace and removing footnote ids, as test/Netcarver/Textile/Test/ParserFixtureTest.php does (lines 57 to 78) Dependencies: php-textile; symfony/yaml to read the fixture (pin the version when building the pool) diff --git a/campaign_files/2026-10-md-media-and-documents/textract/README.md b/campaign_files/2026-10-md-media-and-documents/textract/README.md index 261ab6abcec9d04e1bf21ef5e24a97ff4eea3691..897eb31c8c2ba5e917bb2574a867f3bb690ff9f8 100644 --- a/campaign_files/2026-10-md-media-and-documents/textract/README.md +++ b/campaign_files/2026-10-md-media-and-documents/textract/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Expected plain text of a .docx with paragraphs and nested tables. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** textract command line (textract paragraphs_and_tables.docx) or textract.process() in Python; output compared line by line after the whitespace normalisation of tests/base.py **Licence (SPDX):** MIT @@ -21,8 +19,6 @@ Files: - `tests/docx/paragraphs_and_tables.docx`: Textract test: Word document with paragraphs and nested table extracted to text. Format: .docx input and one expected text file Origin: a Word document with a paragraph before, a 2x3 table with an inner table, and a paragraph after, plus the expected text, written by the textract maintainers; tests/test_docx.py with tests/base.py is the judge - Reference result: expected plain text of a .docx with paragraphs and nested tables - Reference source: tests/docx/paragraphs_and_tables.txt Check: give the agent the .docx and ask for its text in reading order; drop blank lines from both, turn tab, CR and no-break space into spaces, collapse runs of spaces, strip each line, and require the two line lists to be equal Tools: textract command line (textract paragraphs_and_tables.docx) or textract.process() in Python; output compared line by line after the whitespace normalisation of tests/base.py Companions (not included): tests/docx/paragraphs_and_tables.txt; tests/test_docx.py; tests/base.py; LICENSE diff --git a/campaign_files/2026-10-md-media-and-documents/the-black-cat-librivox-reading/README.md b/campaign_files/2026-10-md-media-and-documents/the-black-cat-librivox-reading/README.md index 2e0fcc161a83b2fc4fcceb09f53251cee2c2c0f0..485c69b5d5eb45618e0d1f82e6684df0c0bfd6cb 100644 --- a/campaign_files/2026-10-md-media-and-documents/the-black-cat-librivox-reading/README.md +++ b/campaign_files/2026-10-md-media-and-documents/the-black-cat-librivox-reading/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. -**Answers:** No answer files are included. Reference result: Orthographic text of the story (about 3,900 words) from Poe's Tales (1845). Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Give the agent the audio file only; it writes the transcript; score word error rate (WER) against the reference words, reported as a number with no pass mark; the Wikisource page furniture (header, notes, licence box, footer) is removed and only the work's text is the reference; case and punctuation folded on both sides; the LibriVox preamble and closing credit ("This is a LibriVox recording...") are not in the text and are cut from the scored span; the story has two early printings with small differences and the reader's source is not stated, so the score is a soft reference score. **Licence (SPDX):** CC0-1.0 @@ -21,7 +19,5 @@ Files: - `wikipedia/commons/b/b4/The_Black_Cat.ogg`: Edgar Allan Poe, "The Black Cat", LibriVox reading by Don Morgan (about 33 minutes) with the Wikisource text of the story in Tales (1845). Format: Ogg Vorbis, 14,875,750 bytes, one complete reading of the story by a single male voice Origin: LibriVox volunteer reading of a public domain story - Reference result: orthographic text of the story (about 3,900 words) from Poe's Tales (1845) - Reference source: wikisource/Tales_(Poe)/The_Black_Cat.html, from "For the most wild, yet most homely narrative" to the end Check: give the agent the audio file only; it writes the transcript; score word error rate (WER) against the reference words, reported as a number with no pass mark; the Wikisource page furniture (header, notes, licence box, footer) is removed and only the work's text is the reference; case and punctuation folded on both sides; the LibriVox preamble and closing credit ("This is a LibriVox recording...") are not in the text and are cut from the scored span; the story has two early printings with small differences and the reader's source is not stated, so the score is a soft reference score Companions (not included): wikisource/Tales_(Poe)/The_Black_Cat.html; File_The_Black_Cat.ogg.html; wikisource/The_Black_Cat_(Poe).html diff --git a/campaign_files/2026-10-ml-serving-configs-2/stablehlo-interpreter-tests/README.md b/campaign_files/2026-10-ml-serving-configs-2/stablehlo-interpreter-tests/README.md index 6c662bd938a18cbb15ecbeddf3df6f713d5d2334..e82838c5f0ef52923dc5ddc118a5d9b9918ceac7 100644 --- a/campaign_files/2026-10-ml-serving-configs-2/stablehlo-interpreter-tests/README.md +++ b/campaign_files/2026-10-ml-serving-configs-2/stablehlo-interpreter-tests/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 9 of 9 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: inline_expected_constants. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). expect_eq_const is exact element equality; expect_almost_eq_const passes when every element differs by at most the tolerance (no explicit tolerance in this file, so the default applies), NaN matching NaN; Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_almost_eq_const, which passes when every element (real and imaginary parts for complex types) differs by at most the tolerance (no explicit tolerance in this file, so the default below applies), NaN matching NaN; Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_eq_const: exact element equality of the result with the inline constant, no tolerance. **Licence (SPDX):** Apache-2.0 @@ -21,72 +19,54 @@ Files: - `stablehlo/tests/interpret/convolution.mlir`: StableHLO reference interpreter test convolution.mlir: stablehlo.convolution; 6 functions (si64, si8, padding, batch_group_count 4, feature_group_count 2, feature before batch), 5 check.expect_eq_const and 1 check.expect_almost_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_eq_const and check.expect_almost_eq_const in stablehlo/tests/interpret/convolution.mlir, written by the StableHLO authors against the op specification Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). expect_eq_const is exact element equality; expect_almost_eq_const passes when every element differs by at most the tolerance (no explicit tolerance in this file, so the default applies), NaN matching NaN Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/dot_general.mlir`: StableHLO reference interpreter test dot_general.mlir: stablehlo.dot_general; 6 functions (si64, algorithm attribute, empty contracting dims, mixed operand and result types, f8E3M4, f8E4M3), 4 check.expect_eq_const and 2 check.expect_almost_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_eq_const and check.expect_almost_eq_const in stablehlo/tests/interpret/dot_general.mlir, written by the StableHLO authors against the op specification Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). expect_eq_const is exact element equality; expect_almost_eq_const passes when every element differs by at most the tolerance (no explicit tolerance in this file, so the default applies), NaN matching NaN Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/quantized_ops.mlir`: StableHLO reference interpreter test quantized_ops.mlir: uniform_quantize, uniform_dequantize and add on quantized types; 5 functions (two uniform_quantize, uniform_dequantize, quantize then dequantize, quantized add), 3 check.expect_eq_const and 2 check.expect_almost_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_eq_const and check.expect_almost_eq_const in stablehlo/tests/interpret/quantized_ops.mlir, written by the StableHLO authors against the op specification Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). expect_eq_const is exact element equality; expect_almost_eq_const passes when every element differs by at most the tolerance (no explicit tolerance in this file, so the default applies), NaN matching NaN Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/fft.mlir`: StableHLO reference interpreter test fft.mlir: stablehlo.fft (FFT, IFFT, RFFT and IRFFT over 1, 2 and 3 dimensions, complex and f64); 12 functions (fft, ifft, rfft and irfft in 1-D, 2-D and 3-D), 12 check.expect_almost_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_* op in stablehlo/tests/interpret/fft.mlir, written by the StableHLO authors against the op specification (stablehlo/docs/spec.md) Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_almost_eq_const, which passes when every element (real and imaginary parts for complex types) differs by at most the tolerance (no explicit tolerance in this file, so the default below applies), NaN matching NaN Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/gather.mlir`: StableHLO reference interpreter test gather.mlir: stablehlo.gather (offset, collapsed and batching dimensions); 2 functions (gather_op_test and gather_op_with_batching_dim_test), 2 check.expect_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_* op in stablehlo/tests/interpret/gather.mlir, written by the StableHLO authors against the op specification (stablehlo/docs/spec.md) Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_eq_const: exact element equality of the result with the inline constant, no tolerance Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/reduce_window.mlir`: StableHLO reference interpreter test reduce_window.mlir: stablehlo.reduce_window (window dimensions, strides, base and window dilations, padding); 2 functions (reduce_window and reduce_window_issue_1662), 2 check.expect_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_* op in stablehlo/tests/interpret/reduce_window.mlir, written by the StableHLO authors against the op specification (stablehlo/docs/spec.md) Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_eq_const: exact element equality of the result with the inline constant, no tolerance Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/select_and_scatter.mlir`: StableHLO reference interpreter test select_and_scatter.mlir: stablehlo.select_and_scatter (the max-pool gradient pattern: select by comparison, scatter by addition); 1 function (select_and_scatter_op_test), 1 check.expect_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_* op in stablehlo/tests/interpret/select_and_scatter.mlir, written by the StableHLO authors against the op specification (stablehlo/docs/spec.md) Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_eq_const: exact element equality of the result with the inline constant, no tolerance Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/triangular_solve.mlir`: StableHLO reference interpreter test triangular_solve.mlir: stablehlo.triangular_solve (left and right side, lower and upper, unit diagonal, transpose and adjoint, real and complex); 27 functions over left and right side, lower and upper, unit diagonal and transpose options, 27 check.expect_almost_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_* op in stablehlo/tests/interpret/triangular_solve.mlir, written by the StableHLO authors against the op specification (stablehlo/docs/spec.md) Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_almost_eq_const, which passes when every element (real and imaginary parts for complex types) differs by at most the tolerance (no explicit tolerance in this file, so the default below applies), NaN matching NaN Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. - `stablehlo/tests/interpret/sort.mlir`: StableHLO reference interpreter test sort.mlir: stablehlo.sort (stable multi-operand sort with a comparator region); 2 functions, 3 check.expect_eq_const. Format: MLIR text: func.func bodies with stablehlo ops on constant inputs and inline check.expect_* assertions Origin: hand-written conformance tests of the StableHLO reference interpreter - Reference result: inline_expected_constants - Reference source: the dense<...> constants of each check.expect_* op in stablehlo/tests/interpret/sort.mlir, written by the StableHLO authors against the op specification (stablehlo/docs/spec.md) Check: Run the file through stablehlo-translate --interpret -split-input-file (or translate each function for another backend and compare its outputs with the inline constants). Every check is check.expect_eq_const: exact element equality of the result with the inline constant, no tolerance Companions (not included): stablehlo/tests/CheckOps.td; stablehlo/tests/CheckOps.cpp; stablehlo/reference/Element.cpp; LICENSE Dependencies: stablehlo-translate built from openxla/stablehlo at commit c5e5abcf (RUN line 1 of the unit: stablehlo-translate --interpret -split-input-file). LLVM and MLIR at the revision pinned by build_tools/llvm_version.txt of the same commit. diff --git a/campaign_files/2026-10-ml-serving-configs-2/text-embeddings-inference-candle-snapshots/README.md b/campaign_files/2026-10-ml-serving-configs-2/text-embeddings-inference-candle-snapshots/README.md index b57d7506c871fd370b20576032037a6b6040d073..8f6235297c64760f284cc3f712e00de5a0daf69c 100644 --- a/campaign_files/2026-10-ml-serving-configs-2/text-embeddings-inference-candle-snapshots/README.md +++ b/campaign_files/2026-10-ml-serving-configs-2/text-embeddings-inference-candle-snapshots/README.md @@ -8,8 +8,6 @@ **Self-contained:** 0 of 16 files are self-contained; for the others it is not stated. 16 of 16 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_model_outputs. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Tools are named in the check written per file in the notes. **Licence (SPDX):** Apache-2.0 @@ -21,128 +19,96 @@ Files: - `backends/candle/tests/snapshots/test_bert__bert_single.snap`: Text-embeddings-inference candle backend test_bert.rs::test_bert: all-MiniLM-L6-v2 sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_bert__bert_single.snap: the expected embedding vector for the single input, written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_bert.rs lines 11 to 68; the batch snapshot backends/candle/tests/snapshots/test_bert__bert_batch.snap is checked by the same test Check: Run cargo test -p text-embeddings-backend-candle --test test_bert test_bert at the pinned commit on CPU with default features (the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on a CPU runner); or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding. Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_bert__bert_batch.snap; backends/candle/tests/test_bert.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 62); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub 0.4 (backends/candle/Cargo.toml lines 31 to 33); Rust toolchain from rust-toolchain.toml at the commit (to be pinned); Hugging Face model repo sentence-transformers/all-MiniLM-L6-v2 at main as written in the test (backends/candle/tests/test_bert.rs line 12: download_artifacts(..., None, None)); revision on 2026-10-07: 1110a243fdf4706b3f48f1d95db1a4f5529b4d41 (licence field: apache-2.0; gated: False) - `backends/candle/tests/snapshots/test_bert__emotions_single.snap`: Text-embeddings-inference candle backend test_bert.rs::test_emotions: roberta-base-go_emotions classifier logits (28 emotion labels, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_bert__emotions_single.snap: the expected score vector for the single input, written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_bert.rs lines 143 to 186; the batch snapshot backends/candle/tests/snapshots/test_bert__emotions_batch.snap is checked by the same test Check: Run cargo test -p text-embeddings-backend-candle --test test_bert test_emotions at the pinned commit on CPU with default features (the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on a CPU runner); or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: every score within 0.005 absolute. Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_bert__emotions_batch.snap; backends/candle/tests/test_bert.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 62); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub 0.4 (backends/candle/Cargo.toml lines 31 to 33); Rust toolchain from rust-toolchain.toml at the commit (to be pinned); Hugging Face model repo SamLowe/roberta-base-go_emotions at main as written in the test (backends/candle/tests/test_bert.rs line 144: download_artifacts(..., None, None)); revision on 2026-10-07: d75048347613a25d77de8cf6412eaae9fa7b26be (licence field: mit; gated: False) - `backends/candle/tests/snapshots/test_mpnet__mpnet_single.snap`: Text-embeddings-inference candle backend test_mpnet.rs::test_mpnet: all-mpnet-base-v2 sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_mpnet__mpnet_single.snap: the expected embedding vector for the single input, written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_mpnet.rs lines 11 to 69; the batch snapshot backends/candle/tests/snapshots/test_mpnet__mpnet_batch.snap is checked by the same test Check: Run cargo test -p text-embeddings-backend-candle --test test_mpnet test_mpnet at the pinned commit on CPU with default features (the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on a CPU runner); or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding. Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_mpnet__mpnet_batch.snap; backends/candle/tests/test_mpnet.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 62); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub 0.4 (backends/candle/Cargo.toml lines 31 to 33); Rust toolchain from rust-toolchain.toml at the commit (to be pinned); Hugging Face model repo sentence-transformers/all-mpnet-base-v2 at main as written in the test (backends/candle/tests/test_mpnet.rs line 13: download_artifacts(..., None, None)); revision on 2026-10-07: e8c3b32edf5434bc2275fc9bab85f82640a19130 (licence field: apache-2.0; gated: False) - `backends/candle/tests/snapshots/test_qwen3__qwen3_cpu_single.snap`: Text-embeddings-inference candle backend test_qwen3.rs::test_qwen3: Qwen3-Embedding-0.6B embeddings, CPU snapshot (float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_qwen3__qwen3_cpu_single.snap: the expected embedding vector for the single input, written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_qwen3.rs lines 11 to 52; the batch snapshot backends/candle/tests/snapshots/test_qwen3__qwen3_cpu_batch.snap is checked by the same test Check: Run cargo test -p text-embeddings-backend-candle --test test_qwen3 test_qwen3 at the pinned commit on CPU with default features (the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on a CPU runner); or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding. Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_qwen3__qwen3_cpu_batch.snap; backends/candle/tests/test_qwen3.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 62); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub 0.4 (backends/candle/Cargo.toml lines 31 to 33); Rust toolchain from rust-toolchain.toml at the commit (to be pinned); Hugging Face model repo Qwen/Qwen3-Embedding-0.6B at main as written in the test (backends/candle/tests/test_qwen3.rs line 12: download_artifacts(..., None, None)); revision on 2026-10-07: 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3 (licence field: apache-2.0; gated: False) - `backends/candle/tests/snapshots/test_gte__alibaba_gte_single.snap`: Text-embeddings-inference candle backend test_gte.rs::test_alibaba_gte: gte-base-en-v1.5 (GTE, English) sentence embeddings (CLS pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_gte__alibaba_gte_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_gte.rs line 47; the batch snapshot backends/candle/tests/snapshots/test_gte__alibaba_gte_batch.snap (3 vectors for the three inputs) is asserted at line 36 of the same test, and lines 48 and 49 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_gte, --exact test_alibaba_gte at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_gte__alibaba_gte_batch.snap; backends/candle/tests/test_gte.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo Alibaba-NLP/gte-base-en-v1.5 at main as written in the test (backends/candle/tests/test_gte.rs line 12: download_artifacts("Alibaba-NLP/gte-base-en-v1.5", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: a829fd0e060bb84554da0dfd354d0de0f7712b7f (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_gte__alibaba_new_gte_single.snap`: Text-embeddings-inference candle backend test_gte.rs::test_alibaba_gte_new: gte-multilingual-base (GTE new architecture) sentence embeddings (CLS pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_gte__alibaba_new_gte_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_gte.rs line 92; the batch snapshot backends/candle/tests/snapshots/test_gte__alibaba_new_gte_batch.snap (3 vectors for the three inputs) is asserted at line 81 of the same test, and lines 93 and 94 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_gte, --exact test_alibaba_gte_new at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_gte__alibaba_new_gte_batch.snap; backends/candle/tests/test_gte.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo Alibaba-NLP/gte-multilingual-base at main as written in the test (backends/candle/tests/test_gte.rs line 57: download_artifacts("Alibaba-NLP/gte-multilingual-base", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 9bbca17d9273fd0d03d5725c7a4b0f6b45142062 (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_gte__snowflake_gte_single.snap`: Text-embeddings-inference candle backend test_gte.rs::test_snowflake_gte: snowflake-arctic-embed-m-v2.0 (GTE architecture) sentence embeddings (CLS pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_gte__snowflake_gte_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_gte.rs line 138; the batch snapshot backends/candle/tests/snapshots/test_gte__snowflake_gte_batch.snap (3 vectors for the three inputs) is asserted at line 127 of the same test, and lines 139 and 140 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_gte, --exact test_snowflake_gte at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_gte__snowflake_gte_batch.snap; backends/candle/tests/test_gte.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo Snowflake/snowflake-arctic-embed-m-v2.0 at main as written in the test (backends/candle/tests/test_gte.rs line 102 to 103: download_artifacts("Snowflake/snowflake-arctic-embed-m-v2.0", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 95c2741480856aa9666782eb4afe11959938017f (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_jina__jina_single.snap`: Text-embeddings-inference candle backend test_jina.rs::test_jina_small: jina-embeddings-v2-small-en (JinaBERT, ALiBi) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_jina__jina_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_jina.rs line 46; the batch snapshot backends/candle/tests/snapshots/test_jina__jina_batch.snap (3 vectors for the three inputs) is asserted at line 35 of the same test, and lines 47 and 48 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_jina, --exact test_jina_small at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_jina__jina_batch.snap; backends/candle/tests/test_jina.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo jinaai/jina-embeddings-v2-small-en at main as written in the test (backends/candle/tests/test_jina.rs line 11: download_artifacts("jinaai/jina-embeddings-v2-small-en", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 44e7d1d6caec8c883c2d4b207588504d519788d0 (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_jina_code__jina_code_single.snap`: Text-embeddings-inference candle backend test_jina_code.rs::test_jina_code_base: jina-embeddings-v2-base-code (JinaBERT for code) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_jina_code__jina_code_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_jina_code.rs line 46; the batch snapshot backends/candle/tests/snapshots/test_jina_code__jina_code_batch.snap (3 vectors for the three inputs) is asserted at line 35 of the same test, and lines 47 and 48 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_jina_code, --exact test_jina_code_base at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_jina_code__jina_code_batch.snap; backends/candle/tests/test_jina_code.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo jinaai/jina-embeddings-v2-base-code at main as written in the test (backends/candle/tests/test_jina_code.rs line 11: download_artifacts("jinaai/jina-embeddings-v2-base-code", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 516f4baf13dec4ddddda8631e019b5737c8bc250 (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_modernbert__modernbert_single.snap`: Text-embeddings-inference candle backend test_modernbert.rs::test_modernbert: ModernBERT-base (mean pooling) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_modernbert__modernbert_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_modernbert.rs line 56 (CPU branch; the flash snapshot at line 54 needs features cuda and flash-attn, line 53); the batch snapshot backends/candle/tests/snapshots/test_modernbert__modernbert_batch.snap (3 vectors for the three inputs) is asserted at line 41 (CPU branch; the flash snapshot at line 39 needs features cuda and flash-attn, line 38) of the same test, and lines 58, 59 and 74 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_modernbert, --exact test_modernbert at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_modernbert__modernbert_batch.snap; backends/candle/tests/test_modernbert.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo answerdotai/ModernBERT-base at main as written in the test (backends/candle/tests/test_modernbert.rs line 14: download_artifacts("answerdotai/ModernBERT-base", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 8949b909ec900327062f0ebf497f51aef5e6f0c8 (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_nomic__nomic_single.snap`: Text-embeddings-inference candle backend test_nomic.rs::test_nomic_small: nomic-embed-text-v1.5 (NomicBERT) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_nomic__nomic_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_nomic.rs line 46; the batch snapshot backends/candle/tests/snapshots/test_nomic__nomic_batch.snap (3 vectors for the three inputs) is asserted at line 35 of the same test, and lines 47 and 48 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_nomic, --exact test_nomic_small at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_nomic__nomic_batch.snap; backends/candle/tests/test_nomic.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo nomic-ai/nomic-embed-text-v1.5 at main as written in the test (backends/candle/tests/test_nomic.rs line 11: download_artifacts("nomic-ai/nomic-embed-text-v1.5", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: e9b6763023c676ca8431644204f50c2b100d9aab (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_nomic__nomic_moe_single.snap`: Text-embeddings-inference candle backend test_nomic.rs::test_nomic_moe: nomic-embed-text-v2-moe (NomicBERT mixture of experts) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_nomic__nomic_moe_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_nomic.rs line 90; the batch snapshot backends/candle/tests/snapshots/test_nomic__nomic_moe_batch.snap (3 vectors for the three inputs) is asserted at line 79 of the same test, and lines 91 and 92 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_nomic, --exact test_nomic_moe at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_nomic__nomic_moe_batch.snap; backends/candle/tests/test_nomic.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo nomic-ai/nomic-embed-text-v2-moe at main as written in the test (backends/candle/tests/test_nomic.rs line 55: download_artifacts("nomic-ai/nomic-embed-text-v2-moe", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 1066b6599d099fbb93dfcb64f9c37a7c9e503e85 (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_pplx1__pplx1_cpu_single.snap`: Text-embeddings-inference candle backend test_pplx1.rs::test_pplx1: pplx-embed-v1-0.6b (Perplexity embedding model) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_pplx1__pplx1_cpu_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_pplx1.rs line 47; the batch snapshot backends/candle/tests/snapshots/test_pplx1__pplx1_cpu_batch.snap (3 vectors for the three inputs) is asserted at line 36 of the same test, and lines 48 and 49 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_pplx1, --exact test_pplx1 at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_pplx1__pplx1_cpu_batch.snap; backends/candle/tests/test_pplx1.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo perplexity-ai/pplx-embed-v1-0.6b at main as written in the test (backends/candle/tests/test_pplx1.rs line 12: download_artifacts("perplexity-ai/pplx-embed-v1-0.6b", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 2c4d510dd4a732063c31a0f70193e35067b51fd8 (licence field: mit; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_voyage_nano__voyage_nano_cpu_single.snap`: Text-embeddings-inference candle backend test_voyage_nano.rs::test_voyage_nano: voyage-4-nano (with 2048-wide projection) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_voyage_nano__voyage_nano_cpu_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_voyage_nano.rs line 51; the batch snapshot backends/candle/tests/snapshots/test_voyage_nano__voyage_nano_cpu_batch.snap (3 vectors for the three inputs) is asserted at line 40 of the same test, and lines 52 and 53 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_voyage_nano, --exact test_voyage_nano at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_voyage_nano__voyage_nano_cpu_batch.snap; backends/candle/tests/test_voyage_nano.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo voyageai/voyage-4-nano at main as written in the test (backends/candle/tests/test_voyage_nano.rs line 12: download_artifacts("voyageai/voyage-4-nano", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); revision on 2026-10-07: 67fabc9bef010dabc5f6024aa1b1b6b93410426f (licence field: apache-2.0; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_default_dense_single.snap`: Text-embeddings-inference candle backend test_dense.rs::test_stella_en_400m_v5_default_dense: stella_en_400M_v5 (default Dense head, 1024 wide) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_default_dense_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_dense.rs line 51 to 55; the batch snapshot backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_default_dense_batch.snap (3 vectors for the three inputs) is asserted at line 36 to 40 of the same test, and lines 56 and 57 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_dense, --exact test_stella_en_400m_v5_default_dense at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_default_dense_batch.snap; backends/candle/tests/test_dense.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo dunzhang/stella_en_400M_v5 at main as written in the test (backends/candle/tests/test_dense.rs line 12: download_artifacts("dunzhang/stella_en_400M_v5", None, None)): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); the test also downloads modules.json and the Dense projection folder of the model (common.rs lines 169 to 195); revision (the Hugging Face API resolves this id to NovaSearch/stella_en_400M_v5, the renamed repo) on 2026-10-07: ffeb2b7ee715c226d4ffe5e4619f7dbb48624c20 (licence field: mit; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs - `backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_dense_768_single.snap`: Text-embeddings-inference candle backend test_dense.rs::test_stella_en_400m_v5_dense_768: stella_en_400M_v5 (Dense head 2_Dense_768) sentence embeddings (mean pooling, float32), expected outputs in the insta snapshot. Format: insta YAML snapshot (list of float32 vectors) checked by a Rust integration test Origin: expected outputs recorded by the text-embeddings-inference maintainers from their own engine and checked into the repository as insta snapshots - Reference result: expected_model_outputs - Reference source: backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_dense_768_single.snap: the expected embedding vector for the single input "What is Deep Learning?", written by the maintainers with insta and checked in; asserted by backends/candle/tests/test_dense.rs line 117 to 121; the batch snapshot backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_dense_768_batch.snap (3 vectors for the three inputs) is asserted at line 97 to 101 of the same test, and lines 122 and 123 assert that batch rows 0 and 2 equal the single vector under the same matcher Check: Run cargo test -p text-embeddings-backend-candle --test test_dense, --exact test_stella_en_400m_v5_dense_768 at the pinned commit with default features (CPU; the CI job .github/workflows/test.yaml line 48 runs cargo test --profile=release-debug on the runner group named at lines 31 to 32), with the model revision of dependencies pinned through the Hugging Face cache; or have the agent produce the vectors with its own inference code and compare. Each produced vector must match the snapshot vector under the matcher of the test: cosine similarity above 0.999 per embedding, with the vector widths equal (common.rs line 59). Timings are never answers. Companions (not included): backends/candle/tests/snapshots/test_dense__stella_en_400m_v5_dense_768_batch.snap; backends/candle/tests/test_dense.rs; backends/candle/tests/common.rs; backends/candle/Cargo.toml; Cargo.toml; rust-toolchain.toml; .github/workflows/test.yaml; LICENSE Dependencies: text-embeddings-inference at commit 56964585069318d01ddeeec9454a952a66f1d3a2, crate text-embeddings-backend-candle built with default features (CPU, no CUDA or flash-attn); Rust toolchain 1.92.0 (rust-toolchain.toml line 2: channel = "1.92.0"); candle pinned at git rev 6381023982251959a2c9bab7378b3013304e192b (Cargo.toml lines 60 to 63); dev-dependencies insta at git rev f4f98c0410b91fb5a28b10df98e4422955be9c2c, is_close 0.1.3, hf-hub from the workspace with ureq, serial_test (backends/candle/Cargo.toml lines 31 to 36); Hugging Face model repo dunzhang/stella_en_400M_v5 at main as written in the test (backends/candle/tests/test_dense.rs line 65 to 66: download_artifacts("dunzhang/stella_en_400M_v5", None, Some("2_Dense_768"))): config.json, tokenizer.json and model.safetensors (common.rs lines 156 to 159); the test also downloads modules.json and the Dense projection folder of the model (common.rs lines 169 to 195); revision (the Hugging Face API resolves this id to NovaSearch/stella_en_400M_v5, the renamed repo) on 2026-10-07: ffeb2b7ee715c226d4ffe5e4619f7dbb48624c20 (licence field: mit; gated: False); TEI implements the architecture natively in candle, so no remote code of the model repo runs diff --git a/campaign_files/2026-10-ml-serving-configs-2/tfjs-core-op-tests/README.md b/campaign_files/2026-10-ml-serving-configs-2/tfjs-core-op-tests/README.md index bc50309fab3a975c38557930e0aef0209298f481..d34210e532b0d851cfae8dd62506b17717d62f6b 100644 --- a/campaign_files/2026-10-ml-serving-configs-2/tfjs-core-op-tests/README.md +++ b/campaign_files/2026-10-ml-serving-configs-2/tfjs-core-op-tests/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_tensors. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Run each case on the tfjs backend under test and compare the output with the expected array by expectArraysClose (test_util.ts): each element must satisfy |actual - expected| <= epsilon, where epsilon is 1e-3 on a float32 backend and 1e-1 on a float16 backend; shapes must match. **Licence (SPDX):** Apache-2.0 @@ -21,8 +19,6 @@ Files: - `tfjs-core/src/ops/local_response_normalization_test.ts`: TensorFlow.js op test local_response_normalization_test.ts: 17 localResponseNormalization cases (depth radius, bias, alpha, beta, 3D and 4D inputs, batches, gradients) with inline inputs and expected outputs computed with TensorFlow (Python). Format: TypeScript jasmine test file with inline input and expected arrays Origin: op test of TensorFlow.js written by the TensorFlow.js team, expected values computed with TensorFlow - Reference result: expected_output_tensors - Reference source: the expected arrays written inline in each it() case of the unit; comments in the file (for example line 159, t = tf.random_uniform([1, 3, 3, 8])) show they were computed with TensorFlow's tf.nn.local_response_normalization Check: Run each case on the tfjs backend under test and compare the output with the expected array by expectArraysClose (test_util.ts): each element must satisfy |actual - expected| <= epsilon, where epsilon is 1e-3 on a float32 backend and 1e-1 on a float16 backend; shapes must match Companions (not included): tfjs-core/src/test_util.ts; LICENSE Dependencies: @tensorflow/tfjs-core at commit 7f5309fe with its jasmine test setup. a tfjs backend to test, for example tfjs-backend-cpu. diff --git a/campaign_files/2026-10-ml-serving-configs-2/tflite-micro-examples-golden-checks/README.md b/campaign_files/2026-10-ml-serving-configs-2/tflite-micro-examples-golden-checks/README.md index 446a1de73b1aca4b3970d594cd8ce331c8ec6184..6f1031f8631c07a13668c3a355bedc244ccf7b2b 100644 --- a/campaign_files/2026-10-ml-serving-configs-2/tflite-micro-examples-golden-checks/README.md +++ b/campaign_files/2026-10-ml-serving-configs-2/tflite-micro-examples-golden-checks/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether each file needs other upstream files is not stated. 5 of 5 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_model_outputs; expected_labels. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Tools are named in the check written per file in the notes. **Licence (SPDX):** Apache-2.0 @@ -21,40 +19,30 @@ Files: - `tensorflow/lite/micro/examples/hello_world/models/hello_world_int8.tflite`: TFLite Micro hello_world: float and int8 sine-regression models must reproduce sin(x) within the published epsilon at fixed golden inputs. Format: TFLite flatbuffer models (float32 and int8) with a C++ golden-value test and a Python parity test Origin: example models trained by the TFLM maintainers on synthetic sine data and checked into the repository - Reference result: expected_model_outputs - Reference source: hello_world_test.cc: float model checked at golden inputs {0, 1, 3, 5} (line 90) and int8 model at golden inputs {0.77, 1.57, 2.3, 3.14}, quantised to {-96, -63, -34, 0} (lines 136 to 140), each against sin(x); evaluate_test.py lines 40 and 58 assert the TFLM predictions equal the TFLite interpreter predictions element for element Check: Run each model on CPU with any TFLite-compatible interpreter at the golden inputs: float model, |sin(x) - y| <= epsilon (hello_world_test.cc lines 88 and 96); int8 model, feed the int8 golden inputs, dequantise the output with the output tensor scale and zero point (line 145) and require |sin(x) - y| <= epsilon (lines 133 and 146). For a task, give the model files and the golden inputs and ask for the predictions; optional parity check: TFLM output equals TFLite output exactly (evaluate_test.py lines 40 and 58) Companions (not included): tensorflow/lite/micro/examples/hello_world/models/hello_world_float.tflite; tensorflow/lite/micro/examples/hello_world/hello_world_test.cc; tensorflow/lite/micro/examples/hello_world/evaluate_test.py; LICENSE Dependencies: tflite-micro at commit 5328e0d852c420f9494a53135529677229dfcdd2: the C++ test is built with make -f tensorflow/lite/micro/tools/make/Makefile test__test (reference kernels, x86 host, no hardware); the Python evaluate_test.py runs through the tflite_micro Python interpreter built from the same commit (bazel) and, where it compares, the TensorFlow Lite interpreter (tensorflow 2.x) - `tensorflow/lite/micro/examples/micro_speech/models/micro_speech_quantized.tflite`: TFLite Micro micro_speech: int8 keyword model plus audio preprocessor must label four 1 s clips (no, yes, noise, silence) as published. Format: TFLite flatbuffer models (keyword classifier and int8 audio feature generator) with four 16 kHz WAV clips and labelled Python and C++ tests Origin: keyword model and preprocessor trained by the TFLM maintainers; the README (line 247) says four audio samples from testdata/ are the test inputs; the clips most likely come from the Speech Commands dataset (CC BY 4.0), which the README does not state. - Reference result: expected_labels - Reference source: evaluate_test.py lines 41 to 46: (no, no_1000ms.wav), (yes, yes_1000ms.wav), (silence, noise_1000ms.wav), (silence, silence_1000ms.wav); the argmax category is compared with assertEqual at line 68 (int8 features) and line 90 (float features) Check: Compute the features of each clip with audio_preprocessor_int8.tflite (or the float preprocessor of the same commit), run micro_speech_quantized.tflite on CPU, take argmax over the category names of evaluate.get_category_names() and require exact label equality with the published pairs. For a task, give the models and clips, hide the label list Companions (not included): tensorflow/lite/micro/examples/micro_speech/models/audio_preprocessor_int8.tflite; tensorflow/lite/micro/examples/micro_speech/testdata/no_1000ms.wav; tensorflow/lite/micro/examples/micro_speech/testdata/yes_1000ms.wav; tensorflow/lite/micro/examples/micro_speech/testdata/noise_1000ms.wav; tensorflow/lite/micro/examples/micro_speech/testdata/silence_1000ms.wav; tensorflow/lite/micro/examples/micro_speech/evaluate_test.py; tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc; tensorflow/lite/micro/examples/micro_speech/README.md; LICENSE Dependencies: tflite-micro at commit 5328e0d852c420f9494a53135529677229dfcdd2: the C++ test is built with make -f tensorflow/lite/micro/tools/make/Makefile test__test (reference kernels, x86 host, no hardware); the Python evaluate_test.py runs through the tflite_micro Python interpreter built from the same commit (bazel) and, where it compares, the TensorFlow Lite interpreter (tensorflow 2.x) - `tensorflow/lite/micro/examples/mnist_lstm/trained_lstm_int8.tflite`: TFLite Micro mnist_lstm: float and int8 LSTM digit models must classify ten sample images as their digit, with published float-vs-int8 and TFLM-vs-TFLite bounds. Format: TFLite flatbuffer LSTM models (float32 and int8) with ten 28x28 PNG digit images labelled by file name, checked by a Python test Origin: LSTM models trained by the TFLM maintainers; the ten sample images look like MNIST digits (MNIST is distributed under CC BY-SA 3.0 by its authors); the repository states no source for them. - Reference result: expected_labels - Reference source: evaluate_test.py: sample.png has label k (lines 77 to 88 and 142 to 153); argmax equality at line 88 (float) and line 153 (int8); float vs dequantised int8 output bound at lines 138 to 140; TFLM vs TFLite float bound at line 75 Check: Run both models on CPU over the ten images: argmax must equal the digit in the file name for the float model (line 88) and the int8 model (line 153); the dequantised int8 output must stay within the published bound of the float output for every class (lines 138 to 140); optionally the TFLM float output must not exceed the TFLite output by the published bound (line 75). The int16 class at lines 159 to 224 is not part of the row (its model was not fetched) Companions (not included): tensorflow/lite/micro/examples/mnist_lstm/trained_lstm.tflite; tensorflow/lite/micro/examples/mnist_lstm/samples/sample0.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample1.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample2.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample3.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample4.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample5.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample6.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample7.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample8.png; tensorflow/lite/micro/examples/mnist_lstm/samples/sample9.png; tensorflow/lite/micro/examples/mnist_lstm/evaluate_test.py; LICENSE Dependencies: tflite-micro at commit 5328e0d852c420f9494a53135529677229dfcdd2: the C++ test is built with make -f tensorflow/lite/micro/tools/make/Makefile test__test (reference kernels, x86 host, no hardware); the Python evaluate_test.py runs through the tflite_micro Python interpreter built from the same commit (bazel) and, where it compares, the TensorFlow Lite interpreter (tensorflow 2.x) - `tensorflow/lite/micro/examples/dtln/dtln_inout_data.cc`: TFLite Micro DTLN noise suppression: int8 model output on the checked-in feature frame must equal the 257-value golden reference exactly. Format: C++ array file holding one int8 input feature frame and its golden output, with the int8 TFLite model and a C++ test Origin: model re-trained by Cadence on DNS Challenge data (README lines 7 to 8) and quantised to 8 bits; DNS Challenge data is mostly CC BY 4.0 per its repository; only the model weights and one feature frame are in the unit - Reference result: expected_model_outputs - Reference source: golden_ref array in dtln_inout_data.cc; dtln_test.cc lines 84 to 88 fix the output shape 1x1x257 int8 and line 93 compares every element Check: Run dtln_noise_suppression.tflite on CPU with the input frame of dtln_inout_data.cc; the int8 output tensor must have shape 1x1x257 and equal golden_ref element for element. For a task, give the model and input frame and hide golden_ref. The README (line 9) says the example is not for judging the network quality, so the row checks interpreter correctness only Companions (not included): tensorflow/lite/micro/examples/dtln/dtln_noise_suppression.tflite; tensorflow/lite/micro/examples/dtln/dtln_test.cc; tensorflow/lite/micro/examples/dtln/README.md; LICENSE Dependencies: tflite-micro at commit 5328e0d852c420f9494a53135529677229dfcdd2: the C++ test is built with make -f tensorflow/lite/micro/tools/make/Makefile test__test (reference kernels, x86 host, no hardware); the Python evaluate_test.py runs through the tflite_micro Python interpreter built from the same commit (bazel) and, where it compares, the TensorFlow Lite interpreter (tensorflow 2.x) - `tensorflow/lite/micro/models/person_detect.tflite`: TFLite Micro person_detection: int8 visual wake word model must score the person image above no-person and the reverse. Format: int8 TFLite MobileNet visual wake word model with two 96x96 grayscale BMP test images and a C++ ordering test Origin: visual wake word model trained by the TFLM maintainers; the two test images likely come from the COCO-derived Visual Wake Words data, not stated in the repository; flagged - Reference result: expected_labels - Reference source: person_detection_test.cc: for the person image person_score > no_person_score (line 100); for the no-person image no_person_score > person_score (line 124); the test reads the same images through generated person_image_data and no_person_image_data arrays, the BMP files are their checked-in sources Check: Convert each BMP to the int8 96x96x1 input, run the model on CPU and compare the two class scores: person.bmp must give person > no_person, no_person.bmp the reverse. For a task, give the model and images, hide which image is which Companions (not included): tensorflow/lite/micro/examples/person_detection/testdata/person.bmp; tensorflow/lite/micro/examples/person_detection/testdata/no_person.bmp; tensorflow/lite/micro/examples/person_detection/person_detection_test.cc; LICENSE Dependencies: tflite-micro at commit 5328e0d852c420f9494a53135529677229dfcdd2: the C++ test is built with make -f tensorflow/lite/micro/tools/make/Makefile test__test (reference kernels, x86 host, no hardware); the Python evaluate_test.py runs through the tflite_micro Python interpreter built from the same commit (bazel) and, where it compares, the TensorFlow Lite interpreter (tensorflow 2.x); the BMP to int8 conversion of the generated arrays (tensorflow/lite/micro/tools/generate_cc_arrays.py of the same commit) when the judge rebuilds the C++ test diff --git a/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-testdata/README.md b/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-testdata/README.md index 5ce38c9493e4686fbec5c9ae142d737abf21faa0..a63103936877bde71a247af7ee86e10989a5d0a8 100644 --- a/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-testdata/README.md +++ b/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-testdata/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_tensors. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Run the int8 Conv2D kernel under test on the input, filter and bias arrays of the unit with the quantization set up in conv_test.cc and compare each output element with the golden array as ValidateConvGoldens does (conv_test.cc lines 1010 to 1013, 1107 to 1110 and 1430 to 1433, tolerance argument 1.0). **Licence (SPDX):** Apache-2.0 @@ -21,8 +19,6 @@ Files: - `tensorflow/lite/micro/kernels/testdata/conv_test_data.cc`: TFLite Micro int8 Conv2D test vectors conv_test_data.cc: a 1x32x32x3 input, 8x3x3x3 filter and per channel bias with the 1x16x16x8 golden output, plus 4x4 and 5x5 SAME padding cases with golden outputs. Format: C++ source with const int8 and int32 arrays (inputs, filters, biases, golden outputs) Origin: kernel test vectors of TensorFlow Lite Micro, golden outputs from the TFLite reference kernel - Reference result: expected_output_tensors - Reference source: the golden arrays kConvGoldenOutput1x16x16x8 (line 305), kConvGoldenOutput4x4InputPaddingSame2x2 (line 496) and kConvGoldenOutput5x5InputPaddingSame3x3 (line 500) of the unit, produced by the TensorFlow Lite reference kernel Check: Run the int8 Conv2D kernel under test on the input, filter and bias arrays of the unit with the quantization set up in conv_test.cc and compare each output element with the golden array as ValidateConvGoldens does (conv_test.cc lines 1010 to 1013, 1107 to 1110 and 1430 to 1433, tolerance argument 1.0) Companions (not included): tensorflow/lite/micro/kernels/testdata/conv_test_data.h; tensorflow/lite/micro/kernels/conv_test.cc; LICENSE Dependencies: TensorFlow Lite for Microcontrollers kernel test build at commit 904193e6. diff --git a/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-tests/README.md b/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-tests/README.md index 40fb68e08b1fe91833642af035c3aedcb0ade2f3..26c7721c08bbd846d55a3a9ac021c8d2ed5226bc 100644 --- a/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-tests/README.md +++ b/campaign_files/2026-10-ml-serving-configs-2/tflm-kernel-tests/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_output_tensors. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Run the Softmax kernel under test on each case and compare every output element with the golden value by EXPECT_NEAR (line 263) with the tolerance of the case: 1e-5 for float (line 283), 1.0 default for quantized (line 292), 7.0 for int16 (line 34). **Licence (SPDX):** Apache-2.0 @@ -21,8 +19,6 @@ Files: - `tensorflow/lite/micro/kernels/softmax_test.cc`: TFLite Micro Softmax kernel test softmax_test.cc: float, int8 and int16 softmax cases (1D, 2D, 4D, beta values) with inline inputs and golden outputs. Format: C++ TFLM micro test source with inline arrays Origin: kernel test of TensorFlow Lite Micro written by the TensorFlow authors - Reference result: expected_output_tensors - Reference source: the golden output arrays written inline in each test of the unit (14 TF_LITE_MICRO_TEST cases), computed by the TensorFlow Lite reference softmax Check: Run the Softmax kernel under test on each case and compare every output element with the golden value by EXPECT_NEAR (line 263) with the tolerance of the case: 1e-5 for float (line 283), 1.0 default for quantized (line 292), 7.0 for int16 (line 34) Companions (not included): LICENSE Dependencies: TensorFlow Lite for Microcontrollers kernel test build at commit 904193e6. diff --git a/campaign_files/2026-10-op-rule-calculations-2/sylius-behat-inventory-features/README.md b/campaign_files/2026-10-op-rule-calculations-2/sylius-behat-inventory-features/README.md index 82353dd3d1f31c17beb717667b876ac324012b87..07e5196b0ae4617eea9730b74b08a06692871a7f 100644 --- a/campaign_files/2026-10-op-rule-calculations-2/sylius-behat-inventory-features/README.md +++ b/campaign_files/2026-10-op-rule-calculations-2/sylius-behat-inventory-features/README.md @@ -8,8 +8,6 @@ **Self-contained:** Whether each file needs other upstream files is not stated. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: scenario_then_steps. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Behat with the @api tag runs the scenario through the Sylius API Platform REST API without a browser; the same steps can be driven by hand against the admin REST API **Licence (SPDX):** MIT @@ -21,24 +19,18 @@ Files: - `features/admin/inventory/checkout_inventory/holding_inventory_units_during_checkout_process.feature`: Sylius feature: units on hold and on hand while an order is placed and after it is paid. Format: Gherkin feature file, 2 scenarios Origin: test written by the project authors; products, quantities and prices are made up - Reference result: scenario_then_steps - Reference source: Then steps: 3 on hold and 5 on hand after the order is placed (lines 22-23); 2 on hand and none on hold after payment (lines 31-32) Check: give the stock on hand and the order (placed, or placed and paid) and ask the units on hold and on hand; compare exactly Tools: Behat with the @api tag runs the scenario through the Sylius API Platform REST API without a browser; the same steps can be driven by hand against the admin REST API Dependencies: Sylius at commit c65f37f695af7314ca6ee902380edc55f9cf97a0 (branch 2.1), PHP 8, Symfony, Behat with the Sylius contexts, MySQL or PostgreSQL. - `features/admin/inventory/managing_inventory/verify_reserved_inventory_is_back_in_stock_on_order_cancelling.feature`: Sylius feature: on-hand and on-hold quantities after cancelling new, paid and refunded orders with one or several variants. Format: Gherkin feature file, 7 scenarios Origin: test written by the project authors; products, quantities and prices are made up - Reference result: scenario_then_steps - Reference source: Then steps: on hand and on hold per variant after cancellation (for example lines 28-29, 39-40) Check: give the starting stock per variant, the order lines and the order history (paid, refunded, cancelled) and ask the on-hand and on-hold quantities per variant; compare exactly Tools: Behat with the @api tag runs the scenario through the Sylius API Platform REST API without a browser; the same steps can be driven by hand against the admin REST API Dependencies: Sylius at commit c65f37f695af7314ca6ee902380edc55f9cf97a0 (branch 2.1), PHP 8, Symfony, Behat with the Sylius contexts, MySQL or PostgreSQL. - `features/admin/inventory/managing_inventory/decreasing_inventory_below_on_hold_validation.feature`: Sylius feature: whether on-hand stock may be lowered below the units on hold for a placed, cancelled or paid order. Format: Gherkin feature file, 3 scenarios Origin: test written by the project authors; products, quantities and prices are made up - Reference result: scenario_then_steps - Reference source: Then steps: rejected with stock staying 5 for a placed order (lines 25-26); accepted with stock 2 after cancellation (lines 34-35) and after payment (lines 43-44) Check: give the stock, the held units and the order state and ask whether setting on hand to 2 is accepted and what stock remains; compare the outcome and the count exactly Tools: Behat with the @api tag runs the scenario through the Sylius API Platform REST API without a browser; the same steps can be driven by hand against the admin REST API Dependencies: Sylius at commit c65f37f695af7314ca6ee902380edc55f9cf97a0 (branch 2.1), PHP 8, Symfony, Behat with the Sylius contexts, MySQL or PostgreSQL. diff --git a/campaign_files/2026-10-op-rule-calculations-2/tax-calculator-tests/README.md b/campaign_files/2026-10-op-rule-calculations-2/tax-calculator-tests/README.md index 0bb599e334411f772e7df6f282fc787973536764..2ba921ebc0f173027121531128bb71e1fb1f3ba4 100644 --- a/campaign_files/2026-10-op-rule-calculations-2/tax-calculator-tests/README.md +++ b/campaign_files/2026-10-op-rule-calculations-2/tax-calculator-tests/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: reference_amounts_from_official_style_worked_example. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** a Python script or a spreadsheet file written with openpyxl that applies the section 199A thresholds (joint 315,000, phase-in range 100,000), the SSTB applicable percentage and the W-2 wage and UBIA limits; or Tax-Calculator's own CLI (tc) on a CSV of the six units **Licence (SPDX):** CC0-1.0 @@ -21,8 +19,6 @@ Files: - `taxcalc/tests/test_calculator.py`: US section 199A qualified business income deduction for the six TPC example filing units (2018 law). Format: Python pytest file, 40,866 bytes; the task is test_qbid_calculation (line 813 on), whose six filing units are an inline CSV literal (TPC_FUNITS) with expected pre_qbid_taxinc and qbid Origin: the expected values are the worked numbers of Table 1 in Gale and Krupkin, 'Navigating the New Pass-Through Provisions: A Technical Explanation' (Tax Policy Center, 31 January 2018), copied into the test as the docstring says (lines 815 to 820); IRC section 199A as enacted in 2017. The expected values were not published by a tax authority: the Tax-Calculator developers copied them from a Tax Policy Center paper whose authors worked them out; they are used as given, with no separate official calibration. - Reference result: reference_amounts_from_official_style_worked_example - Reference source: TPC_FUNITS literal in test_calculator.py (lines 830 to 837): pre_qbid_taxinc and qbid for units 1 to 6 (15000.00, 1612.50, 0.00, 15000.00, 10750.00, 10000.00), taken from TPC Table 1 Check: give the agent the six units (filing status joint, wages, business income, SSTB flag, W-2 wages paid by the business, UBIA of property) and the 2018 section 199A rules; it reports each unit's QBI deduction; pass when numpy.allclose(answer, expected) holds with numpy defaults, the check the test itself runs Tools: a Python script or a spreadsheet file written with openpyxl that applies the section 199A thresholds (joint 315,000, phase-in range 100,000), the SSTB applicable percentage and the W-2 wage and UBIA limits; or Tax-Calculator's own CLI (tc) on a CSV of the six units Dependencies: numpy for the allclose check; version filled by relay task 051-T2 diff --git a/campaign_files/2026-10-s-security-detection-2/sshguard-sshguard-on-github/README.md b/campaign_files/2026-10-s-security-detection-2/sshguard-sshguard-on-github/README.md index 0743e2bb69c6ead270f3ff2ab9b1cdd5fa44f56e..b95f85257ba036fe9e4ff48f7e469a74d79a286c 100644 --- a/campaign_files/2026-10-s-security-detection-2/sshguard-sshguard-on-github/README.md +++ b/campaign_files/2026-10-s-security-detection-2/sshguard-sshguard-on-github/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 1 of 1 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: expected_attack_record. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Task: for each log line output the attack record (service code, attacker address, address family, dangerousness) or '*' when it is not an attack, or write the scanner rule that does so. Score: build sshg-parser and run test-sshg-parser; each case's output must equal the expected line exactly (string equality as in parser.c), all 218 cases. **Licence (SPDX):** ISC @@ -21,8 +19,6 @@ Files: - `src/parser/tests.txt`: SSHGuard attack parser test corpus: 218 log lines from SSH, mail, FTP, web, VPN, database and other services with the expected attack record (service, address, family, dangerousness) or no attack. Format: plain text test corpus (33,622 B, 732 lines): for each case an input log line, the expected output line ('service address family dangerousness', or '*' for no attack) and the test type M (expected match); the rules are the lex and yacc grammar in src/parser/attack_scanner.l and attack_parser.y Origin: log lines submitted by users and maintainers for each supported service, with addresses mostly replaced by placeholders, annotated with the expected parser output (recorded, edited) - Reference result: expected_attack_record - Reference source: src/parser/tests.txt: after each input line the next line is the expected parser output (e.g. line 4 '100 6.6.6.0 4 10' = service 100 (the code used throughout the SSH section), address, IPv4, dangerousness 10; '*' = no attack, 37 cases) and the line after that the type M (218 cases); the harness test_next_line in src/parser/parser.c (lines 41 to 90) compares by exact string equality Check: Task: for each log line output the attack record (service code, attacker address, address family, dangerousness) or '*' when it is not an attack, or write the scanner rule that does so. Score: build sshg-parser and run test-sshg-parser; each case's output must equal the expected line exactly (string equality as in parser.c), all 218 cases Companions (not included): COPYING; src/parser/attack_scanner.l; src/parser/attack_parser.y; src/parser/parser.c; src/parser/test-sshg-parser Dependencies: sshg-parser built from SSHGuard at this commit (flex, bison, a C compiler; the harness is src/parser/test-sshg-parser, which runs ./sshg-parser -t < tests.txt and prints TAP): needed only to rerun; version pinned when the pool is built; not packed diff --git a/campaign_files/2026-10-s-security-detection-2/stratus-red-team-detonation-logs/README.md b/campaign_files/2026-10-s-security-detection-2/stratus-red-team-detonation-logs/README.md index 2aa241bc76bc0052fe4b4fc3823f6d834c536501..482e1d74851034d2d0231c757a552267bfbc2baf 100644 --- a/campaign_files/2026-10-s-security-detection-2/stratus-red-team-detonation-logs/README.md +++ b/campaign_files/2026-10-s-security-detection-2/stratus-red-team-detonation-logs/README.md @@ -8,8 +8,6 @@ **Self-contained:** Not self-contained: each file is used together with other files from the same upstream. 32 of 32 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. -**Answers:** No answer files are included. Reference result: attack_technique_label. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Tools are named in the check written per file in the notes. **Licence (SPDX):** Apache-2.0 @@ -21,224 +19,160 @@ Files: - `docs/detonation-logs/aws.defense-evasion.cloudtrail-stop.json`: Stratus Red Team: AWS CloudTrail log of 'Stop CloudTrail Trail' (aws.defense-evasion.cloudtrail-stop). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,239 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.defense-evasion.cloudtrail-stop.md (3,034 B): its Mappings section (MITRE ATT&CK tactic Defense Evasion; Threat Technique Catalog for AWS T1562.008 (Impair Defenses: Disable Cloud Logs)) and its 'Detonation logs' section listing the CloudTrail events `cloudtrail:StopLogging` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic or the catalog technique id T1562.008; exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (cloudtrail:StopLogging). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.defense-evasion.cloudtrail-stop.md; LICENSE - `docs/detonation-logs/aws.persistence.iam-create-admin-user.json`: Stratus Red Team: AWS CloudTrail log of 'Create an administrative IAM User' (aws.persistence.iam-create-admin-user). Format: AWS CloudTrail events, JSON array of 3 record(s) (4,389 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.iam-create-admin-user.md (6,942 B): its Mappings section (MITRE ATT&CK tactic Persistence; Privilege Escalation; Threat Technique Catalog for AWS T1136.003 (Create Account: Create Cloud Account)) and its 'Detonation logs' section listing the CloudTrail events `iam:AttachUserPolicy`, `iam:CreateAccessKey`, `iam:CreateUser` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic (either listed tactic counts) or the catalog technique id T1136.003; exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (iam:AttachUserPolicy, iam:CreateAccessKey, iam:CreateUser). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. the doc page's Detection section: "Identify a call to CreateUser closely followed by AttachUserPolicy with an administrator policy" Companions (not included): docs/attack-techniques/AWS/aws.persistence.iam-create-admin-user.md; LICENSE - `docs/detonation-logs/aws.exfiltration.ec2-share-ebs-snapshot.json`: Stratus Red Team: AWS CloudTrail log of 'Exfiltrate EBS Snapshot by Sharing It' (aws.exfiltration.ec2-share-ebs-snapshot). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,566 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.exfiltration.ec2-share-ebs-snapshot.md (5,056 B): its Mappings section (MITRE ATT&CK tactic Exfiltration; Threat Technique Catalog for AWS T1578.001 (Modify Cloud Compute Infrastructure: Create Snapshot)) and its 'Detonation logs' section listing the CloudTrail events `ec2:ModifySnapshotAttribute` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic or the catalog technique id T1578.001; exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (ec2:ModifySnapshotAttribute). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.exfiltration.ec2-share-ebs-snapshot.md; LICENSE - `docs/detonation-logs/aws.lateral-movement.ec2-instance-connect.json`: Stratus Red Team: AWS CloudTrail log of 'Usage of EC2 Instance Connect on multiple instances' (aws.lateral-movement.ec2-instance-connect). Format: AWS CloudTrail events, JSON array of 3 record(s) (4,403 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.lateral-movement.ec2-instance-connect.md (6,774 B): its Mappings section (MITRE ATT&CK tactic Lateral Movement) and its 'Detonation logs' section listing the CloudTrail events `ec2-instance-connect:SendSSHPublicKey` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic; exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (ec2-instance-connect:SendSSHPublicKey). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.lateral-movement.ec2-instance-connect.md; LICENSE - `docs/detonation-logs/aws.credential-access.secretsmanager-batch-retrieve-secrets.json`: Stratus Red Team: AWS CloudTrail log of 'Retrieve a High Number of Secrets Manager secrets (Batch)' (aws.credential-access.secretsmanager-batch-retrieve-secrets). Format: AWS CloudTrail events, JSON array of 5 record(s) (6,752 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.credential-access.secretsmanager-batch-retrieve-secrets.md (10,323 B): its Mappings section (MITRE ATT&CK tactic Credential Access) and its 'Detonation logs' section listing the CloudTrail events `secretsmanager:BatchGetSecretValue` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic; exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (secretsmanager:BatchGetSecretValue). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.credential-access.secretsmanager-batch-retrieve-secrets.md; LICENSE - `docs/detonation-logs/aws.initial-access.console-login-without-mfa.json`: Stratus Red Team: AWS CloudTrail log of 'Console Login without MFA' (aws.initial-access.console-login-without-mfa). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,283 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.initial-access.console-login-without-mfa.md (4,225 B): its Mappings section (MITRE ATT&CK tactic Initial Access; Threat Technique Catalog for AWS T1078.A001 (Valid Accounts: IAM Users)) and its 'Detonation logs' section listing the CloudTrail events `signin:ConsoleLogin` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic or the catalog technique id T1078.A001; exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (signin:ConsoleLogin). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.initial-access.console-login-without-mfa.md; LICENSE - `docs/detonation-logs/aws.persistence.rolesanywhere-create-trust-anchor.json`: Stratus Red Team: AWS CloudTrail log of 'Create an IAM Roles Anywhere trust anchor' (aws.persistence.rolesanywhere-create-trust-anchor). Format: AWS CloudTrail events, JSON array of 2 record(s) (9,557 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.rolesanywhere-create-trust-anchor.md (11,902 B): its Mappings section (MITRE ATT&CK tactic Persistence; Privilege Escalation) and its 'Detonation logs' section listing the CloudTrail events `rolesanywhere:CreateProfile`, `rolesanywhere:CreateTrustAnchor` as generated by this technique; written by the publisher Check: Classification: given only the JSON log, name the ATT&CK tactic (either listed tactic counts); exact match on the name or id. Detection variant: list the eventName values that evidence the technique; exact set match against the page's Detonation logs list (rolesanywhere:CreateProfile, rolesanywhere:CreateTrustAnchor). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.rolesanywhere-create-trust-anchor.md; LICENSE - `docs/detonation-logs/aws.defense-evasion.cloudtrail-delete.json`: Stratus Red Team: AWS CloudTrail log of 'Delete CloudTrail Trail' (aws.defense-evasion.cloudtrail-delete). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,233 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.defense-evasion.cloudtrail-delete.md (2,849 B): its Mappings section (MITRE ATT&CK tactic Defense Evasion; Threat Technique Catalog for AWS T1562.008 (Impair Defenses: Disable Cloud Logs)) and its 'Detonation logs' section listing the CloudTrail events `cloudtrail:DeleteTrail` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Defense Evasion) or the catalog technique id T1562.008; exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (cloudtrail:DeleteTrail). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.defense-evasion.cloudtrail-delete.md; LICENSE - `docs/detonation-logs/aws.defense-evasion.vpc-remove-flow-logs.json`: Stratus Red Team: AWS CloudTrail log of 'Remove VPC Flow Logs' (aws.defense-evasion.vpc-remove-flow-logs). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,527 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.defense-evasion.vpc-remove-flow-logs.md (2,903 B): its Mappings section (MITRE ATT&CK tactic Defense Evasion) and its 'Detonation logs' section listing the CloudTrail events `ec2:DeleteFlowLogs` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Defense Evasion); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (ec2:DeleteFlowLogs). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.defense-evasion.vpc-remove-flow-logs.md; LICENSE - `docs/detonation-logs/aws.defense-evasion.organizations-leave.json`: Stratus Red Team: AWS CloudTrail log of 'Attempt to Leave the AWS Organization' (aws.defense-evasion.organizations-leave). Format: AWS CloudTrail events, JSON array of 3 record(s) (7,559 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.defense-evasion.organizations-leave.md (9,671 B): its Mappings section (MITRE ATT&CK tactic Defense Evasion; Threat Technique Catalog for AWS T1666.A002 (Modify Cloud Resource Hierarchy: Leave AWS Organization)) and its 'Detonation logs' section listing the CloudTrail events `organizations:LeaveOrganization`, `sts:AssumeRole` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Defense Evasion) or the catalog technique id T1666.A002; exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (organizations:LeaveOrganization, sts:AssumeRole). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.defense-evasion.organizations-leave.md; LICENSE - `docs/detonation-logs/aws.exfiltration.rds-share-snapshot.json`: Stratus Red Team: AWS CloudTrail log of 'Exfiltrate RDS Snapshot by Sharing' (aws.exfiltration.rds-share-snapshot). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,577 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.exfiltration.rds-share-snapshot.md (3,949 B): its Mappings section (MITRE ATT&CK tactic Exfiltration) and its 'Detonation logs' section listing the CloudTrail events `rds:ModifyDBSnapshotAttribute` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Exfiltration); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (rds:ModifyDBSnapshotAttribute). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.exfiltration.rds-share-snapshot.md; LICENSE - `docs/detonation-logs/aws.persistence.iam-backdoor-role.json`: Stratus Red Team: AWS CloudTrail log of 'Backdoor an IAM Role' (aws.persistence.iam-backdoor-role). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,632 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.iam-backdoor-role.md (3,867 B): its Mappings section (MITRE ATT&CK tactic Persistence; Threat Technique Catalog for AWS T1098.003 (Account Manipulation: Additional Cloud Roles)) and its 'Detonation logs' section listing the CloudTrail events `iam:UpdateAssumeRolePolicy` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence) or the catalog technique id T1098.003; exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (iam:UpdateAssumeRolePolicy). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.iam-backdoor-role.md; LICENSE - `docs/detonation-logs/aws.persistence.lambda-backdoor-function.json`: Stratus Red Team: AWS CloudTrail log of 'Backdoor Lambda Function Through Resource-Based Policy' (aws.persistence.lambda-backdoor-function). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,580 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.lambda-backdoor-function.md (3,348 B): its Mappings section (MITRE ATT&CK tactic Persistence) and its 'Detonation logs' section listing the CloudTrail events `lambda:AddPermission20150331v2` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (lambda:AddPermission20150331v2). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.lambda-backdoor-function.md; LICENSE - `docs/detonation-logs/aws.credential-access.ssm-retrieve-securestring-parameters.json`: Stratus Red Team: AWS CloudTrail log of 'Retrieve a High Number of SSM Parameters' (aws.credential-access.ssm-retrieve-securestring-parameters). Format: AWS CloudTrail events, JSON array of 10 record(s) (25,017 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.credential-access.ssm-retrieve-securestring-parameters.md (27,638 B): its Mappings section (MITRE ATT&CK tactic Credential Access) and its 'Detonation logs' section listing the CloudTrail events `ssm:DescribeParameters`, `ssm:GetParameters` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Credential Access); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (ssm:DescribeParameters, ssm:GetParameters). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.credential-access.ssm-retrieve-securestring-parameters.md; LICENSE - `docs/detonation-logs/aws.persistence.sts-federation-token.json`: Stratus Red Team: AWS CloudTrail log of 'Generate temporary AWS credentials using GetFederationToken' (aws.persistence.sts-federation-token). Format: AWS CloudTrail events, JSON array of 2 record(s) (4,478 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.sts-federation-token.md (6,932 B): its Mappings section (MITRE ATT&CK tactic Persistence; Threat Technique Catalog for AWS T1098.001 (Account Manipulation: Additional Cloud Credentials)) and its 'Detonation logs' section listing the CloudTrail events `sts:GetCallerIdentity`, `sts:GetFederationToken` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence) or the catalog technique id T1098.001; exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (sts:GetCallerIdentity, sts:GetFederationToken). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.sts-federation-token.md; LICENSE - `docs/detonation-logs/aws.impact.bedrock-invoke-model.json`: Stratus Red Team: AWS CloudTrail log of 'Invoke Bedrock Model' (aws.impact.bedrock-invoke-model). Format: AWS CloudTrail events, JSON array of 15 record(s) (25,403 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.impact.bedrock-invoke-model.md (30,063 B): its Mappings section (MITRE ATT&CK tactic Impact) and its 'Detonation logs' section listing the CloudTrail events `bedrock:CreateFoundationModelAgreement`, `bedrock:GetFoundationModelAvailability`, `bedrock:InvokeModel`, `bedrock:ListFoundationModelAgreementOffers`, `bedrock:PutFoundationModelEntitlement`, `bedrock:PutUseCaseForModelAccess` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Impact); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (bedrock:CreateFoundationModelAgreement, bedrock:GetFoundationModelAvailability, bedrock:InvokeModel, bedrock:ListFoundationModelAgreementOffers, bedrock:PutFoundationModelEntitlement, bedrock:PutUseCaseForModelAccess). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.impact.bedrock-invoke-model.md; LICENSE - `docs/detonation-logs/aws.execution.ssm-send-command.json`: Stratus Red Team: AWS CloudTrail log of 'Usage of ssm:SendCommand on multiple instances' (aws.execution.ssm-send-command). Format: AWS CloudTrail events, JSON array of 56 record(s) (81,644 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.execution.ssm-send-command.md (86,966 B): its Mappings section (MITRE ATT&CK tactic Execution) and its 'Detonation logs' section listing the CloudTrail events `ssm:DescribeInstanceInformation`, `ssm:GetCommandInvocation`, `ssm:SendCommand` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Execution); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (ssm:DescribeInstanceInformation, ssm:GetCommandInvocation, ssm:SendCommand). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.execution.ssm-send-command.md; LICENSE - `docs/detonation-logs/aws.privilege-escalation.iam-update-user-login-profile.json`: Stratus Red Team: AWS CloudTrail log of 'Change IAM user password' (aws.privilege-escalation.iam-update-user-login-profile). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,214 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.privilege-escalation.iam-update-user-login-profile.md (2,956 B): its Mappings section (MITRE ATT&CK tactic Privilege Escalation) and its 'Detonation logs' section listing the CloudTrail events `iam:UpdateLoginProfile` as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Privilege Escalation); exact match on the name or id. Detection variant: list the eventSource:eventName values that evidence the technique; exact set match against the page's Detonation logs list (iam:UpdateLoginProfile). Leakage to hide when building the task: the page embeds the same log, and the userAgent and some resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.privilege-escalation.iam-update-user-login-profile.md; LICENSE - `docs/detonation-logs/aws.credential-access.ec2-steal-instance-credentials.json`: Stratus Red Team: AWS CloudTrail log of 'Steal EC2 Instance Credentials' (aws.credential-access.ec2-steal-instance-credentials). Format: AWS CloudTrail events, JSON array of 45 record(s) (62,677 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.credential-access.ec2-steal-instance-credentials.md (67,395 B): its Mappings section (MITRE ATT&CK tactic Credential Access; Threat Technique Catalog for AWS T1552.005 (Unsecured Credentials: Cloud Instance Metadata API)) and its 'Detonation logs' section listing the CloudTrail events ec2:DescribeInstances, ssm:DescribeInstanceInformation, ssm:GetCommandInvocation, ssm:SendCommand as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Credential Access) or the catalog technique id T1552.005; exact match on the name or id. Detection variant: flag the 2 records whose userIdentity is the assumed-role session named after the instance (i-786a3A8B5C0d92eF4) and the ssm:SendCommand record that preceded them; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2:DescribeInstances, ssm:DescribeInstanceInformation, ssm:GetCommandInvocation, ssm:SendCommand). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.credential-access.ec2-steal-instance-credentials.md; LICENSE - `docs/detonation-logs/aws.defense-evasion.cloudtrail-event-selectors.json`: Stratus Red Team: AWS CloudTrail log of 'Disable CloudTrail Logging Through Event Selectors' (aws.defense-evasion.cloudtrail-event-selectors). Format: AWS CloudTrail events, JSON array of 1 record(s) (2,392 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.defense-evasion.cloudtrail-event-selectors.md (4,287 B): its Mappings section (MITRE ATT&CK tactic Defense Evasion; Threat Technique Catalog for AWS T1562.008 (Impair Defenses: Disable Cloud Logs)) and its 'Detonation logs' section listing the CloudTrail events cloudtrail:PutEventSelectors as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Defense Evasion) or the catalog technique id T1562.008; exact match on the name or id. Detection variant: the single PutEventSelectors record is the positive; the task is to recognise that its event selectors switch off management-event logging; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (cloudtrail:PutEventSelectors). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.defense-evasion.cloudtrail-event-selectors.md; LICENSE - `docs/detonation-logs/aws.execution.ec2-launch-unusual-instances.json`: Stratus Red Team: AWS CloudTrail log of 'Launch Unusual EC2 instances' (aws.execution.ec2-launch-unusual-instances). Format: AWS CloudTrail events, JSON array of 3 record(s) (8,905 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.execution.ec2-launch-unusual-instances.md (11,329 B): its Mappings section (MITRE ATT&CK tactic Execution; Threat Technique Catalog for AWS T1496.A008 (Resource Hijacking: Compute Hijacking - EC2 Use) and T1578.002 (Modify Cloud Compute Infrastructure: Create Cloud Instance)) and its 'Detonation logs' section listing the CloudTrail events ec2:RunInstances, sts:AssumeRole as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Execution) or the catalog technique id T1496.A008; exact match on the name or id. Detection variant: flag the ec2:RunInstances record for a GPU instance type with maxCount 10, even though it failed; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2:RunInstances, sts:AssumeRole). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.execution.ec2-launch-unusual-instances.md; LICENSE - `docs/detonation-logs/aws.exfiltration.ec2-security-group-open-port-22-ingress.json`: Stratus Red Team: AWS CloudTrail log of 'Open Ingress Port 22 on a Security Group' (aws.exfiltration.ec2-security-group-open-port-22-ingress). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,933 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.exfiltration.ec2-security-group-open-port-22-ingress.md (3,757 B): its Mappings section (MITRE ATT&CK tactic Exfiltration; Threat Technique Catalog for AWS T1562.007 (Impair Defenses: Disable or Modify Cloud Firewall)) and its 'Detonation logs' section listing the CloudTrail events ec2:AuthorizeSecurityGroupIngress as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Exfiltration) or the catalog technique id T1562.007; exact match on the name or id. Detection variant: the single AuthorizeSecurityGroupIngress record is the positive; the task is to read fromPort 22, toPort 22 and the /0 prefix; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2:AuthorizeSecurityGroupIngress). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.exfiltration.ec2-security-group-open-port-22-ingress.md; LICENSE - `docs/detonation-logs/aws.persistence.iam-backdoor-user.json`: Stratus Red Team: AWS CloudTrail log of 'Create an Access Key on an IAM User' (aws.persistence.iam-backdoor-user). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,428 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.iam-backdoor-user.md (3,078 B): its Mappings section (MITRE ATT&CK tactic Persistence and Privilege Escalation; Threat Technique Catalog for AWS T1098.001 (Account Manipulation: Additional Cloud Credentials)) and its 'Detonation logs' section listing the CloudTrail events iam:CreateAccessKey as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence) or the catalog technique id T1098.001; exact match on the name or id. Detection variant: the single CreateAccessKey record is the positive; the task is to see that the key is created for a user other than the caller; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (iam:CreateAccessKey). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.iam-backdoor-user.md; LICENSE - `docs/detonation-logs/aws.persistence.iam-create-backdoor-role.json`: Stratus Red Team: AWS CloudTrail log of 'Create a backdoored IAM Role' (aws.persistence.iam-create-backdoor-role). Format: AWS CloudTrail events, JSON array of 2 record(s) (3,747 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.iam-create-backdoor-role.md (7,058 B): its Mappings section (MITRE ATT&CK tactic Persistence; Threat Technique Catalog for AWS T1098.003 (Account Manipulation: Additional Cloud Roles)) and its 'Detonation logs' section listing the CloudTrail events iam:AttachRolePolicy, iam:CreateRole as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence) or the catalog technique id T1098.003; exact match on the name or id. Detection variant: flag both records; the task is to read the external account principal in assumeRolePolicyDocument and the AdministratorAccess policy ARN; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (iam:AttachRolePolicy, iam:CreateRole). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.iam-create-backdoor-role.md; LICENSE - `docs/detonation-logs/aws.persistence.iam-create-user-login-profile.json`: Stratus Red Team: AWS CloudTrail log of 'Create a Login Profile on an IAM User' (aws.persistence.iam-create-user-login-profile). Format: AWS CloudTrail events, JSON array of 3 record(s) (3,985 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.iam-create-user-login-profile.md (6,313 B): its Mappings section (MITRE ATT&CK tactic Persistence and Privilege Escalation; Threat Technique Catalog for AWS T1098.001 (Account Manipulation: Additional Cloud Credentials)) and its 'Detonation logs' section listing the CloudTrail events iam:CreateLoginProfile, iam:DeleteLoginProfile, sts:GetCallerIdentity as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence) or the catalog technique id T1098.001; exact match on the name or id. Detection variant: flag the CreateLoginProfile record with passwordResetRequired false on a user other than the caller; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (iam:CreateLoginProfile, iam:DeleteLoginProfile, sts:GetCallerIdentity). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.persistence.iam-create-user-login-profile.md; LICENSE - `docs/detonation-logs/aws.credential-access.ec2-get-password-data.json`: Stratus Red Team: AWS CloudTrail log of 'Retrieve EC2 Password Data' (aws.credential-access.ec2-get-password-data). Format: AWS CloudTrail events, JSON array of 34 record(s) (94,254 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.credential-access.ec2-get-password-data.md (97,658 B): its Mappings section (MITRE ATT&CK tactic Credential Access (the page names no catalog technique)) and its 'Detonation logs' section listing the CloudTrail events ec2:GetPasswordData, sts:AssumeRole as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Credential Access) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag the 30 ec2:GetPasswordData records (a burst on many instance ids in seconds); the AssumeRole records are context; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2:GetPasswordData, sts:AssumeRole). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.credential-access.ec2-get-password-data.md; LICENSE - `docs/detonation-logs/aws.execution.ssm-start-session.json`: Stratus Red Team: AWS CloudTrail log of 'Usage of ssm:StartSession on multiple instances' (aws.execution.ssm-start-session). Format: AWS CloudTrail events, JSON array of 29 record(s) (41,692 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.execution.ssm-start-session.md (45,353 B): its Mappings section (MITRE ATT&CK tactic Execution (the page names no catalog technique)) and its 'Detonation logs' section listing the CloudTrail events ssm:DescribeInstanceInformation, ssm:StartSession, ssm:TerminateSession as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Execution) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag the 3 ssm:StartSession records (one user, 3 distinct target instances), exact set match on the records; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ssm:DescribeInstanceInformation, ssm:StartSession, ssm:TerminateSession). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.execution.ssm-start-session.md; LICENSE - `docs/detonation-logs/aws.lateral-movement.ec2-serial-console-send-ssh-public-key.json`: Stratus Red Team: AWS CloudTrail log of 'Usage of EC2 Serial Console to push SSH public key' (aws.lateral-movement.ec2-serial-console-send-ssh-public-key). Format: AWS CloudTrail events, JSON array of 4 record(s) (8,483 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.lateral-movement.ec2-serial-console-send-ssh-public-key.md (11,158 B): its Mappings section (MITRE ATT&CK tactic Lateral Movement (the page names no catalog technique)) and its 'Detonation logs' section listing the CloudTrail events ec2-instance-connect:SendSerialConsoleSSHPublicKey, ec2:EnableSerialConsoleAccess as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Lateral Movement) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag the EnableSerialConsoleAccess record and the 3 SendSerialConsoleSSHPublicKey records; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2-instance-connect:SendSerialConsoleSSHPublicKey, ec2:EnableSerialConsoleAccess). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names contain 'stratus-red-team' or the technique words; strip or rename them, never show the page. Companions (not included): docs/attack-techniques/AWS/aws.lateral-movement.ec2-serial-console-send-ssh-public-key.md; LICENSE - `docs/detonation-logs/aws.credential-access.secretsmanager-retrieve-secrets.json`: Stratus Red Team: AWS CloudTrail log of 'Retrieve a High Number of Secrets Manager secrets' (aws.credential-access.secretsmanager-retrieve-secrets). Format: AWS CloudTrail events, JSON array of 21 record(s) (27,489 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.credential-access.secretsmanager-retrieve-secrets.md (30,208 B): its Mappings section (MITRE ATT&CK tactic Credential Access) and its 'Detonation logs' section listing the CloudTrail events secretsmanager:GetSecretValue, secretsmanager:ListSecrets as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Credential Access) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag all 21 records (the 20 GetSecretValue calls and the ListSecrets call); or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (secretsmanager:GetSecretValue, secretsmanager:ListSecrets). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names carry the string stratus-red-team; strip them. Companions (not included): docs/attack-techniques/AWS/aws.credential-access.secretsmanager-retrieve-secrets.md; LICENSE - `docs/detonation-logs/aws.discovery.ec2-download-user-data.json`: Stratus Red Team: AWS CloudTrail log of 'Download EC2 Instance User Data' (aws.discovery.ec2-download-user-data). Format: AWS CloudTrail events, JSON array of 19 record(s) (51,627 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.discovery.ec2-download-user-data.md (54,566 B): its Mappings section (MITRE ATT&CK tactic Discovery) and its 'Detonation logs' section listing the CloudTrail events ec2:DescribeInstanceAttribute, sts:AssumeRole as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Discovery) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag the 15 DescribeInstanceAttribute records and the 4 AssumeRole records; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2:DescribeInstanceAttribute, sts:AssumeRole). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names carry the string stratus-red-team; strip them. Companions (not included): docs/attack-techniques/AWS/aws.discovery.ec2-download-user-data.md; LICENSE - `docs/detonation-logs/aws.exfiltration.ec2-share-ami.json`: Stratus Red Team: AWS CloudTrail log of 'Exfiltrate an AMI by Sharing It' (aws.exfiltration.ec2-share-ami). Format: AWS CloudTrail events, JSON array of 1 record(s) (1,541 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.exfiltration.ec2-share-ami.md (3,365 B): its Mappings section (MITRE ATT&CK tactic Exfiltration) and its 'Detonation logs' section listing the CloudTrail events ec2:ModifyImageAttribute as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Exfiltration) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag the single ModifyImageAttribute record and name the external account 846424999548 added to launchPermission; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (ec2:ModifyImageAttribute). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names carry the string stratus-red-team; strip them. Companions (not included): docs/attack-techniques/AWS/aws.exfiltration.ec2-share-ami.md; LICENSE - `docs/detonation-logs/aws.persistence.lambda-overwrite-code.json`: Stratus Red Team: AWS CloudTrail log of 'Overwrite Lambda Function Code' (aws.persistence.lambda-overwrite-code). Format: AWS CloudTrail events, JSON array of 1 record(s) (3,198 B) Origin: CloudTrail logs captured from a real detonation of the technique in Datadog's test AWS account with Grimoire, then anonymised with LogLicker (page footnote); published as documentation, not written by hand - Reference result: attack_technique_label - Reference source: companion page docs/attack-techniques/AWS/aws.persistence.lambda-overwrite-code.md (4,870 B): its Mappings section (MITRE ATT&CK tactic Persistence) and its 'Detonation logs' section listing the CloudTrail events lambda:UpdateFunctionCode20150331v2 as generated by this technique; written by the publisher; checked on this round that the eventSource:eventName set of the log equals that list Check: Classification: given only the JSON log, name the ATT&CK tactic (Persistence) (the page maps no catalog technique id, so only the tactic is scored); exact match on the name or id. Detection variant: flag the single UpdateFunctionCode20150331v2 record; or list the eventSource:eventName values that evidence the technique, exact set match against the page's Detonation logs list (lambda:UpdateFunctionCode20150331v2). Leakage to hide when building the task: the page embeds the same log, and the userAgent and resource names carry the string stratus-red-team; strip them. Companions (not included): docs/attack-techniques/AWS/aws.persistence.lambda-overwrite-code.md; LICENSE diff --git a/campaign_files/2026-10-s-security-detection-2/tcp1p-ctf-2024-public-challenges/README.md b/campaign_files/2026-10-s-security-detection-2/tcp1p-ctf-2024-public-challenges/README.md index a599dc652d527d63ae222bc84f7d634a505be699..38044e5eb4e9900a42fcdb2a906355a87109ce66 100644 --- a/campaign_files/2026-10-s-security-detection-2/tcp1p-ctf-2024-public-challenges/README.md +++ b/campaign_files/2026-10-s-security-detection-2/tcp1p-ctf-2024-public-challenges/README.md @@ -8,8 +8,6 @@ **Self-contained:** 1 of 2 files are self-contained; the others are used together with other upstream files. 1 of 2 files name companion files (other upstream files read alongside); companions are not included in this delivery and are listed per file in the notes. Dependencies named upstream are listed per file in the notes. -**Answers:** No answer files are included. Reference result: flag_string. Where the reference result comes from and how a solution is checked is written per file in the notes. - **Tools to open and check:** Check: Exact string match of the submitted flag against the challenge yaml flags value; no tolerance. **Licence (SPDX):** NOASSERTION @@ -21,15 +19,11 @@ Files: - `Forensics/EncryptDecryptFile/dist/dist.zip`: TCP1P CTF 2024 Forensics/EncryptDecryptFile: an encryptor and a Mercurial (.hg) repository ship together; recover the flag by reversing the encryption using the version-control history. Format: one unencrypted ZIP (23,044 B, 31 entries): main.py (the encryptor, read-only, never run) plus a .hg Mercurial repository whose history retains flag.enc before encryption Origin: official competition instance built by the organisers for TCP1P CTF 2024, a public CTF, and released with its flag and official solution (fixture of a competition, not recorded production telemetry) - Reference result: flag_string - Reference source: challenge.yaml line 12 gives TCP1P{introduction_to_hg_a82ffbe612}; solver/README.md documents the hg-history recovery. The flag is hidden from the solver and is not plaintext in the unit ZIP. Check: Exact string match of the submitted flag against the challenge yaml flags value; no tolerance Dependencies: the challenge yaml / FLAG.txt / writeup hold the flag and the official solution path (hidden from the solver); solver tools are free choices (for example an archive, image, pcap or office-format library), version fixed when the pool is built - `Forensics/Skibidi Format/dist/suisei.skibidi`: TCP1P CTF 2024 Forensics/Skibidi Format: decode a bespoke binary container (magic SKB1) per its distributed spec to recover the embedded image carrying the flag. Format: one custom binary file suisei.skibidi (12,440,903 B, magic "SKB1") plus dist/spec.html (18,383 B, no