CONTRIBUTIONS.md

Upstream contribution record

Snapshot: 30 October 2026. Account: Afloat16. Display dates: Asia/Shanghai; machine-readable timestamps retain UTC.

82 verified merged author PRs · 24 upstream repositories · 15 latest verified top-30 contributor-list positions

TD-MPC2 adds one separately verified authored-code adoption, for 25 upstream projects with accepted contributions. Across the merged PRs, 31 are in 8 repositories with more than 10,000 current stars. The five profile highlights are an editorial selection from this set.

Verification method

• The first author:Afloat16 is:pr is:merged -user:Afloat16 search returned 80 results with incomplete_results=false. An independent all-status author-PR search returned 509 unique PRs across six pages, ordered by creation time; 82 had pull_request.merged_at. Geoopt #258 and #259 were missing from the first search index. Every counted PR was then checked at its direct REST endpoint for merged=true, author Afloat16 and a merge timestamp. Default-branch author-commit lists and commit-to-PR associations were cross-checked for additional omissions. The result is a verified inventory at retrieval time, not a guarantee against future indexing changes.
• Only public upstream repositories are counted. Account-owned PRs, unmerged PRs and maintainer-authored integration PRs are excluded. Counts include code, tests and documentation; they do not imply 82 separate bug fixes.
• Stars come from official repository metadata and belong to the upstream projects. They are not stars earned by my own repositories and do not measure an individual contribution's quality.
• Contributor positions are one-based positions in GitHub's /repos/{owner}/{repo}/contributors?per_page=100&anon=false responses, ordered by commit contributions, including bots and excluding anonymous contributors. Pagination offsets are included when later pages are retrieved. Equal counts retain GitHub's returned order. These are not PR-count rankings, maintainer roles, lines-of-code rankings or quality assessments.
• GitHub may cache contributor data. Its email-to-account attribution also has documented limits. — means no verified entry in the pages retrieved, not zero commits and not a fabricated numerical rank. The snapshot records retrieval status, pages, counts and source URLs.
• The full index’s Current author commits column uses the independently enumerated /commits?author=Afloat16 default-branch history, refreshed on 30 October. This can differ from contributor statistics and is not the count used to determine the contributor position. The full index retains both metrics.
• Contributor refresh succeeded for all 25 of 25 repositories, with 25 exact positions. All positions were rechecked on 30 October; pagination was followed for LeRobot, Haystack, Kornia and TileLang. All 15 top-30 merged-PR projects have current verification dates. TD-MPC2 is separately listed at #3. The machine-readable snapshot preserves per-repository retrieval times, page offsets and source URLs.
• The homepage preserves top-30 placements first and sorts each group by current stars. The five selected merges require current stars >10,000, verified authorship and merged state, and an inspected production diff with a substantive correctness or state-management change. They are sorted by merged_at, not PR creation or last update time, and are not claimed to be the newest five consecutive PRs.

Sources: Contributor API documentation · Author PR search · Current source snapshot · Previous record.

This refresh updates public metadata and retains the previously inspected selected-diff evidence. It does not rerun upstream test suites or enlarge the original PRs' validation claims.

Five recent selected merges

Merged	Upstream project · ★	Engineering contribution	PR / merged commit
2026-10-27	REA 409,561 ★	Reject malformed UTF-8 before CDP discovery decodes JSON, preserving target titles and URLs.	#4524 095a51c
2026-10-23	DeepFace 23,613 ★	Align face-region bounds and anti-spoofing ROIs with exclusive image-crop coordinates.	#1658 72aca24
2026-10-22	LeRobot 28,587 ★	Apply the established pairwise RoPE correction to VLA-JEPA's training-only world predictor.	#4835 f9e9d73
2026-10-22	REA 409,561 ★	Complete short positional ZIP reads and report premature EOF as an integrity failure.	#4414 82e4e03
2026-10-16	Haystack 26,881 ★	Isolate resumed execution counters so pipeline snapshots remain reusable after retries and breakpoints.	#13076 34ab691

REA #4524

Protocol input integrity; complete-body validation, split multibyte input and recovery regressions.

Scope: Ordinary validation correction, not a new security capability or performance claim. Loopback tests do not establish real-browser qualification.

Evidence: merged PR · upstream commit; merge commit author: Afloat16.

DeepFace #1658

Cross-path geometry contract, singleton/full-frame cases and anti-spoof ROI regression coverage.

Scope: No recognition-accuracy or anti-spoofing-quality improvement claim; broad optional model/backend suite was not run.

Evidence: merged PR · upstream commit; merge commit author: Afloat16.

LeRobot #4835

Rotary position geometry; norm, inner-product and relative-position invariants; single-frequency broadcasting.

Scope: Correction originated in echosprint's V-JEPA 2 PR #15; not an original algorithm. Training-only predictor, checkpoint outputs/losses change; no pretrained-quality or real-robot performance claim.

Evidence: merged PR · upstream commit; merge commit author: Afloat16.

REA #4414

Low-level I/O correctness; offset loop, truncation handling and bounded archive extraction regressions.

Scope: No general throughput or full-platform qualification claim; tests inject short reads on real descriptors.

Evidence: merged PR · upstream commit; merge commit author: Afloat16.

Haystack #13076

Checkpoint state ownership, replay determinism, downstream-failure retry and prior-snapshot isolation.

Scope: Focused pipeline checks only; no claim of full repository/integration suite validation.

Evidence: merged PR · upstream commit; merge commit author: Afloat16.

Recent candidates retained in the full record but not highlighted:

• tile-ai/tilelang #3357: 8,852 stars is below the strict >10,000 threshold.
• recommenders-team/recommenders #2388: Valid contribution, but a basic nonpositive parameter check is less representative than checkpoint state ownership; retained in full record.
• morluto/rea #3919: Valid contribution, but straightforward client-list deduplication is less representative; retained in full record.

Complete merged-PR index

Upstream repository	Stars	Position · checked	Contributor-list count	Current author commits	Merged PRs	Source PRs
serengil/deepface	23,613	#9 2026-10-30	18	18	18	#1658, #1655, #1651, #1653, #1652, #1649, #1654, #1648, #1646, #1647, #1643, #1644, #1634, #1638, #1636, #1633, #1631, #1630
lucidrains/denoising-diffusion-pytorch	10,721	#9 2026-10-30	1	1	1	#373
Eigenwise/atomic-agents	6,313	#28 2026-10-30	2	2	1	#292
petercorke/robotics-toolbox-python	3,557	#6 2026-10-30	11	11	11	#801, #800, #799, #798, #797, #796, #795, #794, #793, #792, #719
leggedrobotics/rsl_rl	3,106	#5 2026-10-30	3	3	3	#244, #240, #239
pypose/pypose	1,620	#15 2026-10-30	6	6	4	#437, #415, #416, #417
Denys88/rl_games	1,408	#24 2026-10-30	1	1	1	#384
geoopt/geoopt	1,102	#3 2026-10-30	10	10	10	#254, #253, #255, #256, #258, #259, #257, #267, #266, #262
Toni-SM/skrl	1,100	#3 2026-10-30	3	3	3	#469, #467, #475
dfki-ric/pytransform3d	707	#15 2026-10-30	2	2	2	#389, #383
naver/roma	647	#2 2026-10-30	4	4	4	#33, #32, #30, #29
lucidrains/rectified-flow-pytorch	489	#3 2026-10-30	1	1	1	#50
neka-nat/kinpy	181	#2 2026-10-30	3	3	3	#29, #28, #27
danijar/elements	73	#3 2026-10-30	2	2	2	#18, #19
stephane-caron/qpmpc	55	#2 2026-10-30	5	5	3	#17, #16, #15
morluto/rea	409,561	#74 2026-10-30	3	3	3	#4524, #4414, #3919
huggingface/lerobot	28,587	#224 2026-10-30	1	1	1	#4835
deepset-ai/haystack	26,881	#365 2026-10-30	1	1	1	#13076
recommenders-team/recommenders	21,971	#49 2026-10-30	6	6	2	#2388, #2389
huggingface/sentence-transformers	19,236	#54 2026-10-30	3	3	3	#4127, #4116, #4113
kornia/kornia	11,459	#196 2026-10-30	2	2	2	#5517, #5474
tile-ai/tilelang	8,852	#210 2026-10-30	1	1	1	#3357
rasbt/mlxtend	5,190	#95 2026-10-30	1	1	1	#1207
geomstats/geomstats	1,528	#82 2026-10-30	1	1	1	#2131

Changes since 23 October 2026

Project	Verified merged author PRs	Newly verified PRs
REA	2 → 3	#4524
DeepFace	17 → 18	#1658
TileLang	0 → 1	#3357
RSL-RL	2 → 3	#244
RL Games	0 → 1	#384
Geoopt	4 → 10	#254, #253, #255, #256, #258, #259
qpmpc	1 → 3	#17, #16

The verified inventory increased from 69 to 82 PRs, and from 22 to 24 merged-PR repositories. Earlier snapshots remain available unchanged.

Other accepted work

These changes entered upstream through maintainer integration. Their original author PRs were closed without being directly merged, so they are excluded from the 82 merged-PR total.

Project / original PR	Adoption evidence	Attribution
TD-MPC2 #98	Implementation cherry-pick; the authored commit remains in the verified default-branch history. Target-Q history and evaluation mode are preserved during model conversion.	Commit author is Afloat16. This separately adopted contribution brings the accepted-project total to 25; current stars: 993, contributor-list position #3 / 1 commit contribution.
Robotics Toolbox #735	Adopted by merged #883, 22 October 2026; squash commit. Circular particle-heading moments.	Maintainer-authored squash retains Co-authored-by: Afloat <132809511+Afloat16@users.noreply.github.com>.
Robotics Toolbox #736	Adopted by merged #805, 16 October 2026 (Asia/Shanghai); squash commit. Normalized-time quintic trajectories and the Hermite alternative.	Maintainer-authored squash retains the same Afloat co-author trailer.

Robotics Toolbox is already one of the 24 merged-PR repositories and is counted once in the 25-project total. Maintainer-authored PRs #883 and #805 are not added as Afloat16-authored merged PRs. Open work, including Denoising Diffusion #390 and REA #3940, is not counted as merged.