Seventeen fixes, merged by the people who lead the projects.

I built and submitted a full lab for the textbook, “The Same-Bits Contract”, on why a model at temperature 0 does not give the same answer twice (pull request #2175). It follows the book’s lab protocol, its math is a new module of the book’s simulator, and it passes the book’s own test suites (2,658 simulator tests, 0 failures) and a full student run in the browser. On the proposal, a collaborator on the textbook: “Yes, the feature looks great and the idea sounds cool to me”, referring it to the book’s author (discussion).

Merged twelve of my fixes in 22 days; the other five, at the US Department of Energy national laboratories, the Massachusetts Institute of Technology and the Barcelona Supercomputing Center, are below. IBM Quantum: Tristan Müller, first author of the Relay-BP decoder, chose my fix for decodings that changed with batch order (1,970 of 2,000 shots → 0) and merged it (#49): “Thank you! Happy with this, will merge it.” Tencent’s ncnn, its on-device AI inference framework: its creator, nihui, extended the tests, approved and merged my fix for NaN outputs of the Mish activation on Vulkan GPUs: “Thanks for your contribution !” NVIDIA's Ising-Decoding maintainer, Ivan Basov, after several review rounds: “Merged, thank you for the contribution and for working through the review rounds.” Apple's MLX maintainer reviewed, approved and merged my fix for a data race in the distributed backend that links Macs for multi-machine training. Meta, on a memory-safety fix in the kernels on-device LLMs run in ExecuTorch: “LGTM! Thanks for the fix!”, then a second fix to the same kernels, merged by Meta's Jacob Stevens. Carnegie Mellon's XGrammar, the default structured-output engine of vLLM, SGLang and NVIDIA TensorRT-LLM, approved and merged my fix for a crash in its jump-forward decoding. Hugging Face, approving two fixes to the reward rubrics that grade agents in OpenEnv: “177 rubric tests pass.” PyZX's co-creator: “Thanks for the fix! This was a regression from an earlier PR.” JaxMARL's lead author, Alex Rutherford, on a fix to its RL rollouts: “Good find!!”

Energy and national security: Kokkos #9738, the GPU performance-portability library of the US Department of Energy national laboratories (Exascale Computing Project): sort now compiles on NVIDIA CUDA and AMD HIP for types without a default constructor. Merged by Damien Lebrun-Grandié, who leads the Kokkos project at Los Alamos National Laboratory, after approvals from him and Daniel Arndt, a Kokkos developer and computational scientist at Oak Ridge National Laboratory, and listed under Bug Fixes in the changelog for the upcoming Kokkos 5.4 release; reviewed by Andrey Prokopenko of Oak Ridge National Laboratory, who proposed my next Kokkos fix (5 November 2026). Weather and climate: ecCodes #684, the weather-data library of the European Centre for Medium-Range Weather Forecasts (intergovernmental, 35 states, headquartered in Reading, UK): reversed-scan Gaussian grids put about 6,000 points at wrong coordinates, now none; PSyclone #3668, the code generator the UK Met Office and the Science and Technology Facilities Council’s Hartree Centre build for the Met Office’s next-generation LFRic model: Fortran BIND(C) is no longer dropped; its maintainer: “otherwise it looks good”. National supercomputing: ReFrame #3750, the testing framework of the Swiss National Supercomputing Centre (ETH Zurich): a regression test on its 4.10 milestone. Particle physics: Rucio #8862, the data management system CERN (the European Organization for Nuclear Research) built for the Large Hadron Collider experiments: eleven commands, upload and download among them, reported success when they had failed; now they return the real error code. Academic research: MEEP #3364, the Massachusetts Institute of Technology’s electromagnetics simulator: merged by Ardavan Oskooi, MEEP’s original lead developer, on an issue where Prof. Steven G. Johnson asked for a pull request; #3363, on an issue Prof. Johnson filed, merged by Oskooi on 6 November: “Option 1 of #3170 (raising a ValueError) is the correct approach because in Python, failing fast and explicitly is strongly preferred over silently mutating inputs.” National supercomputing: Autosubmit #3405 and #3406, the climate-model workflow manager of the Barcelona Supercomputing Center, Spain’s national supercomputing centre: on PostgreSQL, job timestamps overflowed 32-bit columns and runs crashed; its maintainer Bruno P. Kinoshita invited the fix for the 4.2.0 release (“as we need it for the 4.2.0 release”), merged it with a thank-you in the 4.2.0 changelog, and merged a second fix he asked for (6 November 2026). Oceans: emsarray #233, the ocean and coastal model data library of the Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia’s national science agency: forecast datasets got the wrong time coordinate; in review. Each tested before and after with the project’s own tests, in public runs.

I sent CARA my independent reproduction of its Monte Carlo collision probabilities on all 53 of its test conjunctions. Its operations team lead referred it to CARA's analysis team: it is now under review by NASA CARA, which will report its findings.

Scouted by a Talent Investor from my public work alone, citing tessel's GEMM at cuBLAS parity, switchyard's mixture-of-experts serving, and a negative result I chose to publish (a compression pipeline that aimed for 8× smaller and measured −6.3%).

kessler screens every pair of the 16,768 active satellites, about 60,000 close approaches a day; the full 32,593-object catalogue, all 531 million pairs, takes 28 seconds on one T4 GPU.

quorumdb ran 10,000 simulated clusters through 610,064 crashes and 144,354 network partitions with zero safety violations; faultline ran 1.26 million operations under 288 injected faults against four etcd releases.

C++, CUDA, Rust, Python and Swift, with kernels for NVIDIA GPUs, Apple Metal and Google TPUs; studies of 24 languages on 15 production tokenizers and 33,869 scams in 49 languages; Populace has 1,600+ npm downloads.

On Vulkan GPUs, Tencent’s on-device inference framework computed Mish as x·tanh(log(1+e^x)), and e^x overflowed for large inputs, so drivers with a strict tanh returned NaN (30 NaN outputs in the reproduction, none after). Its creator, nihui, added a commit extending the tests, approved and merged it (5 November 2026, India time): “Thanks for your contribution !”

Relay-BP, the quantum error-correction decoder from IBM Quantum (© IBM), decoded the same shot differently depending on batch order: serial and parallel batches disagreed on 1,970 of 2,000 shots. Its first author, Tristan Müller, chose this fix and merged it (5 November 2026); now 0.

MEEP’s remaining-time estimate fell to 0.25× the real time; now 0.97–1.02×. Merged by Ardavan Oskooi, MEEP’s original lead developer. #3363 makes a silent setting override raise an error instead; Oskooi merged it as “the correct approach”.

On PostgreSQL, Autosubmit stored job timestamps such as 20261027164231 in 32-bit columns, so runs crashed with NumericValueOutOfRange; they are now 64-bit, checked on PostgreSQL 16. Invited and merged for the 4.2.0 release by maintainer Bruno P. Kinoshita, with a thank-you in the changelog; #3406, at his request, makes a test that failed on busy CI runners wait for its process to exit.

The socket threads of MLX's distributed ring backend read their task queues outside the lock while other threads appended to them. ThreadSanitizer found 2–4 races on every run before the fix and none after; the ring tests pass on 8 ranks.

The 2-bit and 4-bit embedding kernels used by on-device LLMs read out of bounds on an invalid token id; added the missing check, with tests that fail without it.

XGrammar, the default structured-output engine of vLLM, SGLang and NVIDIA TensorRT-LLM, crashed in jump-forward decoding whenever a grammar forced only part of a UTF-8 character (for example, a choice between “é” and “è”): the bug behind its issue #132, closed earlier with the feature switched off. It now returns only whole, valid characters, which the grammar still forces, so results stay correct; four new tests fail without the fix. Approved and merged by maintainer Linzhang Li (29 October 2026).

In the reward rubrics RL environments use to grade agents, composite rubrics ran their pre-hooks after scoring their children (a timing hook read 0.01 ms for 400 ms of work), and a rubric handing off to an async judge recorded an unawaited coroutine as its score. 9 of 13 new tests fail without the fixes.

jax.grad through lax.reduce_window with a custom reduction has failed since 2021 (#7815). Its derivative is now a fold over the window offsets, built only from transposable primitives in the order XLA evaluates the reduction; the 242 new test cases fail without it. Google assigned it for review to Matthew Johnson, one of JAX’s original authors (21 October 2026).

The public-model tests, offline smoke test and docs looked only for .pt checkpoints, so anyone using the SafeTensors models NVIDIA publishes on Hugging Face failed them; they now load the published files out of the box. Merged by maintainer Ivan Basov (30 October 2026): “thank you for the contribution and for working through the review rounds.” #99, split out at his request, makes the documented checkpoint setting work: “Merged, thank you for the follow-up and for adding the precedence test” (2 November 2026).

The same kernels checked zero points against the output type instead of the scales' type: valid models with fp16 scales and fp32 output aborted, and the .out variants read past the end of their scales under AddressSanitizer. Six new tests fail without the fix. Revised after maintainer review so the export kernels also report the scales' dtype; approved and merged by Meta's Jacob Stevens (28 October 2026).

The FlashInfer prefill tests compared non-causal attention with a causal reference and passed only because of a loose tolerance; fixed so they catch the bug.

DeepSeek-V3 tool calls streamed with empty arguments whenever multi-token prediction delivered the closing fence and end tokens in one chunk. Found with a differential check of every tool-call parser, streaming against one-shot; two new tests fail without the fix and all 263 parser tests pass with it.

DaCe, the data-centric compiler from ETH Zurich's Scalable Parallel Computing Lab, compiled numpy.rint to C's round, which rounds halfway values away from zero, so compiled programs returned 3 for 2.5 where NumPy returns 2; fixed for float and complex types, with a regression test that fails without the fix. Also reported that max/min reductions drop NaN (issue #2730). Its maintainer (28 October 2026): “Seems all good, once we have the CI calmer, I will let these workflows run, once green can be merged.”

PyZX, the ZX-calculus library for quantum circuit optimisation, put an X rotation on the wrong qubit after rewriting two CNOTs as a SWAP, so its optimisers returned a different unitary for 25 of 100 random Clifford circuits; found by checking optimised circuits against the originals, fixed, with a regression test. Merged by PyZX co-creator John van de Wetering: “Thanks for the fix! This was a regression from an earlier PR that changed how XPhase gates were handled in optimization.”

The rollout wrapper behind 8 of JaxMARL's Q-learning baselines (QMIX, VDN, IQL, PQN-VDN, SHAQ, TransfQMIX) replaced the first observation of every new episode with zeros after an auto-reset, so agents chose their first action blind and trained on it. Two of three new tests fail without the fix. Approved and merged by JaxMARL's lead author, Alex Rutherford (26 October 2026): “Good find!!”

Composing a single 4-bit Nunchaku LoRA ignored its strength: 0.5, or 0 to switch it off, came back at full strength with no error. Only strength 1 now skips scaling; three of four new tests fail without the fix.

Deterministic mode fills torch.empty with NaN in eager PyTorch but not under the Inductor compiler; fixed without filling Inductor's own buffers, which is why an earlier fix was reverted.

Re-enabled 21 torch.compile tests skipped for bugs long fixed, and documented why 11 more must stay skipped, each checked on CPU and T4 GPUs.

A vocabulary extension for 24 Indian languages that keeps every other token (cs.CL). Read the paper (PDF).

Finding and removing a silent training–inference mismatch in a from-scratch GRPO stack (cs.LG). Read the report (PDF).

An open two-body Monte Carlo that agrees with NASA's on all 53 of its test conjunctions, 15–19× faster per core (physics.space-ph).

The multilingual scam-protection gap, measured on 33,869 reported scams in 49 languages.

On Hugging Face, for 24 Indian languages; loads with one line of transformers.

Simulated users that drive an app through its real API; on npm with no runtime dependencies.