Research window: 2026-11-09 04:30 ~ 2026-11-10 04:30 (Beijing time, past 24 hours; contiguous with the cutoff of the previous 11-09 report).

Sources: GitHub key repository and organization push verification (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler, with per-repository checks of commits, PRs, Releases and tags within the window, branch-level atom fallback, organization-level commits counted precisely by window), ruyi package manager Releases and open PRs, packages-index, openRuyi packaging repository SPECS and release specification repository, ruyiai-daily daily intelligence archive, RuyiSDK developer community forum, three-page snapshot of the RuyiAI official homepage, openRuyi distribution news page, Google News RSS targeted Chinese and English queries (via proxy), arXiv search.


This Issue’s Index

  • Today’s Highlight: triton-riscv ships its first official release tag, install path moves from source to wheel (11-09)
  • Today’s Highlight: buddy-mlir lowers an entire f32 YOLOv8 graph to VIR, byte-identical output on K3 (11-10)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 triton-riscv: after switching to torch 2.15.0, seven flaggems test cases fail; switched to defensive reads (11-09)
    • 1.2 triton-riscv: lit crashes under Python 3.14; CI changed to prefer pip-installed lit (11-09)
    • 1.3 pytorch, tilelang-riscv, llvm-project: zero commits in window (11-10 status)
  • II. RuyiSDK Toolchain and Distributions
    • 2.1 openRuyi: KDE Plasma 6.8.2 batch upgrade, forty desktop packages follow same day (11-09)
    • 2.2 openRuyi: seven Python ecosystem security updates, covering urllib3, tornado, pymongo (11-09)
    • 2.3 openRuyi: Wayland stack plus scattered updates to wget2, llvm bindings, etc. (11-09)
    • 2.4 openRuyi: new release repo created, bilingual template defines requirement IDs and priorities (11-09)
    • 2.5 ruyi: no new pre-release for 0.54, three fix PRs stalled since 11-04 (11-10 status)
  • III. Vector Extensions and Operator Optimization
    • 3.1 buddy-mlir: convolution vectorized along output channels, FHWC filters first converted to HWCF (11-10)
    • 3.2 buddy-mlir: adds rank-0 input support, last-dim-preserving reduce, and ops reading linalg.index (11-10)
    • 3.3 buddy-mlir: nightly stuck at 0.1.4.dev20261105, four commits not yet packaged (11-10 status)
    • 3.4 Vector benchmarks and the three translation repos remain silent (11-10 status)
  • IV. Buddy Compiler Updates
    • 4.1 Three points of review convergence: parallel row memory dependencies, FMA rounding, reduce dispatch (11-10)
    • 4.2 Four AME- and RAX-related PRs still under review (11-10 status)
  • V. Community, Events, and Official Homepage
    • 5.1 Official homepage three pages unchanged, two blog posts, twenty-three documentation chapters (11-10 status)
    • 5.2 No new forum posts in window, latest still the 11-06 repost (11-10 status)
    • 5.3 ruyiai-daily archive stops at the 11-08 issue (11-10 status)
    • 5.4 Zero hits in external news, arXiv latest stops at 11-05 (11-10 status)
  • VI. Trend Observations
    • 6.1 What this round fills in is the delivery pipeline, not operators
    • 6.2 Byte-identical output is becoming the acceptance criterion for VIR lowering
    • 6.3 Desktop stack and security updates take over the packaging cadence

Today’s Highlight: triton-riscv cuts its first formal release tag, install path shifts from source to wheel (11-09)

Date: 2026-11-09 Source: Release tag release/20261108-6895885, PR #105, PR #107

The release/20261108-6895885 tag cut at 06:59 Beijing time on 11-09 is the repo’s first formal release tag: all eighteen prior tags carried the nightly/ prefix, and the repo had never had a release/ tier. The release ships two wheels, one for x86_64 (manylinux_2_27 and 2_28) and one for riscv64 (manylinux_2_39), both cp312-abi3, with version string triton-3.9.0.

The tag points at commit 6895885, the merge commit for the 11-06 “add codegen library for AMD” change (#102), the same point as nightly/20261106-6895885. The day’s three commits (#105, #106, #107) all landed after the tag and are not in this package. That means the “first formal release” actually packages last weekend’s code state; the release action itself is the entire content of this tag.

#105, landed the same day as the tag, rewrites the README to make the install section wheel-first instead of source-first: it first gives the wheel path and the package index hosting it, then the LLVM archive they require and a variable pointing at it; the source path is compressed into make setup (create .venv, clone a pinned Triton, download pinned LLVM and the buddy build), and the previous instructions (“create a virtualenv, install dependencies, fifteen variables in the env script”) are reduced to a short note.

The quick start is also rewritten: instead of just printing ok: True, it compiles a kernel for riscv64-unknown-linux-gnu, keeps the object file and greps the disassembly, printing the vsetivli, vle32.v, vfadd.vv, vse32.v emitted by the backend. For a compiler project, that demonstrates “the backend really emits vector instructions” far better than a single assertion.

Nightly builds also gain a unique version suffix .dev: previously every build shared Triton's own version number, and the package index rejects duplicate versions. The wheel publish step now runs inside the wheel job to avoid re-downloading artifacts, and is gated on the presence of the PYPI-family secret, continuing without failure when the package index side is not yet ready.

#107 then corrects this: the package index does not yet carry triton and the buddy tool, so the address given in the README’s binary install would fail; it now points at the two projects’ respective GitHub releases, installs the wheel from disk, and states the source of the LLVM archive. The same commit also fixes a lit crash on Python 3.14.


Today’s Highlight: buddy-mlir lowers an entire f32 YOLOv8 graph to VIR, byte-identical output on K3 (11-10)

Date: 2026-11-09 (merged 01:07 Beijing time on 11-10) Source: PR #1041, merge commit 18c66ec

PR #1041 takes the f32 YOLOv8 graph from #1042 end-to-end through VIR: first the Linalg-to-VIR lowering, then the VIR-to-vector-dialect lowering. Previously this pass had no convolution pattern, and three other classes of ops in the graph failed outright — a single change adds both convolution and those three classes.

The scale: fourteen files, 1,646 lines added, 115 removed, thirteen commits. Author zhanghb97 opened it at 08:57 on 11-09 and merged it that evening.

Both convolution families are covered: 2D NHWC with HWCF and 2D NHWC with FHWC. Vectorization runs along the output channel F (the last dimension of the output and of the HWCF filter); for each (kh, kw, c) the input element is broadcast into one vir.fma per output pixel, so every vector access is contiguous and there is no cross-lane reduction. FHWC filters are first copied in parallel into an HWCF buffer; output rows go into scf.parallel with one vir.set_vl per row over the F lanes.

The three previously failing op classes are also handled. First, rank-0 inputs: a memref fed directly into linalg.generic triggers invalid reassociation in memref.expand_shape, so it is loaded once and then broadcast. Second, linalg.reduce preserving the last dimension, e.g. the 1x16x4x2100, dimensions=[1], addf-and-maximumf combination in DFL softmax: this is vectorized along the last dimension, with slices along the reduced dimension merged element-wise, generating no vir.reduce. Third, named ops reading linalg.index, e.g. the linalg.map iota in upsampling: these get the same scalar loop as the equivalent linalg.generic.

The on-device results are clear: on K3, the raw output of f32 YOLOv8 pose changes by at most 6e-5, and the gap to PyTorch remains within 6e-4 (bus) and 8e-4 (zidane), unchanged from before; the detect model’s detections are unchanged and speed does not regress; the f32 and IME dynamic library outputs are byte-identical to before the change.


I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V Porting)

1.1 triton-riscv: Seven flaggems test cases fail after switching to torch 2.15.0; fixed with defensive reads (11-09)

Date: 2026-11-09 Source: PR #106, commit cfdf4f7

The riscv64 CI installs the torch bundled in the index, and the arrival of the latest 2.15.0 broke seven test cases in the flaggems suite. After the fix, 3,826 tests pass and two are skipped.

Five of the median failures stem from name tensors: median.py implements torch’s named tensor support, but this 2.15.0 build lacks Tensor.names, so the null-check function throws AttributeError before the kernel even executes. The fix is to read that attribute defensively, treating builds without this API as “no names,” which matches the answer the existing check gives for “tensors whose names are all None.” The author also explains why this code cannot simply be deleted: upstream flaggems still calls names without protection, and named tensors did exist in earlier torch versions.

The two sinh fp16 cases were originally compared with a relative and absolute tolerance of 1e-4 each, which fp16 cannot meet (measured relative difference 6.6e-4); fp32 keeps 1e-4. At the same time, torch 2.10 was removed from the riscv64 installation, which now follows the index again. For a CI that tracks upstream torch, this kind of maintenance is a cost that cannot be outsourced: whenever upstream changes, the suite must be realigned.

1.2 triton-riscv: lit crashes under Python 3.14; CI now prefers the pip-installed lit (11-09)

Date: 2026-11-09 Source: PR #107, commit 209274f

The failing job was the Python 3.14 task in CI. Triton uses which to find lit on PATH, and the copy in the LLVM tree calls main at module level; Python 3.14 defaults to the forkserver start method on Linux, so re-importing crashes outright.

The fix is lightweight: put the build interpreter’s bin directory first on PATH so the lookup hits the pip-installed lit first. This is the same kind of cleanup as 1.1 after “CI follows upstream,” and explains why two of the three commits are environment fixes rather than compiler fixes.

1.3 pytorch, tilelang-riscv, llvm-project: zero commits in the window (11-10 status)

Date: 2026-11-10 (status) Source: pytorch riscv branch, tilelang-riscv, llvm-project

None of the three repos had commits in the window. The head commit of pytorch’s riscv branch is still the 11-04 sync batch; tilelang-riscv is stopped at 11-03 06:13 UTC; llvm-project’s main is stopped at 11-02 14:48 UTC. This issue’s RuyiAI stack increments all come from the three triton-riscv commits and one merge into buddy-mlir.


II. RuyiSDK Toolchain and Distributions

2.1 openRuyi: KDE Plasma 6.8.2 bulk upgrade, forty desktop packages follow on the same day (11-09)

Date: 2026-11-09 Source: plasma-workspace commit, kwin commit, commit list

The packaging repo landed its largest batch of the window on this day: a wholesale version bump to KDE Plasma 6.8.2, with forty packages following on the same day — from kwin, plasma-workspace, and plasma-desktop to breeze, discover, powerdevil, kinfocenter, and kscreenlocker, plus libraries such as libplasma, libksysguard, and libkscreen, and supporting components including layer-shell-qt, kwayland6, kpipewire, and knighttime.

For a RISC-V distribution, a full desktop-stack version bump is one of the scenarios most likely to expose architecture issues: the compositor path, hardware decoding integration, graphics resources, and the package management backend all have to be rebuilt and re-run on riscv64. Batches like this rarely come up all green on the first pass; the fixes over the following days are worth watching more than the bump itself.

Wayland-related components also moved forward in the same window: wayland to 1.26.0, wayland-protocols to 1.50, plasma-wayland-protocols to 1.23.0, and wlroots to 0.20.2. Upgrading protocols and compositor libraries in lockstep is a necessary step to keep the desktop stack from getting stuck on half-old protocols.

2.2 openRuyi: Seven Python ecosystem security updates covering urllib3, tornado, pymongo (11-09)

Date: 2026-11-09 Source: urllib3 to 2.8.0, python-truststore added, commit list

In a batch of commits at 12:07 Beijing time, seven were tagged as security fixes: urllib3 to 2.8.0 (three CVEs fixed), tornado to 6.6 (three CVEs and two GHSAs), pymongo to 4.19.0 (four CVEs), dnspython to 2.9.1 (one CVE), websockets to 17.4, trio to 0.35.0, and aioquic to 1.7.2 (six GHSAs).

What these share is the networking and async stack: HTTP clients, DNS, WebSocket, QUIC, and async runtimes. They are not desktop components, but they determine whether containers, CI, and cloud-native toolchains can run securely on riscv64 — a layer that is invisible day to day but affects the entire stack when something goes wrong.

The same batch also added python-truststore (a packaging approach that uses the system certificate store instead of bundled certificates), merged pymongo’s BSON module into the main package, and had pydantic-extra-types use it for integration testing. In addition, httpx2 and httpcore2 both moved to 2.13.0, idna to 3.20, langgraph-sdk to 0.4.11, and uv-dynamic-versioning to 0.14.1.

2.3 openRuyi: Wayland stack plus scattered updates to wget2, LLVM bindings, and more (11-09)

Date: 2026-11-09 Source: wlroots to 0.20.2, wget2 to 2.3.0, LLVM binding fix

Beyond the two main threads of desktop and security, there were a few scattered updates that evening: wlroots to 0.20.2, wget2 to 2.3.0, and openruyi-linux-build gained versioned documentation installation and a fix for devel file conflicts.

Worth calling out separately are the Python bindings for the llvm22, llvm23, and llvm24 packages: they now use text substitution to point clang’s bindings at a versioned libclang. In a distribution where multiple LLVM versions coexist, bindings that point at an unversioned name will overwrite each other once a second version is installed; changes like this look trivial but directly determine whether multiple toolchains can coexist.

2.4 openRuyi: New release repo defines requirement numbering and priority via bilingual templates (11-09)

Date: 2026-11-09 Source: release repo, commit list

openRuyi-Project created a release repo on 11-04, and today Jingwiw pushed eight commits in a row to codify the distribution’s requirement management rules as documentation: requirement numbering rules, writing principles, scope classification (compute and verification), issue priority, plus bilingual Chinese-English issue templates and per-item standard documents.

This repo holds no code, only rules. For a distribution that releases monthly, pinning down “what counts as a requirement, how it is numbered, and how priority is set” is a prerequisite for turning judgments scattered across issues and meetings into a reviewable process. Once the rules are in place, entries in the monthly release notes should be traceable to specific requirement numbers — and that is verifiable.

2.5 ruyi: 0.54 has no new pre-release; three fix PRs stalled since 11-04 (11-10 status)

Date: 2026-11-10 (status) Source: ruyi Release list, open PR list

The main ruyi repo saw zero commits and zero releases in the window; the latest pre-release is still 0.54.0-beta.20261102 from 11-02, and the latest stable is still 0.53.0 from 09-24. The three fix PRs opened on 11-04 (ignore relative XDG base directories, handle trailing empty tables in the manifest, and expose repo.disabled via config get) are unchanged in status.

The guidance from the biweekly progress report in issue 080 was that 0.54 would be released around November 11 — tomorrow is that observation point. If only a beta exists by then, it means the blocker for the stable release is not in code commits.


III. Vector Extensions and Operator Optimization

3.1 buddy-mlir: Convolution vectorized along output channels; FHWC filters first converted to HWCF (11-10)

Date: 2026-11-09 (merged 11-10 01:07 Beijing time) Source: PR #1041, merge commit 18c66ec

The convolution lowering strategy deserves separate attention, because it determines the room available for all subsequent tuning. The choice is to vectorize along the output channel F rather than reduce along the input channel C: for each (kh, kw, c), one input element is broadcast into one vir.fma per output pixel.

The direct benefit is that every vector memory access is contiguous, and there is no cross-lane reduction—reductions require horizontal instructions plus a subsequent reduction tree, which is costly at narrow VLEN. The cost is that the filter layout must match: filters in FHWC order must first be copied in parallel into an HWCF buffer, adding one extra transfer in exchange for contiguous access in the inner loop.

Output rows use scf.parallel, with one vir.set_vl per F lane per row. The previous VIR-to-vector lowering emitted a serial affine.for for the leading static dimension of dynamic vectors, and the parallelization pass could not see through vector.load; this is now changed to scf.parallel, provided the memory accesses permit it. In addition, the approach of one iteration per output pixel is retained in the Linalg-to-VIR lowering, and “multiple iterations sharing a single vector load” is left to the backend to merge within a vector setting that carries results, leaving register decisions to the target.

3.2 buddy-mlir: rank-0 inputs, last-dimension-preserving reduce, and operators reading linalg.index added (11-10)

Date: 2026-11-09 (merged 11-10 01:07 Beijing time) Source: PR #1041, merge commit 18c66ec

The three classes of failing operators each have their own pitfalls. rank-0 inputs (a single-element memref fed directly into linalg.generic) reported an invalid reassociation at shape expansion; this is now changed to load once and then broadcast.

Last-dimension-preserving reduce is more easily overlooked: DFL softmax’s 1x16x4x2100, dimensions=[1], with combiners addf and maximumf, previously made the pass error out directly. The new approach vectorizes along the last dimension and merges slices along the reduced dimension element by element, without generating vir.reduce; the forms that could previously be lowered remain unchanged.

The third class is named operators that read linalg.index, such as linalg.map iota in upsampling. Named operators have no generic op body to copy, so they were previously skipped; they now get the same scalar loop as the equivalent linalg.generic.

3.3 buddy-mlir: nightly stuck at 0.1.4.dev20261105, four commits not yet released (11-10 status)

Date: 2026-11-10 (status) Source: buddy-mlir Release list

The latest nightly is still 0.1.4.dev20261105 from 11-05. Since then, the 11-06 SPM merge, the two 11-08 VLEN changes, and the 11-09 VIR convolution merge have all gone unreleased.

Given the previous cadence (a release the day after new commits), this is now a consecutive slip. The nightly is a dependency of triton-riscv, and renovate automatically tracks buddy-mlir’s nightly version, so a stalled release means downstream cannot pick up these changes.

3.4 Vector benchmark and translation repos remain silent (11-10 status)

Date: 2026-11-10 (status) Source: VecIntrinBench, IntrinTrans, FBGEMM

The three repos had zero commits in the window, with their latest commits stuck at 02-13, 06-25, and 07-13 respectively. This issue’s vector increments remain concentrated on buddy-mlir’s VIR lowering, with no return to benchmarks and translation.


IV. Buddy Compiler Updates

4.1 Three Points Resolved in Review: Parallel-Row Memory Dependencies, FMA Rounding, reduce Dispatch (11-10)

Date: 2026-11-09 (merged 11-10 01:07 Beijing time) Source: PR #1041

All three review comments turned into code changes. Parallel-row memory dependencies: there are no SSA dependencies between rows, but there may be memory dependencies, so scf.parallel is used only when the rows’ memory accesses can be proven independent — every write is a vir.store and the stored vector carries these rows; the written elements do not overlap (unit layout, or all strides static and mutually nested after sorting by stride); and any access that may overlap the written memref must use the same base, indices, and rank. Otherwise it falls back to serial affine.for. The new test file covers four cases: parallelizable, scalar read/write shared across rows, memref with stride 0, and reading data another row is writing.

The FMA point is subtler: VIR-to-vector lowering lowers vir.fma to a vector fma in the main loop, but splits it into multiply plus add in the tail loop, so the same op gives different rounding on vector lanes versus tail lanes. The test uses a = 1 + 2^-23, b = 1 - 2^-23, c = -1: vector lanes get -2^-46, tail lanes get 0. The tail now uses math.fma, so both sides agree, and the test checks both. This change introduces rounding-level differences on tail lanes; the upper bound on f32 YOLOv8 output change on K3 is the 6e-5 mentioned earlier.

The reduce dispatch conditions were also narrowed: the predicate function now aligns exactly with what the old implementation actually supported (element type f32, combiners addf or maxnumf, dimensions per rank, etc.), so reduces that should not take the fast path are no longer swallowed by it.

Date: 2026-11-10 (status) Source: PR #1035, PR #1038, PR #1039, PR #1040

The four previously open PRs saw no merges and no updates during the window: descriptor offsets for AME tile addressing, Linalg-to-XTAME matmul-family lowering, the instruction-counting XtameBench, and a llama-bench-style RAX benchmark. Together these form the path from AME “running” to AME “correct and comparable,” but they have not kept pace with #1041.


V. Community, Events, and Official Site

5.1 No Changes to the Three Official Site Pages; Two Blog Posts, Twenty-Three Doc Chapters (11-10 status)

Date: 2026-11-10 (status) Source: RuyiAI official site

The index, blog, and docs page snapshots match the previous issue, with no new blog posts or doc chapters. The blog index still has two posts (Triton’s RISC-V port, PyTorch RISC-V CI practices), and the docs have twenty-three chapters.

5.2 No New Forum Posts in the Window; Latest Is Still the 11-06 Repost (11-10 status)

Date: 2026-11-10 (status) Source: Latest RuyiSDK community topics

No new topics in the window. The latest is still the 11-06 repost on Infineon’s RISC-V automotive MCU, and the replies to the 11-07 hwprobe baseline discussion have not continued. Biweekly progress issue 081 is not out yet: issue 080 was finalized on 11-05, so on the biweekly cadence the next one is due around 11-19.

5.3 ruyiai-daily Archive Stops at the 11-08 Issue (11-10 status)

Date: 2026-11-10 (status) Source: ruyiai-daily commit list

The last commit to the daily intelligence repo was 11-08 06:21 (Beijing time), corresponding to the 11-08 issue; neither the 11-09 nor the 11-10 issue has been generated. The repo previously produced one issue per day; this two-day gap looks like the pipeline stopped rather than a lack of content to collect.

5.4 Zero Hits in External News; arXiv Latest Stops at 11-05 (11-10 status)

Date: 2026-11-10 (status) Source: Google News, arXiv RISC-V search

The general RISC-V crawl returned only items like personal-learning CPU repos, ESP32 dev boards, and Infineon AURIX RC1 during the window, unrelated to the Ruyi community; the five targeted queries — RuyiAI, RuyiSDK, Ruyi, openRuyi, Buddy Compiler — returned zero hits. The latest paper in the arXiv RISC-V search is SensorGuard-RV from 11-05; no new papers in the window.


VI. Trend Observations

6.1 This round fills in the delivery pipeline, not operators

Of the three triton-riscv commits, two are installation and test pipeline changes (README switched to wheel-first, flaggems updated for the new torch), and one is CI Python 3.14 compatibility; none touch the compiler itself. Together with Sunday’s first official release tag, the theme of the day is “making it installable, and installable correctly.”

Rounds like this occur in any project, but are easily misread as a lack of progress. In fact, moving from source installs to wheel installs marks the boundary between “developers can build it” and “users can install it”: the source path requires users to prepare LLVM, the buddy build, and dozens of environment variables themselves, while the wheel path needs just one command. For an ecosystem like RISC-V that has long relied on “building it yourself,” the marginal value of this step exceeds optimizing another operator.

6.2 Byte-for-byte identity is becoming the acceptance criterion for VIR lowering

The two changes on 11-08 used load-time validation to block VLEN mismatches and made host sizing follow the real VLEN, and today’s #1041 adds “f32 and IME dynamic library output byte-for-byte identical to before the change” to the verification criteria. All three changes point to the same standard: lowering may swap implementations, but must not change numerical values.

This standard is stricter than “precision within tolerance” and better suited to compilers: tolerance can mask ordering errors, byte-for-byte comparison cannot. The cost is that every change needs a reproducible baseline, and a dual-core machine like K3 (with different VLENs across its two clusters) happens to provide the hardest baseline.

6.3 The desktop stack and security updates take over the packaging cadence

Of the sixty-five commits, forty are KDE Plasma 6.8.2, seven are Python security updates, and the rest are scattered version bumps such as Wayland, wget2, and llvm bindings. The AI stack had only four commits this day, while the distro side had over sixty.

This ratio shows that openRuyi’s current bottleneck is keeping up with upstream, not building new things: once upstream releases a desktop or language runtime version, the packaging repo must catch up entirely within a day or two; meanwhile the AI stack’s cadence revolves around models and toolchains, moving weekly or even per release.


Source Verification Table

Source Verification Result
RuyiAI-Stack key repos triton-riscv: three commits (#105 docs, #106 tests, #107 CI), one official release tag release/20261108-6895885 (two wheels); pytorch, tilelang-riscv, llvm-project: zero commits; remaining repos have no commits in the window
ruyisdk org ruyi, packages-index: zero commits, zero releases; ruyi’s three open PRs stalled since 11-04; wechat-articles: no commits in the window; remaining repos have no updates
openRuyi-Project Packaging repo: sixty-five commits (forty KDE Plasma 6.8.2 packages, seven Python security batch packages, Wayland and wget2, etc.); firmware repo: no new tags, no new commits; release spec repo: eight commits; news page: no additions, monthly edition still at 2026.09
buddy-compiler buddy-mlir: one merge (#1041, fourteen files, thirteen commits); nightly still at 0.1.4.dev20261105; four PRs still under review
RuyiSDK Forum No new topics in the window; latest is still the 11-06 repost; biweekly progress issue 081 not yet published
RuyiAI official homepage index, blog, and docs pages all unchanged; 2 blog posts, 23 documentation chapters
ruyiai-daily Archive stops at the 11-08 issue; 11-09 and 11-10 issues not generated
Site repo No new contributor index commits in this window
Google News Five targeted queries returned zero hits; general RISC-V crawls were all irrelevant items
arXiv No new papers in the window; latest stops at 11-05

Full Source List