Status: living verification artifact for final.md section 27.
Last checked: 2026-07-07 from the macOS repository checkout for source-list shape, toolchain/source freshness, and newly added systems inputs. The last full URL reachability pass was 2026-07-03.
final.md uses the Source Matrix as its evidentiary backbone. This document
tracks whether those sources resolve, what each source should be used for, and
where reviewers should be careful not to over-claim. It supplements final.md
without replacing or editing it.
This file also carries temporal toolchain and accelerator freshness snapshots. Those snapshots are not repo pins. They are review inputs for agents who need to make current-version claims in performance, language-boundary, compiler, runtime, accelerator, or close-to-metal PRs.
Source-quality order for beaterOS design decisions:
- Primary papers, official specifications, and official vendor docs.
- Mature OS/security references and reproducible benchmarks.
- Company blogs for direction and product constraints.
- Speculation only when clearly marked as speculation.
Command shape:
sed -n '/^## 27\\. Source Matrix/,/^## 28\\. Final Strategic Recommendation/p' final.md |
rg -o 'https?://[^ )]+' |
while read url; do
curl -L --max-time 20 --connect-timeout 10 \
-A 'beaterOS-source-audit/0.1' -o /dev/null -s -w '%{http_code}' "$url"
doneCurrent result:
- 86 URLs are extracted from the current
final.mdsection 27. - The 2026-07-03 full reachability pass covered an earlier 51-URL source list. URLs added after that pass must not be described as fully reachability-audited until a new pass records them.
Result from the then-current source list:
- 51 URLs were extracted from
final.mdsection 27 at that revision. - 44 returned HTTP 200 from the command-line reachability pass.
- 7 returned HTTP 403 to
curlbut resolved in a browser-style fetch:https://wiki.osdev.org/Main_Pagehttps://wiki.osdev.org/Beginner_Mistakeshttps://wiki.osdev.org/Required_Knowledgehttps://openai.com/index/introducing-operator/https://openai.com/index/computer-using-agent/https://openai.com/index/chatgpt-agent-system-card/https://www.intel.com/content/www/us/en/developer/tools/trust-domain-extensions/overview.html
Interpretation: no final.md section 27 URL was found dead in the 2026-07-03
pass, but HTTP reachability is not the same as scholarly endorsement, benchmark
quality, or implementation correctness. New sources added after that pass must
be audited independently before their reachability is treated as checked.
Issue #16 flagged five structurally unusual 2026 arXiv IDs. Each resolved to the named paper on arXiv as of this audit.
| final.md entry | Verified source metadata | beaterOS use | Caveat |
|---|---|---|---|
Agent Operating Systems (AOS): Integrating Agentic Control Planes into, and Beyond, Traditional Operating Systems, 2606.01508 |
arXiv page exists; submitted 2026-06-01; title matches; subjects include cs.CR and cs.AI | Agent-control-plane framing, AOS responsibilities, OS abstraction gaps | Treat as a recent preprint, not settled architecture |
Qualixar OS: A Universal Operating System for AI Agent Orchestration, 2604.06392 |
arXiv page exists; submitted 2026-04-07; title matches; subjects include cs.AI, cs.MA, cs.SE | Multi-agent orchestration, routing, compatibility, dashboard/product surface | Application-layer orchestration, not a kernel trust model |
Toward Securing AI Agents Like Operating Systems, 2605.14932 |
arXiv page exists; submitted 2026-05-14; title matches; subject cs.CR | Security analogy between agents and OSs; privilege separation; resource mediation | Under-submission preprint; verify claims against beaterOS threat model |
CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents, 2601.09923 |
arXiv page exists; submitted 2026-01-14; revised 2026-06-04; title matches | CUA isolation, planning/execution separation, prompt-injection resistance | Security/utility tradeoffs need scenario tests before adoption |
OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks, 2606.29537 |
arXiv page exists; submitted 2026-06-28; title matches | Long-horizon computer-use benchmark pressure: hidden state, many tool calls, verification debt | Benchmark results must not become product requirements without local replication |
Conclusion for issue #16: the flagged IDs are valid and should stay in the Source Matrix. They should be marked as recent research inputs, not canonical requirements.
- Agent OS papers justify an agent control plane, but beaterOS should avoid replacing kernel mechanisms before it has a safe local runtime and sandbox lane.
- CUA and browser-use sources reinforce the need for manifest-before-execution, action-bound approvals, trusted sandbox resolution, and screenshots/DOM/a11y receipts.
- OSWorld-style benchmarks show that long-horizon agents fail on hidden state, verification, cross-source reasoning, and repeated tool calls; release gates need trace-property checks rather than final-answer-only scores.
- OS architecture sources reinforce minimal TCB, explicit interfaces, capability-style authority, and measured close-to-metal performance.
- Protocol and vendor sources are useful for compatibility direction, but authority must still be enforced outside model output and outside tool claims.
- Crypto, TEE, stablecoin, and post-quantum sources are input material for threat-model and schema agility; they are not reasons to invent cryptography.
These sources inform the metal-touching roadmap added to final.md §8.1.1 and
§8.1.2. They are design inputs, not implementation commitments.
| Source | beaterOS use | Caveat |
|---|---|---|
Linux sched_ext documentation |
Linux add-on experiments for policy-aware scheduling before native scheduler work | Linux-specific; not portable policy truth |
Linux io_uring and zero-copy receive documentation |
Ring-buffer IO, batching, completion records, and copy-budget discipline | API is Linux-specific and security-sensitive |
| eBPF/XDP documentation | Early packet filtering, tracing, observability, and enforcement hooks on Linux | Requires verifier limits and platform abstraction |
| Rust-for-Linux documentation | Memory-safe kernel-adjacent implementation path | Kernel Rust APIs remain evolving |
| DPDK poll-mode driver documentation | When user-space polling and direct descriptors beat interrupt-heavy paths | Burns CPU and weakens isolation if used casually |
| SPDK documentation | Zero-copy, asynchronous, lockless storage path inspiration | Specialized storage workloads only |
| Firecracker documentation | Minimal device model and microVM isolation for risky agent lanes | Depends on KVM/Linux host primitives |
| seL4 documentation | Capability microkernel and verification model for high-assurance future appliances | Broad hardware/device support is not free |
| CHERI architecture material | Hardware memory capabilities and compartmentalization direction | Research/availability constraints remain |
| Linux CXL memory-tiering background | Memory hierarchy planning for hot context, cold provenance, trace archives, and far-memory experiments | Linux- and hardware-dependent; not a near-term requirement |
These sources inform final.md accelerator-fabric language. They are inputs for
portable contracts, not reasons to bind beaterOS to one vendor SDK.
| Source | beaterOS use | Caveat |
|---|---|---|
| NVIDIA CUDA Programming Guide and CUDA Toolkit pages: https://docs.nvidia.com/cuda/cuda-programming-guide/index.html and https://developer.nvidia.com/cuda/toolkit | GPU execution model, streams, memory movement, launch overhead, kernel optimization vocabulary, and CUDA 13.x feature tracking | CUDA-specific; keep OS contract vendor-neutral |
| NVIDIA MIG User Guide: https://docs.nvidia.com/datacenter/tesla/mig-user-guide/latest/index.html | Hardware GPU partitioning and tenant isolation model for accelerator scheduling | Only supported on specific NVIDIA datacenter GPUs |
| AMD ROCm and HIP documentation: https://rocm.docs.amd.com/en/latest/index.html and https://rocm.docs.amd.com/projects/HIP/en/latest/index.html | Portable GPU backend input for AMD devices, HIP/CUDA portability, profiling, compatibility, and isolation review | Linux/device matrix matters; keep OS contract vendor-neutral |
| Intel oneAPI Level Zero specification: https://oneapi-src.github.io/level-zero-spec/ | Low-level direct-to-device runtime vocabulary for discovery, memory, command queues, and explicit accelerator control | Vendor/runtime-specific; SYCL/Level Zero behavior must be locally measured |
| OpenXLA, StableHLO, and JAX docs: https://openxla.org/ and https://openxla.org/stablehlo/spec and https://docs.jax.dev/ | Portable accelerator compiler shape, StableHLO as an ML compiler interchange layer, and CPU/GPU/TPU array execution vocabulary | Framework/compiler behavior changes; conformance must be local |
| Google Cloud TPU documentation and TPU architecture docs: https://docs.cloud.google.com/tpu/docs and https://docs.cloud.google.com/tpu/docs/system-architecture-tpu-vm | TPU as custom ASIC/pod resource for matrix-heavy ML workloads through VM/GKE/Vertex surfaces | Cloud/provider-specific; APIs and generations change |
| Groq LPU architecture documentation: https://groq.com/lpu-architecture | Deterministic token-generation silicon and low-jitter inference as a distinct accelerator class | Vendor-specific claims need benchmarked local validation |
| Apple Metal and Core AI documentation: https://developer.apple.com/metal/whats-new/ and https://developer.apple.com/videos/play/wwdc2026/324/ | Local GPU, neural-accelerator, tensor, and on-device model deployment direction for macOS and Apple Silicon | Public low-level access varies by framework and entitlement |
These sources support docs/optimization-agent-playbook.md and the language
freshness notes in docs/sota-systems-engineering.md. They are not repo pins;
they tell agents where to verify current facts before making toolchain claims.
Rows in this table carry their own Verified on dates; re-check the linked
official sources before treating any row as a current-version claim.
The following primary-source facts were re-checked for the metal OS blueprint and optimization guidance:
- Rust official release feed shows Rust 1.96.1 on 2026-06-30.
- LLVM project page shows LLVM 22.1.8 released on 2026-06-16.
- Swift official forum/install pages show Swift 6.3.3 on 2026-06-30.
- Python official downloads show Python 3.14.6 on 2026-06-10.
- NVIDIA CUDA Toolkit archive shows CUDA Toolkit 13.3.1 in June 2026.
- AMD ROCm/HIP documentation shows HIP 7.x material and current ROCm programming, profiling, compatibility, and isolation guidance.
- Intel oneAPI Level Zero documentation shows the current low-level direct device runtime specification.
- Apple Metal documentation shows Metal 4 and 2026 Apple Silicon ML/tensor acceleration direction.
- Google Cloud TPU docs show TPU7x/Ironwood as the current TPU architecture.
- Linux kernel documentation covers
sched_ext, and Linux/io_uring documentation covers ring-buffer and zero-copy-oriented IO models. - Linux kernel Rust documentation and Rust-for-Linux remain the primary source trail for Rust kernel-adjacent work.
These facts are inputs to review. They do not change the repo-pinned Rust baseline or create a committed CUDA, Metal, TPU, LPU, NPU, or kernel backend.
Use this table when a PR says "latest", "current", "new compiler", "new
runtime", "GPU optimized", "TPU optimized", "LPU optimized", "Apple Silicon
optimized", or similar. The Repo baseline column says what beaterOS actually
builds or tests against today; the Upstream version/status column is only a
dated source snapshot.
| Component | Upstream version/status | Repo baseline | Primary source | Source type | Source date | Verified on | Optimization relevance | Claim boundary |
|---|---|---|---|---|---|---|---|---|
| Rust | Rust 1.96.1 | rust-toolchain.toml pins 1.93.1; Cargo.toml declares rust-version = "1.93" |
https://blog.rust-lang.org/2026/06/30/Rust-1.96.1/ | Official release blog | 2026-06-30 | 2026-07-07 | Compiler, Cargo, stdlib, MIR/LLVM backend behavior, safety fixes | New Rust use is not automatic; benchmark and compatibility evidence required before changing repo baseline |
| LLVM | LLVM 22.1.8 | Indirect through Rust/Apple/vendor toolchains unless explicitly invoked | https://llvm.org/ | Official project release page | 2026-06-16 | 2026-07-07 | Backend, sanitizer, C/C++ interop, vectorization, target support | LLVM version alone does not prove Rust, Apple Clang, CUDA, or vendor compiler behavior |
| Zig | 0.16.0 release; 0.17.0-dev snapshots visible on download page | No beaterOS TCB baseline | https://ziglang.org/download/ | Official download page | 2026-07-05 snapshot observed | 2026-07-07 | Freestanding and cross-compilation probes | Experimental only until toolchain stability and reviewer depth are proven |
| Swift | Swift 6.3.3 | No authority-path baseline; Apple-native/platform integration only | https://forums.swift.org/t/announcing-swift-6-3-3/87888 | Official project forum announcement | 2026-06-30 | 2026-07-07 | Apple platform APIs, UI/platform integration, possible embedded/platform experiments | Swift is not the policy, journal, receipt, or scheduler authority boundary |
| Go | Go 1.26.4 artifacts on download page | No beaterOS authority baseline | https://go.dev/dl/ | Official download page | Dynamic release page | 2026-07-07 | Non-TCB infrastructure daemon/tooling checks | Not for policy, journals, receipts, or scheduler authority |
| Python | Python 3.14.6 | Host python3 for bounded scripts and local gates |
https://www.python.org/downloads/ | Official download page | 2026-06-10 | 2026-07-07 | Audit, validation, research, replay scripts | Python scripts must remain bounded and non-authoritative |
| CUDA Toolkit | CUDA Toolkit 13.3.1 shown in CUDA archive/download surfaces | No committed CUDA backend; Linux CUDA lane is experimental readiness metadata | https://developer.nvidia.com/cuda-toolkit-archive | Official vendor archive | 2026-06 | 2026-07-07 | GPU kernels, streams, launch overhead, occupancy, memory movement, profiler compatibility | CUDA is a backend behind beaterOS admission/telemetry/receipt contracts, not the OS contract |
| NVIDIA MIG | Latest MIG guide | No committed MIG backend; partitioning input for future GPU lanes | https://docs.nvidia.com/datacenter/tesla/mig-user-guide/latest/index.html | Official vendor documentation | Dynamic latest page | 2026-07-07 | Hardware GPU partitioning and tenant isolation | Only certain NVIDIA datacenter GPUs support it; fallback isolation required |
| AMD ROCm/HIP | ROCm/HIP latest documentation includes HIP 7.x material | No committed ROCm backend | https://rocm.docs.amd.com/en/latest/index.html and https://rocm.docs.amd.com/projects/HIP/en/latest/index.html | Official vendor documentation | Dynamic documentation | 2026-07-07 | AMD GPU backend vocabulary, HIP portability, profiler/isolation/compatibility review | ROCm/HIP is a backend behind beaterOS contracts, not the OS contract |
| Intel Level Zero | oneAPI Level Zero 1.17.24 specification | No committed Level Zero backend | https://oneapi-src.github.io/level-zero-spec/ | Official specification | Dynamic latest page | 2026-07-07 | Low-level direct-to-device runtime, command queues, memory, device discovery | Vendor/runtime-specific; require local telemetry and fallback |
| Apple Metal | Metal 4 direction on Apple Developer "What's new" page | Apple Silicon readiness lane is metadata only; no committed Metal backend | https://developer.apple.com/metal/whats-new/ | Official vendor documentation | Dynamic 2026 page | 2026-07-07 | Local GPU, tensors, quantization, MPS/Core ML adjacent placement, Apple Silicon profiling | Framework placement can be opaque; record limitations and CPU fallback |
| Cloud TPU | Cloud TPU docs and TPU7x/Ironwood architecture pages | No committed TPU backend | https://docs.cloud.google.com/tpu/docs | Official cloud/provider documentation | Dynamic documentation | 2026-07-07 | Matrix-heavy ML accelerator scheduling, pod/VM/GKE provider constraints | Provider-specific; admission, spend, telemetry, and receipts remain beaterOS-owned |
| Groq LPU | LPU architecture documentation | No committed LPU backend | https://groq.com/lpu-architecture | Vendor architecture documentation | Dynamic vendor page | 2026-07-07 | Deterministic low-jitter inference silicon as a distinct accelerator class | Vendor claims require measured validation and fallback |
| OpenXLA/StableHLO/JAX | Portable compiler/interchange and array execution docs | No committed XLA backend | https://openxla.org/ and https://openxla.org/stablehlo/spec and https://docs.jax.dev/ | Official project documentation | Dynamic documentation | 2026-07-07 | TPU/GPU compiler portability, graph lowering, backend placement vocabulary | Local conformance and backend telemetry required before claims |
Track announced language/compiler/backend changes that are not beaterOS pins but can invalidate optimization assumptions.
| Component | Effective window | Source | Verified on | Impact | Required response |
|---|---|---|---|---|---|
Rust nvptx64-nvidia-cuda target |
Announced 2026-05-01; scheduled for Rust 1.97 | https://blog.rust-lang.org/2026/05/01/nvptx-baseline-update/ | 2026-07-07 | Raises PTX ISA and GPU architecture baselines for Rust-generated NVIDIA GPU PTX, affecting older CUDA drivers and GPUs | Any Rust GPU backend PR must record target CPU/PTX ISA, CUDA driver floor, supported GPU floor, scalar/CPU fallback, and rollback plan before changing repo baseline |
| Source | beaterOS use | Caveat |
|---|---|---|
| Rust release blog, Rust 1.96.1, 2026-06-30: https://blog.rust-lang.org/2026/06/30/Rust-1.96.1/ | Current Rust release verification and Cargo/rustup update provenance | The workspace baseline is pinned separately in rust-toolchain.toml; use the newer release only when a PR explicitly justifies and measures the toolchain change |
| LLVM project home/release page, LLVM 22.1.8, 2026-06-16: https://llvm.org/ | Compiler/backend, sanitizer, C/C++/Rust backend, and toolchain-version checks | LLVM version alone does not prove a Rust, Apple Clang, or vendor compiler behavior |
| Zig download page, 0.16.0 release and 0.17.0-dev snapshots: https://ziglang.org/download/ | Freestanding/cross-compilation experiment tracking | Zig remains non-TCB until stability and reviewer coverage are proven |
| Swift.org macOS install page and Swift 6.3.3 announcement: https://swift.org/install/macos/ and https://forums.swift.org/t/announcing-swift-6-3-3/87888 | Apple-native platform integration and Swift build-tooling awareness | Swift is not the beaterOS authority boundary |
| Go downloads page, Go 1.26.4 artifacts: https://go.dev/dl/ | Non-TCB infrastructure daemon/tooling version checks | Go is not used for policy, journals, receipts, or scheduler authority paths |
| Python downloads page, Python 3.14.6, 2026-06-10: https://www.python.org/downloads/ | Audit/research script runtime freshness | Python scripts must remain bounded and non-authoritative |
| NVIDIA CUDA Programming Guide: https://docs.nvidia.com/cuda/cuda-programming-guide/index.html | GPU programming model, memory hierarchy, streams, launch/occupancy vocabulary | CUDA is a backend, not the OS contract |
| AMD ROCm/HIP documentation: https://rocm.docs.amd.com/en/latest/index.html and https://rocm.docs.amd.com/projects/HIP/en/latest/index.html | AMD GPU backend and HIP portability checks | ROCm/HIP claims need local compatibility, profiler, isolation, and fallback evidence |
| Intel Level Zero specification: https://oneapi-src.github.io/level-zero-spec/ | Low-level accelerator runtime and command-queue vocabulary | Level Zero claims need local device/runtime telemetry and fallback evidence |
These sources support final.md §14.9 and
docs/design/eval-statistical-method.md. They define how beaterOS release gates
avoid single-run point estimates for probabilistic agents.
| Source | beaterOS use | Caveat |
|---|---|---|
tau-bench, arXiv:2406.12045, submitted 2024-06-17: https://arxiv.org/abs/2406.12045 |
pass^k reliability metric for repeated agent trials; motivates reporting reliability across k consecutive successes | Benchmark domains are not beaterOS requirements; the metric is the reusable input |
Adding Error Bars to Evals, arXiv:2411.00640, submitted 2024-11-01: https://arxiv.org/abs/2411.00640 |
Error bars, paired model comparisons, and experiment-planning discipline for language-model evals | Preprint guidance; beaterOS still needs local gate calibration |
| Sequential Testing for Early Stopping of Online Experiments, SIGIR 2015: https://dl.acm.org/doi/10.1145/2766462.2767729 | Sequential stopping pattern so expensive multi-trial agent evals can stop once evidence is decisive | Online-experiment setting, not agent-specific; gate configs must declare stopping rules before runs |
When adding a source:
- Record title, URL, source type, date checked, and why it matters.
- Prefer official links over secondary commentary.
- For arXiv papers, verify that the ID resolves to the stated title and record submission or revision date.
- Mark preprints, marketing posts, and speculative sources explicitly.
- Do not cite a source as evidence for claims it does not make.
- Add or update a doc-health check when citations become part of a release gate.