Graphsignal
graphsignal/graphsignal
Profile AI inference workloads (vLLM, SGLang, TensorRT-LLM, PyTorch, any GPU application) with the Graphsignal profiler and read the results from its local /signals JSON endpoint.
Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs.
Install Compute Sanitizer
$ npx skills add facebookexperimental/triton --skill compute-sanitizer -a claude-codeProject install by default; add -g for ~/.claude/skills/.
$ gh skill install facebookexperimental/triton compute-sanitizer --agent claude-codeProject scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.claude/skills/compute-sanitizer .claude/skills/compute-sanitizer && rm -rf skills-srcUse ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.
Claude Code skills documentation · loads skills from .claude/skills/
Install the "compute-sanitizer" agent skill from https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizer into .claude/skills/compute-sanitizer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "compute-sanitizer", then confirm the skill loads.Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$skill-installer install https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizerType this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
$ npx skills add facebookexperimental/triton --skill compute-sanitizer -a codexProject install goes to .agents/skills/; add -g for ~/.codex/skills/.
$ gh skill install facebookexperimental/triton compute-sanitizer --agent codexProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .agents/skills && cp -r skills-src/.claude/skills/compute-sanitizer .agents/skills/compute-sanitizer && rm -rf skills-srcUse ~/.agents/skills/ instead of .agents/skills for a personal install.
Codex skills documentation · loads skills from .agents/skills/
Install the "compute-sanitizer" agent skill from https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizer into .agents/skills/compute-sanitizer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "compute-sanitizer", then confirm the skill loads.Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add facebookexperimental/triton --skill compute-sanitizer -a cursorProject install goes to .agents/skills/; add -g for ~/.cursor/skills/.
$ gh skill install facebookexperimental/triton compute-sanitizer --agent cursorProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .cursor/skills && cp -r skills-src/.claude/skills/compute-sanitizer .cursor/skills/compute-sanitizer && rm -rf skills-srcUse ~/.cursor/skills/ instead of .cursor/skills for a personal install.
Cursor skills documentation · loads skills from .cursor/skills/, .agents/skills/, .claude/skills/, .codex/skills/
Install the "compute-sanitizer" agent skill from https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizer into .cursor/skills/compute-sanitizer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "compute-sanitizer", then confirm the skill loads.Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gemini skills install https://github.com/facebookexperimental/triton.git --path .claude/skills/compute-sanitizer--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
$ npx skills add facebookexperimental/triton --skill compute-sanitizer -a gemini-cliProject install goes to .agents/skills/; add -g for ~/.gemini/skills/.
$ gh skill install facebookexperimental/triton compute-sanitizer --agent gemini-cliProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .gemini/skills && cp -r skills-src/.claude/skills/compute-sanitizer .gemini/skills/compute-sanitizer && rm -rf skills-srcUse ~/.gemini/skills/ instead of .gemini/skills for a personal install, then run /skills reload.
Gemini CLI skills documentation · loads skills from .gemini/skills/, .agents/skills/
Install the "compute-sanitizer" agent skill from https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizer into .gemini/skills/compute-sanitizer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "compute-sanitizer", then confirm the skill loads.Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gh skill install facebookexperimental/triton compute-sanitizerInstalls for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
$ npx skills add facebookexperimental/triton --skill compute-sanitizer -a github-copilotProject install goes to .agents/skills/; add -g for ~/.copilot/skills/.
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .github/skills && cp -r skills-src/.claude/skills/compute-sanitizer .github/skills/compute-sanitizer && rm -rf skills-srcUse ~/.copilot/skills/ instead of .github/skills for a personal install. Commit .github/skills so cloud agent and code review can use it.
GitHub Copilot skills documentation · loads skills from .github/skills/, .claude/skills/, .agents/skills/
Install the "compute-sanitizer" agent skill from https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizer into .github/skills/compute-sanitizer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "compute-sanitizer", then confirm the skill loads.GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add facebookexperimental/triton --skill compute-sanitizer -a opencodeOpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
$ gh skill install facebookexperimental/triton compute-sanitizer --agent opencodeProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .opencode/skills && cp -r skills-src/.claude/skills/compute-sanitizer .opencode/skills/compute-sanitizer && rm -rf skills-srcUse ~/.config/opencode/skills/ instead of .opencode/skills for a personal install.
OpenCode skills documentation · loads skills from .opencode/skills/, .claude/skills/, .agents/skills/
Install the "compute-sanitizer" agent skill from https://github.com/facebookexperimental/triton/tree/main/.claude/skills/compute-sanitizer into .opencode/skills/compute-sanitizer/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "compute-sanitizer", then confirm the skill loads.OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Facts
- Skill name
compute-sanitizer- GitHub stars
- 201
- Token cost
- ~1.6k tokens
- SKILL.md length
- 683 words
- Files
- 1
- Skills in repo
- 18
- Repo updated
- First seen
- Licence
- MIT
At a glance
Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs.
- Works in 4 steps: Locate the binary. It is often not on… → Get the reproduce command (a… → Pin a healthy GPU. Run bash… → …
- A kernel produces wrong results
- SKILL.md covers The four tools, Setup, Invocation and Triton/TLX specifics, plus 2 more sections
- Calls python and bash
What it does
Compute Sanitizer is an agent skill from facebookexperimental/triton, published by the product's own GitHub organization. Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs. Use when a kernel produces wrong results, crashes with an illegal/misaligned access, or is suspected of a shared-memory data race or invalid barrier usage — especially warp-specialized (WS) kernels using mbarriers, named barriers, TMA copies, or MMA accumulators. This is a runtime check: it runs the real kernel via its reproduce command, so it needs a working GPU…
Its SKILL.md is about 1.6k tokens, which your agent loads only when the skill is triggered. It is a single SKILL.md file with no bundled scripts.
It sits in AI & LLM Engineering. It works with NVIDIA AI Platform and CUDA. The repository describes itself as: Github mirror of trition-lang/triton repo. The licence is MIT.
When your agent uses it
- A kernel produces wrong results
- Crashes with an illegal/misaligned access
- Is suspected of a shared-memory data race
- Invalid barrier usage — especially warp-specialized (WS) kernels using mbarriers
Example prompts
- “/compute-sanitizer”
Requirements
- Python 3
Workflow steps
4 steps, taken from the first numbered list in SKILL.md.
- Locate the binary. It is often not on PATH. Prefer
- Get the reproduce command (a pytest/python invocation). Use the
- Pin a healthy GPU. Run bash third_party/tlx/find_working_gpu.sh, take a
- Enable source mapping. Triton emits line info by default; keep it on
What it can do on your machine
Read from SKILL.md and the folder at commit 37301d4. It shows what the files ask for, not the result of running them.
Tool permissions
Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.
From allowed-tools in the SKILL.md frontmatter.
Runs code
Shell commands in SKILL.md call:
pythonbash
From the folder's file list and the shell code blocks in SKILL.md.
Network
No URLs in SKILL.md.
From URLs in SKILL.md, links to its own repository left out.
Credentials
Names no API keys, tokens, secrets or passwords.
From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.
Context cost
Compute Sanitizer loads about 1.6k tokens when it runs. Until then it costs about 144 tokens; SKILL.md has 683 words of instructions outside code blocks.
Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.
Safety
Auto-check passed
The automated check found no risky patterns in SKILL.md.
Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.
SKILL.md
The full file from facebookexperimental/triton at commit 37301d4, republished under its MIT licence (© facebookexperimental). 683 words, ~1,639 tokens.
Download SKILL.mdSave it as
.claude/skills/compute-sanitizer/SKILL.md (or your agent's skills folder).- name
- compute-sanitizer
- description
- Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs. Use when a kernel produces wrong results, crashes with an illegal/misaligned access, or is suspected of a shared-memory data race or invalid barrier usage — especially warp-specialized (WS) kernels using mbarriers, named barriers, TMA copies, or MMA accumulators. This is a runtime check: it runs the real kernel via its reproduce command, so it needs a working GPU and is 10-100x slower than a normal run.
Compute Sanitizer
compute-sanitizer is NVIDIA's runtime correctness checker. It instruments the
real kernel launch and reports memory and synchronization errors that static IR
analysis cannot see. Use it to confirm (or rule out) out-of-bounds accesses,
uninitialized reads, shared-memory data races, and illegal barrier usage in
Triton/TLX kernels.
It complements the static investigations: barrier-visualization reasons about
the intended barrier protocol from the IR, while compute-sanitizer observes
what actually happens at runtime. When a race or sync error fires, cross-check
the two.
The four tools
Select with --tool; memcheck is the default.
- memcheck — out-of-bounds and misaligned global/local/shared accesses, plus device-side allocation/leak errors. First tool to run; catches the classic TMA-past-end and bad-index bugs.
- racecheck — hazards on shared memory (RAW / WAR / WAW). Highly relevant to WS producer/consumer kernels that stage data through SMEM buffers. Does NOT check global-memory races.
- synccheck — illegal use of barriers and synchronization primitives
(divergent
bar.sync, mismatched arrive/wait counts, invalid cluster/async barrier usage). Directly relevant to mbarrier / named-barrier WS code. - initcheck — reads of uninitialized global memory. Useful when output has nondeterministic garbage rather than a hard crash.
Setup
- Locate the binary. It is often not on PATH. Prefer
$COMPUTE_SANITIZER_BINif set, elsecommand -v compute-sanitizer, else/usr/local/cuda/bin/compute-sanitizer, else the newest/usr/local/cuda-*/bin/compute-sanitizer. - Get the reproduce command (a
pytest/pythoninvocation). Use the smallest failing shape/config — sanitizer overhead makes large grids impractical. Prefer a single test id over a whole file. - Pin a healthy GPU. Run
bash third_party/tlx/find_working_gpu.sh, take aWORKING_GPUSindex, and prependCUDA_VISIBLE_DEVICES=<idx>. See thedebug-failing-gpuskill. If a run hangs past your timeout, runthird_party/tlx/killgpu.sh. - Enable source mapping. Triton emits line info by default; keep it on
(do not set
TRITON_DISABLE_LINE_INFO=1) so errors map tofile:line. Pass--show-backtrace yesfor host+device backtraces.
Invocation
General form (the sanitizer wraps the whole command, env vars go before it):
CUDA_VISIBLE_DEVICES=<idx> \
<compute-sanitizer> --tool memcheck \
--error-exitcode 1 \
--show-backtrace yes \
--log-file <out_dir>/memcheck.log \
--target-processes all \
python -m pytest -s -x "<test_id>"Useful flags:
--error-exitcode 1— return non-zero when any error is found (scripting).--log-file <path>— capture the report (%pexpands to PID if needed).--target-processes all— follow child processes (pytest workers, subprocs).--kernel-name-exclude/--kernel-name <regex>— filter to the Triton kernel (names look like_attn_fwd_...,matmul_kernel, etc.) to cut noise.--launch-timeout <s>— bound a single launch.- memcheck:
--leak-check full,--padding <bytes>(catch off-by-a-few OOB). - racecheck:
--racecheck-report all(hazards + analysis). --print-limit 0— do not truncate the error list while triaging.
Run the tools in order — memcheck → racecheck → synccheck → initcheck — each into its own log. Stop early only with a stated reason (e.g. memcheck already found the crashing OOB, or an earlier tool hung).
Show full SKILL.md (273 more words)Show less
Triton/TLX specifics
- First run JIT-compiles the kernel (slow, unrelated to sanitizer). Let the compile complete; do not mistake compile time for a hang.
- racecheck is shared-memory only. WS kernels stage operands through SMEM, so a missing producer/consumer barrier shows up here as a WAR/RAW hazard on the SMEM buffer. A global-memory race will NOT be reported.
- synccheck + mbarriers. Invalid named-barrier / mbarrier / cluster-barrier
usage (e.g. wrong arrive count, divergent participation) surfaces here. Pair
with
barrier-visualizationto map the offending barrier to a partition. - TMA / async copies. Out-of-bounds TMA tile accesses crash with an illegal
instruction at runtime rather than masking — memcheck localizes the launch;
also see the
tma-illegal-instructionskill for the structural launcher-bug pattern. - Possible noise. Some async/TMA/
cp.asyncpaths may emit benign warnings or unsupported-feature notes on certain driver/toolkit versions. Note them as caveats; do not treat a warning as a confirmed bug without corroboration.
Interpreting output
- Each tool ends with
========= ERROR SUMMARY: N errors.0 errors= clean for that tool. racecheck printsRACECHECK SUMMARY. - A real finding includes the access kind (e.g. Invalid global read of size
16), the address, the kernel name, and — with line info — the
file:line. Capture the first error; later ones are often cascades. - "Clean" is a result, not a failure. Report it plainly: a clean memcheck + racecheck + synccheck materially narrows a wrong-result bug toward compiler/logic rather than memory/sync.
Reporting
Return a compact matrix plus triage:
tool | exit | errors | first finding (kernel @ file:line) | mapped?
memcheck | 0 | 0 | - | n/a
racecheck | 1 | 3 | WAR hazard on smem buf @ k.py:142 | yes
synccheck | 0 | 0 | - | n/a
initcheck | 0 | 0 | - | n/aThen state the most likely root cause (OOB vs. shared-memory race vs.
uninitialized read vs. illegal sync), the implicated TLX/Triton construct, and
the next action (e.g. narrow with a kernel filter, or hand a race/sync hit to
barrier-visualization).
Do not run performance benchmarks.
© facebookexperimental, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file
Files
Just SKILL.md in .claude/skills/compute-sanitizer of facebookexperimental/triton.
Open the folder on GitHubat commit 37301d4
Compare with similar skills
Compute Sanitizer next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.
| Skill | Stars | Used in | Tokens | Auto-check | Licence | Repo updated |
|---|---|---|---|---|---|---|
| Compute Sanitizer this skillfacebookexperimental/triton | 201 | — | ~1.6k | Automated safety check: Pass | MIT | |
| Graphsignalgraphsignal/graphsignal | 257 | — | ~6.3k | Automated safety check: Pass | Apache-2.0 | |
| LLM Torch Profiler Trace AnalysisBBuf/AI-Infra-Auto-Driven-SKILLS | 938 | — | ~2.8k | Automated safety check: Pass | None | |
| Optimize OpCVCUDA/CV-CUDA | 2.7k | — | ~834 | Automated safety check: Pass | Custom licence | |
| Cutlass SkillslowlyC/agent-gpu-skills | 169 | — | ~1.3k | Automated safety check: Pass | MIT | |
| Setup Workshop Nemoclawbrevdev/workshop-build-an-agent | 146 | — | ~5.2k | Automated safety check: Pass | Apache-2.0 |
Similar skills
LLM Torch Profiler Trace Analysis
BBuf/AI-Infra-Auto-Driven-SKILLS
Analyzes Torch Profiler traces from SGLang, vLLM and TensorRT-LLM servers into kernel attribution, overlap and fusion tables.
Optimize Op
CVCUDA/CV-CUDA
Drive a single-operator optimization campaign per .agents/guidance/OPTIMIZATIONGUIDELINES.md, with a deterministically enforced definition-of-done and versioned MR summary.
Cutlass Skill
slowlyC/agent-gpu-skills
Write, debug, and optimize CUTLASS, CuTe, and CuTeDSL GPU kernels from local upstream source, examples, and headers.
Setup Workshop Nemoclaw
brevdev/workshop-build-an-agent
Set up the NVIDIA "Build an Agent" DevX workshop as a working JupyterLab environment from INSIDE a locked-down OpenShell/NemoClaw sandbox, and hand the user the token URL + access commands.
Cv Deploy
LMIXR/CV_Deployment_skill
基于 helpfile 工程经验，协助 agent 配置 CV 主机和边缘设备环境、编译视觉与推理依赖、接入摄像头视频并打包部署服务。适用于 Ubuntu、CentOS、Windows、macOS、Jetson、树莓派和 RK3399 的 CV 工程实施与故障排查，以及相关移动端配套工具；模型训练和纯算法设计不属于本技能主线。
More from facebookexperimental/triton
All 18 skills in this repo- Official
Amd Att Trace
facebookexperimental/triton
Collect, validate, package, and inspect rocprofv3 Advanced Thread Trace bundles for AMD GPU kernels.
- Official
Ir Override Ablation
facebookexperimental/triton
Design and run Triton TTGIR debugging ablations using iroverride.
- Official
Tlx Kernel Optimization Agent
facebookexperimental/triton
Execute the TLX Kernel Optimization Agent CLI on a Triton or TLX kernel.
- Official
Debug Failing GPU
facebookexperimental/triton
Recover from GPU-busy / GPU-unavailable failures. An agent skill from facebookexperimental/triton.
- Official
Ir Debugging
facebookexperimental/triton
Debug Triton compilation by dumping IR at each stage (TTIR, TTGIR, LLVM, PTX).
- Official
Kernel Perf Testing
facebookexperimental/triton
Run TLX kernel performance benchmarks on Hopper, Blackwell, and AMD (gfx950/CDNA4, gfx1250) GPUs.
Related
Works with
Categories
Questions about Compute Sanitizer
What does Compute Sanitizer do?
Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs. Compute Sanitizer is an agent skill from facebookexperimental/triton, published by the product's own GitHub organization. Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs.
When should I use Compute Sanitizer?
Compute Sanitizer fits situations like: A kernel produces wrong results; crashes with an illegal/misaligned access; is suspected of a shared-memory data race; invalid barrier usage — especially warp-specialized (WS) kernels using mbarriers.
How do I install Compute Sanitizer in Claude Code?
Run `npx skills add facebookexperimental/triton --skill compute-sanitizer -a claude-code`. Or copy the skill folder (.claude/skills/compute-sanitizer in facebookexperimental/triton) into .claude/skills/compute-sanitizer in your project. Claude Code loads it when a task matches its description.
How do I install Compute Sanitizer in Codex?
Run `npx skills add facebookexperimental/triton --skill compute-sanitizer -a codex`. Or copy the skill folder (.claude/skills/compute-sanitizer in facebookexperimental/triton) into .agents/skills/compute-sanitizer in your project. Codex loads it when a task matches its description.
Can I use Compute Sanitizer in Cursor, Gemini CLI or GitHub Copilot?
Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add facebookexperimental/triton --skill compute-sanitizer -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/compute-sanitizer, .gemini/skills/compute-sanitizer, .github/skills/compute-sanitizer and .opencode/skills/compute-sanitizer in your project.
What does Compute Sanitizer need to run?
Going by SKILL.md and its folder, Compute Sanitizer needs the command-line tools its instructions call (python and bash). Our summary lists: Python 3.
Does Compute Sanitizer access the network?
SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.
Is Compute Sanitizer safe to install?
Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.
What licence does Compute Sanitizer use?
Compute Sanitizer is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.
How many tokens does Compute Sanitizer use?
About 1.6k tokens (SKILL.md is roughly 6.6k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.
What are the alternatives to Compute Sanitizer?
Skills that share tags, products or a category with Compute Sanitizer: Graphsignal (graphsignal/graphsignal, 257 stars), LLM Torch Profiler Trace Analysis (BBuf/AI-Infra-Auto-Driven-SKILLS, 938 stars), Optimize Op (CVCUDA/CV-CUDA, 2.7k stars) and Cutlass Skill (slowlyC/agent-gpu-skills, 169 stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.
Who maintains Compute Sanitizer?
facebookexperimental (a GitHub organization, an official publisher) maintains it in facebookexperimental/triton, which has 201 GitHub stars. The repository holds 18 skills in this directory. The repository was last updated on October 11, 2026.
Source: facebookexperimental/triton on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.