GPU Wiki

In every natural-language Wiki request, state the true target product separately from the authoritative runtime architecture; never infer either from a scheduler label or masked device name. If the true product is unavailable, say so rather than substitute another GPU. Ask for that product's hardware specification and relevant architecture/ISA facts alongside the operator, DSL and bottleneck. Distinguish measurements from hypotheses. Do not reduce the question to arbitrary keywords.

python3 tools/sandbox.py --kind wiki-query "Target B200, runtime sm_100. Triton normalization is measured DRAM-bound; which fusion strategies apply? Include B200 hardware specifications and relevant architecture/ISA facts." --brief
python3 tools/sandbox.py --kind wiki-query --file scratch/wiki-question.txt
python3 tools/sandbox.py --kind wiki-search --arch sm_120 --dsl triton --coverage
python3 tools/sandbox.py --kind wiki-hardware --product sm120

Use the tools/sandbox.py --kind wiki-* commands above from an Agent workspace. The Supervisor owns stores, the Wiki-only model, timeouts, concurrency and query telemetry; never override their host paths or provider configuration. Repeating an identical query in this Episode reuses the same completed response and query_id.

For repairable: true with error.code: "request_not_started", wait at least retry_after_seconds before retrying the unchanged query with backoff. The query executor has not started. Do not blindly retry transport failures or unknown outcomes. For wiki_queue_timeout, wiki_query_in_progress, wiki_query_timeout or wiki_unavailable, wait at least retry_after_seconds and retry the same read-only query unchanged. Do not submit parallel copies or wrap the call in a shorter shell timeout. If Wiki remains unavailable, continue with other tools and report the blocker. A Wiki failure is not a GPU environment failure.

Check coverage before restrictive type filters. Read the returned source, scope, payload and notes. A zero-match fallback sample is labelled, not advice for your question. A missing hardware spec is not permission to substitute another GPU. When research recording is enabled, record query_id and canonical wiki_id in the existing Experiment Journal when knowledge materially affects an experiment; retrieval alone is not adoption.