Produce and semantically review evidence-grounded Yuansheng Root Cause Blueprint v1-lite candidates from validated perf evidence and a confirmed hardware profile. Use when Yuansheng Trace requests a per-function Blueprint candidate or an independent review of one.
Produce and semantically review evidence-grounded Yuansheng Root Cause Blueprint v1-lite candidates from validated perf evidence and a confirmed hardware profile. Use when Yuansheng Trace requests a per-function Blueprint candidate or an independent review of one.
Write a Root Cause Blueprint
Bind the candidate to the supplied software, test case, ranked function, and
confirmed hardware profile.
Use only the supplied perf statistics, annotate output, metadata, hardware
profile, and their recorded hashes. Do not read source code or acquire new
evidence.
Distinguish observations from hypotheses. Copy measured values exactly and
keep unsupported facts as null with the required machine-readable gap
codes.
Keep every rvv_pattern and pattern_confidence value null, and include
pattern_catalog_unavailable. Do not name or infer a Pattern.
Do not invent a source file, line, code location, metric, baseline, or
hardware capability. Keep recommendations within what the evidence can
support.
Set needs_human_review to true and allow_auto_forward_to_ys_craft to
false.
Return strict JSON conforming to the supplied v1-lite Schema and preserve the
workflow-requested claim-to-evidence references.
For semantic review, compare every material claim with the immutable candidate
and supplied evidence. Report unsupported, conflicting, or uncited claims
without modifying the candidate. Require a new candidate and full validation
when correction is necessary.