skills.rest
WangWindow
WangWindowCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

aog-prior-art-verify

Scan, stage, build, and classify prior-art arch35 operator candidates with digest-bound provenance.

Updated Sep 15, 2026
Search Tags:#provenance#prior art#build verification#cann#sha-256#ascendc#operator migration
One-click install
npx skills add https://github.com/WangWindow/CANN-BatchMatMulMaxsum --skill aog-prior-art-verify-wangwindow
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: aog-prior-art-verify
Source: https://github.com/WangWindow/CANN-BatchMatMulMaxsum/tree/main/.agents/skills/aog-prior-art-verify
Command: npx skills add https://github.com/WangWindow/CANN-BatchMatMulMaxsum --skill aog-prior-art-verify-wangwindow

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve? During arch22 to arch35 Ascend C operator migrations, existing target-architecture implementations may already exist, but reusing them blindly risks treating unverified code as ground truth. This Skill provides a controlled pipeline to discover, stage, build, measure, and learn from prior-art candidates without ever replacing the independent truth source or customer-facing verification. ## Core Features & Use Cases - Provenance-bound staging: scan_prior_art.py detects candidate implementations (upstream arch35, shared-common, V220 entries, workspace stashes) and stage_candidate.py copies only digest-authorized files into .prior_art_candidate/ with a SHA-256 manifest. - Isolated build and verification: build_candidate.py compiles the candidate in a unique container-side checkout of ops-nn, and verify_candidate.py measures it against the freshly provisioned independent truth (npubench bundle or live arch22 capture). - Advisory classification and learning: classify.py emits CANDIDATE_PASS / PRECISION_GAP / PERF_GAP / DET_GAP / BUILD_GAP verdicts, and learn_extract.py mechanically extracts API-surface signals into reviewable knowledge candidates. - Use Case: When migrating an operator such as adaptive_avg_pool3d, run the pipeline to test whether the upstream arch35 sibling implementation passes precision and performance floors, then feed the resulting seed or counterexamples into the standard worker brief. ## Quick Start Ask the agent to run the aog-prior-art-verify pipeline for the current operator so it scans for prior art, stages and builds the candidate, and reports an advisory verdict against the fresh independent truth.

Frequently Asked Questions about aog-prior-art-verify

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I verify a prior-art candidate during an arch22 to arch35 operator migration?▼

Run the pipeline in order: scan_prior_art.py to discover candidates, stage_candidate.py to copy digest-authorized files, build_candidate.py to compile in an isolated container checkout, then verify_candidate.py and classify.py against the freshly provisioned independent truth.

Can a prior-art candidate replace the independent truth source?▼

No. A target, sibling, or archive candidate is always prior art, never an independent truth. The npubench KernelBench-style task bundle or a fresh arch22 live capture is mandatory for every run, and cached captures or candidate outputs cannot substitute.

What verdicts does the prior-art classification produce?▼

classify.py emits CANDIDATE_PASS, CANDIDATE_PRECISION_GAP, CANDIDATE_PERF_GAP, CANDIDATE_DET_GAP, or CANDIDATE_BUILD_GAP. Each verdict is advisory only and feeds the worker brief; it never writes customer-facing verification results.

Does the scan use upstream arch35 sources by default?▼

No. Target-architecture discovery detectors run only when OPGEN_PRESTAGE_ARCH35=1 is explicitly set. Under the default, the scan prefers V220-pure entries and suppresses cross-arch apt dispatchers to avoid leaking arch35 code into generation.

What happens when the candidate build fails or times out?▼

Build failure and timeout are valid outcomes, not exceptions. They are recorded as CANDIDATE_BUILD_GAP counterexamples, and the standard worker generation and fresh truth capture continue normally.

Where are the prior-art verification outputs stored?▼

All outputs are workspace-local and advisory: .prior_art_candidate/manifest.json, the build directory, verification_prior_art.json, prior_art_verdict.json, and prior_art_learn.md. Candidate evidence is never copied into the customer-facing verification.json.

Related Skills

Explore More Software Engineering Skills→
walm00walm00
18

context-routing

Search documents across BCOS zones and assemble task-specific context bundles.

Community
Intermediate
ShineBreakerShineBreaker
2

adversarial-review-trigger

Attacks task completion claims by constructing counterexamples against worker handoff evidence.

Community
Advanced
AvatarGanymedeAvatarGanymede

gui-learn

Extracts reusable lessons from GUI development sessions into a persistent knowledge base.

Community
Advanced
CAPHTECHCAPHTECH
1

contradiction-check

Detect contradictory definitions and inconsistent constraints across codebases and requirements.

Official
Advanced