메뉴
Enforces docs/4-research/honesty-audit.md policy by scanning PR diffs, READMEs, and presentation material for removed/unverified claims (87% MNIST, 30-states-as-fact, energy multipliers vs NAND/GPUs). Use before committing docs, PRs, or release notes that include accuracy or efficiency numbers.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Audits Go code for goroutine-to-widget access without fyne.Do(...) wrapping, the project's most common GUI freeze cause. Use when reviewing a PR that adds goroutines, async I/O, or simulation tickers in any pkg/gui/ or shell package.
Runs build flows for the canonical Fyne shell and module commands on this Go 1.25 monorepo. Use when building, packaging, or debugging build failures in cmd/fecim-lattice-tools or module GUI commands.
Relentlessly interviews the user about a proposed FeCIM physics, simulation, or GUI change before any code is written — covering source citation, educational-vs-validated framing, TDD-RED test, thread-safety, and honesty-audit alignment. Use when starting a non-trivial change to physics, ISPP, crossbar, or GUI logic.
Runs the FeCIM test matrix (full, race, module-scoped, coverage, golden regen) and interprets physics regression failures using the 5 known bug patterns. Use when running tests, debugging test failures, or regenerating physics golden files.
Verifies and formats FeCIM physics/measurement claims against the project's published-source list and docs/4-research/honesty-audit.md. Use when adding a numeric claim, accuracy figure, or device-parameter assertion to code, docs, PR descriptions, or commit messages.
Surveys FeCIM domain knowledge by searching references/, citations/, docs/4-research/, and the local Cognee KG, then synthesizes a cited research note. Use when investigating a physics topic, evaluating a paper, or grounding a design decision in literature.
| name | fecim-honesty-audit |
| description | Enforces docs/4-research/honesty-audit.md policy by scanning PR diffs, READMEs, and presentation material for removed/unverified claims (87% MNIST, 30-states-as-fact, energy multipliers vs NAND/GPUs). Use before committing docs, PRs, or release notes that include accuracy or efficiency numbers. |
Scan a diff, README, or PR description for the project's removed/unverified claim list before it lands. See tools/fecim-skills/_shared/fecim-context.md (Honesty-audit policy).
Take input. Either:
git diff main...HEAD -- '*.md' README.md docs/Regex-scan for trigger phrases (case-insensitive):
\b87%?\s*MNIST\b\b30\s*(analog\s+)?(states|levels)\b (followed by device|hardware|fact → red flag; with simulation default|configurable → ok)vs\.?\s*(NAND|GPUs?|DRAM) near energy/power numbers\bX×\s*(less|more)\s*(energy|power)\bMNIST|CIFAR|accuracy|efficiencyClassify each hit as one of:
verified — has a published citation in the project's list.educational-default — properly labeled as a simulation default.removed-unverified — matches the audit's removal list; must change.For removed-unverified, suggest the approved rephrasing:
Output a structured report:
File:line: <claim text>
Status: verified | educational-default | removed-unverified
Suggested change: <wording or DELETE>
Exit 0 if all clean, exit 1 (or print CHANGES REQUESTED) if any removed-unverified. Report exact file:line hits, counts by status, and whether the scan used rg or grep fallback.
our chip achieves 87% MNIST accuracy at 1000× lower energy than GPUs.
Expected: 2 hits flagged, both removed-unverified; suggests removal or reframing.Audit is observation — TDD: N/A. Wording changes are documentation-only and qualify under CLAUDE.md's TDD: N/A carve-out.