scientific-rust-differential-test-design
Design differential and invariant tests for scientific Rust against a characterized independent reference, including floating-point edge cases and reproducible failing inputs.
- License 许可
- MIT
- 最近更新
- 2026-10-30
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- scientific-rust-differential-test-design
- description
- Design differential and invariant tests for scientific Rust against a characterized independent reference, including floating-point edge cases and reproducible failing inputs.
03 SKILL.md 正文
Read numerical contracts. Establish the reference contract first; use the reference-characterization skill when it is genuinely missing. Identify the slice's scientific claim, reference uncertainty and intended precision/feature/optimized-build matrix.
Design paired cases over the justified input domain, including near-zero outputs, boundaries, scale extremes, ill-conditioned regions and invalid inputs. Specify conversions, units, shape/length checks and nonfinite handling before any tolerance expression. State absolute/relative/norm or statistical criteria with a scientific rationale. Include at least one case that exposes a plausible implementation defect; passing identical copies of an algorithm does not provide independence.
Combine reference comparisons with applicable analytical solutions, convergence tests, conservation/metamorphic properties and negative cases. Explain why each property follows from this method. For randomized cases record seeds and streams, retain/minimize failing examples, and distinguish deterministic regression from statistical assurance. Never widen tolerances or discard failures without investigating their cause and obtaining any required criteria revision.
Fit the cases into existing project tests and named scientific-differential / scientific-invariants commands. For a design request, produce the test matrix and proposed commands without editing code. When implementation is requested, add tests within the allowed slice and run only authorized checks. Report actual results, missing reference coverage and resource limits; absent mandatory commands must remain visible as skipped/blocked checks.
04 同仓库更多 Skills
example-review
Review synthetic example changes for their stated acceptance criteria.
second-reader
Independent second reader for supply-chain and data documents (RFQs, supplier quotes, emails, PDFs including scanned, spreadsheets, CSV/JSON). Extracts values with source evidence, runs deterministic checks, compares quotes against an RFQ and a bid price, and briefs the human on what they may have missed while they review the same material. Every step is logged in a SQLite ledger.
second-reader-offline
The no-AI, no-network variant of second-reader for confidential client documents. Reads with local rules and OCR, checks and compares with code, and produces a self-contained HTML report on the user's own machine. The assistant must not read the documents or the report.
review
Review an engineering plan, diff, code or architecture from a general or selected professional perspective, challenging assumptions with evidence and no automatic implementation edits.
review-diff
Compatibility entry point for evidence-based diff review; establish the intended baseline and inspect consequential changes read-only.
plan-feature
Plan a new subsystem, cross-cutting feature, migration or explicitly requested design before implementation. Routine local edits can proceed directly.