菜单
Testing discipline for drmSEM — pure-logic kernels run without drmTMB, integration tests gated by skip_if_not_installed, every effect/estimand/d-sep rule needs a known-DGP recovery test, evidence tiers recorded in the validation ledger. Use when adding tests or any new estimand/rule.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Keep drmSEM CRAN-clean — R CMD check 0 errors/0 warnings, examples that need drmTMB wrapped in \dontrun, Remotes for the GitHub engine, no compiled code in drmSEM itself, DESCRIPTION hygiene, ASCII, reasonable file sizes, reproducible RNG. Use before release or when touching DESCRIPTION/examples/CI.
Reader-facing docs for drmSEM — README, the vignette, and pkgdown reference sync. Pair equations + R syntax + interpretation, show component-labelled output and the distribution-mediated decomposition, and tell users what to try when a path/family is unsupported. Use when writing user-facing prose.
THE core skill. Enforces component-labelled paths — never describe or compute a non-mu path as a mean effect. Use for ANY prose, docs, message, or code that states what a drmSEM path MEANS.
How drmSEM talks to drmTMB — ONLY through R/extractors.R. The drmTMB return-shape facts, the dpar:term coefficient convention, and the rule to confirm parameterizations against a live fit before trusting samplers.
Build basis sets and run any-component d-separation / Fisher's C correctly — adjacency = direct parent on ANY component, causal ordering, conditioning on parents, LRT df bookkeeping, honest framing as drmSEM's definition.
R package craft for drmSEM — roxygen2 on every export, S3 conventions, the document/test/check dev loop, NAMESPACE sync, cli messaging, tests shipped with implementation, small focused commits.
| name | validation-harness |
| description | Testing discipline for drmSEM — pure-logic kernels run without drmTMB, integration tests gated by skip_if_not_installed, every effect/estimand/d-sep rule needs a known-DGP recovery test, evidence tiers recorded in the validation ledger. Use when adding tests or any new estimand/rule. |
Two test tiers, kept strictly separate so the suite runs on a toolchain-free machine and still proves correctness where it counts.
test-dsep-kernels.R, test-effect-kernels.R, test-utils.R.drmTMB absent. If a kernel needs a fit, it is in the
wrong tier — refactor the pure math out of the adapter.test-integration.R. First line of the file/block:
skip_if_not_installed("drmTMB").drmTMB nodes, build a drm_sem, run paths(), dsep(),
fisher_c(), effects end to end. Keep these small and fast.Per AGENTS.md design rule 1: do NOT add a public effect type, estimand, or d-separation rule without a simulation test that recovers a KNOWN data-generating process.
set.seed() always; reuse
helper-dgp.R).skip_if_not_installed guarded.sigma effect and
report it as mu).After validating a claim, record it in docs/memory/VALIDATION_LEDGER.md with a
tier:
validated — recovery test passes against a known DGP.experimental — implemented, smoke-tested, recovery still pending.assumed — relies on an unverified drmTMB parameterization (link to the
drmtmb-adapter confirmation TODO).
Each entry: claim, test file/line, tier, date. Update the ledger in the same
change that adds the test; also note it in docs/memory/AGENT_LOG.md.Keep recovery DGPs small (low n, few draws) so the suite stays fast; use a larger N only when a tolerance genuinely requires it. Cover ordinary, edge, and malformed inputs (Curie's mandate).