Conciseness

The body is dense and information-rich — nearly every line carries protocol specifics (schemas, step gates, file layouts) with no concept explanations Claude already knows, matching anchor 4's 'efficient; minor instances of over-explanation'. The trims that would reach anchor 5 are the design-rationale asides: the note on why invariant counts are not written ("不写条数：条数会随规范件演进而漂"), the blockquote distinguishing the plugin-side config.mode from session self-determination, and the '世界模型不是记忆库' framing sentence.

Actionability

For an instruction-only skill, guidance is concrete: exact file paths (ledger/YYYY-MM-DD.jsonl, runs/<session-id>.jsonl, current.yaml), full field-level schemas (event_id, prediction_id, falsifier, time_horizon...), parseable time_horizon formats ("2026-10-15T12:00:00Z", "+30m/+12h/+3d/+2w"), an 11-step named pipeline, and mechanical tool-risk tiers. It falls short of anchor 5 because the executing machinery it cites — the adapters (mind/distill.py, mind/lp_evaluator.py), the governance/proposal path, and the guard implementation — is external and unspecified in the bundle, leaving gaps in how key steps actually execute.

Workflow Clarity

The 执行时序 section gives a strictly ordered 11-step sequence ("顺序即验收" — order is acceptance) with explicit validation checkpoints (PREDICTION_EVALUATED confirmed/refuted/partial; guards refuse consequential actions without a bound prediction) and feedback loops (failed prediction → residual label + one of three explicit revision paths), plus a session-end checklist. It does not reach anchor 5 because part of the authoritative sequencing is delegated to instruction/generated.md §6, which is not in the bundle, leaving ambiguity about which form governs.

Progressive Disclosure

Sections are clearly headed and well-ordered (模式, 不变量, 状态层, 执行时序, 各环要点, Canonical 写入门槛, 会话结束契约), but the skill ships no bundle files at all, while the body references many external files that are not in the bundle (instruction/generated.md §1/§3/§5/§6, governance.yaml, dsh/world-model/cordis.patch.yml, mind/distill.py, mind/lp_evaluator.py, proposals/) — dangling navigation targets. Additionally, ~50 lines of inline schema/field definitions (状态层 block, episode and prediction schemas) are content that would typically live in a reference file. This fits anchor 3: structure present, references clearly named but unusable, and reference-type content inlined; not anchor 2 since structure and signaling are real, not 'minimal' or 'buried'.

Specificity

The description enumerates a chain of concrete protocol actions ("恢复持久状态→建立竞争模型→显式预测→真实观察→修订模型→主动取证→决策→持久化" — restore state, build competing models, explicit predictions, real observations, revise models, active forensics, decisions, persist), which is several specific actions rather than vague filler. It falls short of anchor 5 because the actions are process-meta verbs with no concrete artifacts or object types (no file formats, tools, or deliverables named), and '主动取证' (active forensics) and '建立竞争模型' are somewhat abstract jargon.

Completeness

Both questions are explicitly answered: the 'what' is the full closed-loop protocol chain with named stages, and the 'when' is a multi-condition trigger clause with an exclusion boundary ("用于多步、高风险、跨会话、存在竞争解释或需要持久世界模型的真实任务，不用于简单一次性任务" — use for multi-step/high-risk/cross-session tasks with competing interpretations, NOT for simple one-off tasks). This matches anchor 5's explicit what-and-when with concrete trigger conditions; the when-clause is more specific than anchor 4's example, not less.

Trigger Term Quality

Trigger terms are present and relevant ("多步" multi-step, "高风险" high-risk, "跨会话" cross-session, "存在竞争解释" competing interpretations, "简单一次性任务" simple one-off tasks), but the vocabulary is protocol-internal jargon ("世界模型" world model, "竞争解释") rather than natural user phrasing, and no synonyms or common variations are covered (e.g. 'high-stakes', 'long-running', 'multi-session'). It fits anchor 3: some relevant keywords but missing common variations — not anchor 4's 'good keyword coverage', and clearly above anchor 2's one-or-two generic keywords.

Distinctiveness Conflict Risk

The niche is distinct — a persistent world-model closed-loop protocol with prediction/falsification gating is a clear specialty unlikely to collide with file-manipulation or tooling skills, and the explicit exclusion ("不用于简单一次性任务") sharpens the boundary. It does not reach anchor 5 because the trigger characteristics 'multi-step, high-risk tasks' are broad meta-conditions that could overlap with general planning or reasoning skills.