kc-beta 0.5.5 → 0.6.0

package/template/skills/zh/meta-meta/evolution-loop/SKILL.md

@@ -241,12 +241,18 @@ description: Drive continuous improvement of skills and workflows through the di
 
 ### 停止条件
 
-当一轮迭代同时满足以下三个条件时，停止循环：
+当一轮迭代**同时**满足以下三个条件时，停止循环：
 
 1. 修正量 < 总测试案例的 5%。
 2. 新模式数 = 0。
 3. 回归数 = 0。
 
+**或者**满足单独的准确率收敛条件（D5，2026-04-23 新增）：
+
+4. 连续两轮迭代之间的整体准确率变化 < 1%。即 `|accuracy[N+1] - accuracy[N]| < 0.01`。
+
+条件 4 是为了防止观察到的过度迭代模式——从 v5 一直迭代到 v12，每轮都在 0.5% 的精度范围内来回波动。当模型已经达到"足够好"时，继续迭代只会消耗 token，不会带来实质改进。一旦准确率趋于稳定，应当进入下一阶段（蒸馏/生产）。
+
 如果修正量在连续两轮迭代之间**增加**，这是回归信号。暂停循环，先诊断原因再继续——上一轮的修复可能正在破坏系统的稳定性。
 
 ### 预期收敛速度

package/template/skills/zh/meta-meta/rule-extraction/SKILL.md

@@ -59,6 +59,79 @@ Rules will be distilled into workflows (see `skill-to-workflow`). Design with di
 ### Catalog Versioning
 When rules change (additions, modifications, deprecations), version the entire rule catalog as a unit. Individual rule versions track specific rules; the catalog version tracks the coherent set. Record the catalog version in `versions.json` alongside individual rule versions.
 
+## Granularity Calibration (read before extracting)
+
+A well-extracted rule catalog has **10-20 rules per typical regulation PDF**
+(2025 banking/insurance disclosure regs, 30-80 pages). Over-extraction into
+60-100 rules per regulation signals you're treating every clause as its own
+rule — which downstream consumers (skill-authoring, workflow-run) can't
+distinguish meaningful checks from boilerplate.
+
+If your first pass produces more than ~25 rules for a single regulation:
+- **Merge rules that share evidence and fail together** (e.g., "must disclose
+  X" and "must disclose Y" where both come from the same required-fields
+  table → one rule: "must disclose the required-fields list including X, Y").
+- **Drop procedural language** that isn't checkable against a report
+  (definitions, scope statements, references to other regs that just
+  transitively apply).
+- **Keep only checkable obligations, prohibitions, and thresholds** — the
+  things where you can read a sample report and say pass or fail.
+
+### Sample "good" rule
+
+```json
+{
+  "id": "R014",
+  "source_ref": "信披办法 §15.2",
+  "description": "季报应在季度结束后 15 个工作日内披露。",
+  "applicable_sections": ["公募产品"],
+  "severity": "high",
+  "machine_checkable": true,
+  "falsifiability_statement": "披露日期晚于季度结束后第 15 个工作日，则不合规",
+  "test_case_stub": "读取季报的披露日期 + 对应季度末日期，计算工作日差值"
+}
+```
+
+Note: one pass/fail outcome, a single `source_ref` to a specific clause,
+clear applicability scope. Skill-authoring can write `check_r014.py` from
+this alone.
+
+### Cross-regulation dedup (when working across multiple PDFs)
+
+If the developer user provides N regulations, rules from later regs often
+duplicate cross-cutting requirements already captured by earlier ones
+(e.g., 资管新规 2018 generic disclosure rule vs. 信披办法 2025's specific
+version). Before emitting a rule from reg-N:
+
+1. **Check the existing catalog.** Use `rule_catalog` (operation: list) to
+   see what's already there. Skip if a rule with equivalent scope + intent
+   exists.
+2. **Prefer the newer / more specific source_ref** when rules overlap.
+3. **If you merged rules**, record the consolidated sources in `source_ref`:
+   e.g., `"信披办法 §15.2 + 资管新规 §24"`.
+
+### Delegation to sub-agents
+
+If you dispatch extraction to sub-agents (one per regulation), the sub-agent
+inherits ONLY its `task_description` — it cannot see your conversation or
+existing catalog. Therefore, when composing the brief:
+
+- **Specify the target count band** explicitly: "Extract 10-20 atomic
+  rules from this regulation."
+- **Include a sample rule** in the brief body (paste the JSON above
+  verbatim) so the sub-agent's calibration matches yours.
+- **Name every regulation the sub-agent should process.** If AGENT.md
+  lists 10 core regulations, the brief must list all 10 by name, not
+  "the core regs" as a pronoun — LLMs composing long structured briefs
+  frequently drop items (observed in session 6304673afaa0 where reg 02
+  was silently omitted).
+- **State the dedup contract**: "Rules already in the parent's catalog
+  (R001–Rnnn) should NOT be re-extracted. If a requirement is already
+  covered, skip it." Then pass the current catalog's ID ranges.
+- **Prefer `rule_catalog` create operations over sandbox_exec writes to
+  catalog.json.** rule_catalog uses workspace file locking (B9);
+  sandbox_exec bypasses it and races with other writers.
+
 ## Extraction Strategies
 
 ### Strategy 1: Structured Input (Developer User Provides Rules)