eduevidence 6.2.0 → 6.3.0

package/domains/policy/copy/module_labels.json

@@ -0,0 +1,168 @@
+{
+  "domain": "policy",
+  "purpose": "Domain-sensitive module/UI labels (policy voice). target_learners displays as 目标人群; intervention is 干预方案.",
+  "labels": {
+    "outcome_separation_title": {
+      "zh": "结果分离 · 过程产出 ≠ 政策效果",
+      "en": "Outcome Separation · Process output ≠ policy effect"
+    },
+    "outcome_sep_note": {
+      "zh": "将不同结果类型分开裁决，避免把试点期产出或效率提升直接等同于真实政策效果。",
+      "en": "Outcomes are adjudicated separately so pilot throughput or efficiency gains are not silently treated as real policy effects."
+    },
+    "outcome_group_task": {
+      "zh": "过程 / 近端产出",
+      "en": "Process / proximal output"
+    },
+    "outcome_group_learning": {
+      "zh": "效果 / 目标结果",
+      "en": "Effect / target outcome"
+    },
+    "outcome_group_risk": {
+      "zh": "风险 / 实施风险",
+      "en": "Risk / implementation risk"
+    },
+    "outcome_group_other": {
+      "zh": "其他结果",
+      "en": "Other outcomes"
+    },
+    "method_guard": {
+      "zh": "产出 vs 效果护栏",
+      "en": "Output vs effect guard"
+    },
+    "applicability": {
+      "zh": [
+        "适用于谁",
+        "适用辖区",
+        "适用结果",
+        "适用条件",
+        "目标人群",
+        "目标情境"
+      ],
+      "en": [
+        "Suitable for",
+        "Jurisdiction",
+        "Outcomes",
+        "Conditions",
+        "Target population",
+        "Target context"
+      ]
+    },
+    "intervention_learners": {
+      "zh": "目标人群",
+      "en": "Target population"
+    },
+    "intervention_population": {
+      "zh": "目标人群",
+      "en": "Target population"
+    },
+    "intervention_duration": {
+      "zh": "试点时长",
+      "en": "Pilot duration"
+    },
+    "intervention_policy": {
+      "zh": "干预使用规则",
+      "en": "Intervention usage policy"
+    },
+    "intervention_rule": {
+      "zh": "干预规则",
+      "en": "Intervention rule"
+    },
+    "intervention_activities": {
+      "zh": "活动",
+      "en": "Activities"
+    },
+    "intervention_check": {
+      "zh": "结果检查",
+      "en": "Outcome check"
+    },
+    "intervention_stop": {
+      "zh": "停止条件",
+      "en": "Stop conditions"
+    },
+    "intervention_timeline": {
+      "zh": "干预方案时间线信息图",
+      "en": "Intervention plan timeline infographic"
+    },
+    "evaluation_measures": {
+      "zh": [
+        "基线",
+        "试点结束",
+        "持续监测",
+        "扩展评估"
+      ],
+      "en": [
+        "Baseline",
+        "Pilot end",
+        "Durability",
+        "Scale-out"
+      ]
+    },
+    "evaluation_metrics": {
+      "zh": [
+        "过程指标",
+        "效果指标",
+        "风险指标"
+      ],
+      "en": [
+        "Process metrics",
+        "Effect metrics",
+        "Risk metrics"
+      ]
+    },
+    "evaluation_question": {
+      "zh": "研究问题",
+      "en": "Research question"
+    },
+    "evaluation_threshold": {
+      "zh": "成功阈值",
+      "en": "Success threshold"
+    },
+    "evaluation_plan": {
+      "zh": "分析计划",
+      "en": "Analysis plan"
+    },
+    "evaluation_figure": {
+      "zh": "评价设计信息图",
+      "en": "Evaluation design infographic"
+    },
+    "svg_intervention_title": {
+      "zh": "干预方案时间线",
+      "en": "Intervention plan timeline"
+    },
+    "svg_intervention_desc": {
+      "zh": "各试点阶段的短名称与活动数量；完整干预使用规则见阶段说明块。",
+      "en": "Short phase names and activity counts; full intervention usage rules are in the phase blocks."
+    },
+    "svg_evaluation_title": {
+      "zh": "评价设计流程",
+      "en": "Evaluation design flow"
+    },
+    "svg_evaluation_desc": {
+      "zh": "基线、试点结束、持续监测与扩展评估的评价流程；完整指标与分析计划见评估章节。",
+      "en": "Evaluation flow across baseline, pilot end, durability and scale-out; full metrics and analysis plan are in the evaluation section."
+    },
+    "svg_workflow_desc": {
+      "zh": "从问题框架、检索、抓取验证、证据抽取、反方质疑、方法审计、裁决到适用性与干预评价的完整流程。",
+      "en": "Research flow from framing, retrieval, fetch/verify, extraction, challenge, method audit and adjudication to applicability and intervention evaluation."
+    },
+    "decision_kpi": {
+      "zh": [
+        "决策",
+        "置信度",
+        "证据最充分的结果",
+        "最不确定的结果",
+        "主要风险",
+        "来源数量"
+      ],
+      "en": [
+        "Decision",
+        "Confidence",
+        "Best-supported outcome",
+        "Most uncertain outcome",
+        "Main risk",
+        "Sources"
+      ]
+    }
+  }
+}

package/domains/policy/copy/risk_constructs.json

@@ -0,0 +1,33 @@
+{
+  "domain": "policy",
+  "purpose": "Outcome grouping + separation guard constructs (process output is not policy effect). No teaching vocabulary.",
+  "outcome_groups": {
+    "task": [
+      "cost_effectiveness",
+      "feasibility"
+    ],
+    "learning": [
+      "policy_effectiveness",
+      "equity"
+    ],
+    "risk": [
+      "implementation_risk"
+    ]
+  },
+  "separation_title": {
+    "zh": "结果分离 · 过程产出 ≠ 政策效果",
+    "en": "Outcome Separation · Process output ≠ policy effect"
+  },
+  "separation_note": {
+    "zh": "将不同结果类型分开裁决，避免把试点期产出或效率提升直接等同于真实政策效果。",
+    "en": "Outcomes are adjudicated separately so pilot throughput or efficiency gains are not silently treated as real policy effects."
+  },
+  "method_guard": {
+    "zh": "产出 vs 效果护栏",
+    "en": "Output vs effect guard"
+  },
+  "guard_principle": {
+    "zh": "过程产出提升 ≠ 政策效果实现；ADOPT 需要独立效果、持续性与公平证据。",
+    "en": "Process output is not policy effect; ADOPT requires independent effect, durability and equity evidence."
+  }
+}

package/domains/policy/copy/section_titles.json

@@ -0,0 +1,186 @@
+{
+  "domain": "policy",
+  "purpose": "Section titles/leads + default full-report chapter plan (policy voice; no teaching vocabulary).",
+  "section_titles": {
+    "zh": {
+      "01": "01 执行决策",
+      "02": "02 结果证据概览",
+      "03": "03 证据矩阵",
+      "04": "04 证据裁决",
+      "05": "05 方法学审计",
+      "06": "06 冲突分析",
+      "07": "07 主张-证据追溯",
+      "08": "08 适用性",
+      "09": "09 干预方案",
+      "10": "10 评价方案",
+      "11": "11 基准测试",
+      "12": "12 来源与溯源"
+    },
+    "en": {
+      "01": "01 Executive Decision",
+      "02": "02 Outcome Evidence Overview",
+      "03": "03 Evidence Matrix",
+      "04": "04 Evidence Tribunal",
+      "05": "05 Methodology Audit",
+      "06": "06 Conflict Analysis",
+      "07": "07 Claim-Evidence Trace",
+      "08": "08 Applicability",
+      "09": "09 Intervention Plan",
+      "10": "10 Evaluation Plan",
+      "11": "11 Benchmark",
+      "12": "12 Sources & Provenance"
+    }
+  },
+  "section_leads": {
+    "zh": {
+      "01": "本节先给结论：最终怎么裁决、置信度多高、靠哪几条证据。",
+      "02": "一图看清：哪些政策效果有支持证据、哪些被反驳。",
+      "03": "每条证据来自哪项研究、测了什么、方向与质量如何；可筛选、可搜索。",
+      "04": "证据允许主张什么、不允许主张什么；缺失的关键证据是什么。",
+      "05": "研究质量可靠吗？哪些方法学问题让结论打折。",
+      "06": "不同研究为何结论不同；分歧出在哪一环。",
+      "07": "从结论到证据到原始来源，每一步都可追查。",
+      "08": "结论适用于谁、什么人群与效果、需要什么条件。",
+      "09": "试点怎么分阶段放开干预规则；什么情况必须叫停。",
+      "10": "如何验证效果：指标、对照、成功阈值。",
+      "11": "EduEvidence 自身基准表现：引用精度与成本。",
+      "12": "每篇文献是谁、出自哪里、如何获取。"
+    },
+    "en": {
+      "01": "The verdict first: what we decide, at what confidence, on which evidence.",
+      "02": "At a glance: which policy outcomes have supporting evidence, which are contradicted.",
+      "03": "Where each piece of evidence comes from, what it measures, its direction and quality — filterable and searchable.",
+      "04": "What the evidence lets us claim, what it does not, and what is still missing.",
+      "05": "How reliable are these studies, and which methodological concerns discount the conclusions.",
+      "06": "Why studies disagree — and where exactly they diverge.",
+      "07": "Every step from conclusion to evidence to source stays traceable.",
+      "08": "Who the conclusion applies to, for which population and outcomes, under what conditions.",
+      "09": "How intervention rules phase in during a pilot, and when we must stop.",
+      "10": "How we verify real effects: metrics, comparison, success threshold.",
+      "11": "How EduEvidence itself performs: citation precision and cost.",
+      "12": "Who wrote each cited study, where it came from, how it was fetched."
+    }
+  },
+  "default_chapters": [
+    {
+      "key": "decision",
+      "title_zh": "结论、裁决与研究边界",
+      "title_en": "Decision, Adjudication & Research Boundary",
+      "lead_zh": "先明确最终裁决与研究边界，再解释为什么。",
+      "lead_en": "State the final adjudication and research boundary before explaining why.",
+      "modules": [
+        "decision",
+        "scope"
+      ]
+    },
+    {
+      "key": "evidence",
+      "title_zh": "关键证据与结果分离",
+      "title_en": "Key Evidence & Outcome Separation",
+      "lead_zh": "把过程产出、真实效果、持续性与风险放在同一证据地图中，但不混为一谈。",
+      "lead_en": "Place process outputs, real effects, durability and risk on one evidence map without conflating them.",
+      "modules": [
+        "retrieval",
+        "outcomes",
+        "evidence"
+      ]
+    },
+    {
+      "key": "quality",
+      "title_zh": "证据可信度、反证与方法审计",
+      "title_en": "Evidence Quality, Counterevidence & Method Audit",
+      "lead_zh": "检查证据为什么可信、哪里冲突，以及哪些结论必须降级。",
+      "lead_en": "Examine why evidence is credible, where it conflicts, and which conclusions require downgrading.",
+      "modules": [
+        "quality",
+        "conflicts",
+        "trace"
+      ]
+    },
+    {
+      "key": "action",
+      "title_zh": "适用范围与政策行动",
+      "title_en": "Applicability & Policy Action",
+      "lead_zh": "把可外推范围、护栏和政策动作连接到具体证据。",
+      "lead_en": "Connect applicability, guardrails and policy actions to specific evidence.",
+      "modules": [
+        "applicability",
+        "intervention"
+      ]
+    },
+    {
+      "key": "evaluation",
+      "title_zh": "试点设计、评估与停止条件",
+      "title_en": "Pilot, Evaluation & Stop Conditions",
+      "lead_zh": "用独立效果测量验证试点，并预先写清停止条件。",
+      "lead_en": "Validate the pilot with independent effect measures and pre-specified stop conditions.",
+      "modules": [
+        "evaluation"
+      ]
+    },
+    {
+      "key": "sources",
+      "title_zh": "来源、溯源与附录",
+      "title_en": "Sources, Traceability & Appendix",
+      "lead_zh": "保留原始来源、URL、证据 ID 和获取信息，确保可回查。",
+      "lead_en": "Preserve original sources, URLs, evidence IDs and retrieval metadata for auditability.",
+      "modules": [
+        "sources"
+      ]
+    }
+  ],
+  "brief_blocks": {
+    "zh": {
+      "decision": [
+        "先看结论",
+        "该不该做、置信度多高、最关键的证据边界在哪。"
+      ],
+      "lieflat": [
+        "Lieflat 实证手作画廊",
+        "AI 按数据形状从 Lieflat 目录选型编排；每张图的数字都可溯源到 result.json。"
+      ],
+      "outcomes": [
+        "过程产出 ≠ 政策效果",
+        "只展示真正有解释力的结果分离；正向、负向与零效应按 effect_direction 编码。"
+      ],
+      "tribunal": [
+        "证据裁决",
+        "支持、不确定、被反驳与缺失证据分开放置，不把长段落平铺在同一层。"
+      ],
+      "action": [
+        "从证据到行动",
+        "适用性、护栏、停止条件与评价连成一条可执行路径。"
+      ],
+      "sources": [
+        "关键来源",
+        "摘要页只列最关键的来源；完整溯源在完整报告中展开。"
+      ]
+    },
+    "en": {
+      "decision": [
+        "Decision first",
+        "What to do, how confident we are, and the most important evidence boundary."
+      ],
+      "lieflat": [
+        "Lieflat Editorial Gallery",
+        "Charts selected and composed by AI from the Lieflat catalog; every number traces back to result.json."
+      ],
+      "outcomes": [
+        "Process output ≠ policy effect",
+        "Only informative outcome separation; positive, negative and null effects use effect_direction."
+      ],
+      "tribunal": [
+        "Evidence tribunal",
+        "Supported, uncertain, contradicted and missing evidence stay separated instead of flattened into long prose."
+      ],
+      "action": [
+        "Evidence to action",
+        "Applicability, guardrails, stop conditions and evaluation form one executable path."
+      ],
+      "sources": [
+        "Key sources",
+        "Only the key sources in the brief; full traceability expands in the full report."
+      ]
+    }
+  }
+}

package/domains/policy/copy/terminology.json

@@ -0,0 +1,64 @@
+{
+  "domain": "policy",
+  "purpose": "Domain terminology (policy voice). target_learners → 目标人群; intervention → 干预方案; no teaching vocabulary.",
+  "methodology_labels": {
+    "zh": {
+      "evidence_level": "证据等级",
+      "causal_identification": "因果识别",
+      "external_validity": "外部效度",
+      "cost_evidence": "成本证据",
+      "stakeholder_representation": "利益相关者代表",
+      "implementation_evidence": "实施证据",
+      "equity_analysis": "公平性分析",
+      "effect_size_reported": "效应量报告",
+      "uncertainty_quantified": "不确定性量化",
+      "comparator_clarity": "对照与反事实清晰",
+      "publication_bias_risk": "发表偏倚风险",
+      "generalizability_claims": "可推广性声明"
+    },
+    "en": {
+      "evidence_level": "Evidence level",
+      "causal_identification": "Causal identification",
+      "external_validity": "External validity",
+      "cost_evidence": "Cost evidence",
+      "stakeholder_representation": "Stakeholder representation",
+      "implementation_evidence": "Implementation evidence",
+      "equity_analysis": "Equity analysis",
+      "effect_size_reported": "Effect size reported",
+      "uncertainty_quantified": "Uncertainty quantified",
+      "comparator_clarity": "Comparator clarity",
+      "publication_bias_risk": "Publication-bias risk",
+      "generalizability_claims": "Generalizability claims"
+    }
+  },
+  "field_labels": {
+    "zh": {
+      "target_learners": "目标人群",
+      "target_population": "目标人群",
+      "ai_usage_policy": "干预使用规则",
+      "ai_usage_rule": "干预规则",
+      "task_vs_learning_guard": "产出 vs 效果护栏",
+      "learning_metrics": "效果指标",
+      "process_metrics": "过程指标",
+      "risk_metrics": "风险指标",
+      "task_performance": "过程产出",
+      "learning_gain": "目标效果",
+      "retention": "持续性",
+      "transfer": "扩散效果"
+    },
+    "en": {
+      "target_learners": "Target population",
+      "target_population": "Target population",
+      "ai_usage_policy": "Intervention usage policy",
+      "ai_usage_rule": "Intervention rule",
+      "task_vs_learning_guard": "Output vs effect guard",
+      "learning_metrics": "Effect metrics",
+      "process_metrics": "Process metrics",
+      "risk_metrics": "Risk metrics",
+      "task_performance": "Process output",
+      "learning_gain": "Target effect",
+      "retention": "Durability",
+      "transfer": "Spillover"
+    }
+  }
+}

package/engine/capabilities.py

@@ -19,20 +19,32 @@ class CapabilitySpec:
     deterministic_local: bool
     scientific_gate: str | None
     workflow_ids: tuple[str, ...] = ()
+    experimental: bool = False
 
 
 _REGISTRY: dict[str, CapabilitySpec] = {}
 
+#: Jev / SemDecide Tier-0 acceleration IDs (opt-in via --experimental).
+#: Not role-owned scientific work: omitted from capability_registry() by
+#: default so the protocol role-ownership gate stays unchanged. Pass
+#: include_experimental=True when wiring skill/workflows/experimental-jev.md.
+EXPERIMENTAL_CAPABILITY_IDS: frozenset[str] = frozenset({
+    "content_screen", "semantic_rerank", "field_extract",
+    "classify_check", "claim_verify",
+})
+
 
 def _register(capability_id: str, *, input_contracts: tuple[str, ...],
               output_contracts: tuple[str, ...], deterministic_local: bool,
-              scientific_gate: str | None = None) -> None:
+              scientific_gate: str | None = None,
+              experimental: bool = False) -> None:
     _REGISTRY[capability_id] = CapabilitySpec(
         capability_id=capability_id,
         input_contracts=input_contracts,
         output_contracts=output_contracts,
         deterministic_local=deterministic_local,
         scientific_gate=scientific_gate,
+        experimental=experimental,
     )
 
 
@@ -90,10 +102,50 @@ _register("intervention_design", input_contracts=("decision-snapshot",),
 _register("evaluation_design", input_contracts=("decision-snapshot",),
           output_contracts=("evaluation",), deterministic_local=False)
 
-
-def capability_registry() -> dict[str, CapabilitySpec]:
-    """Return the frozen capability registry."""
-    return dict(_REGISTRY)
+# Jev experimental Tier-0 (opt-in --experimental; mode 1/2/3).
+# Acceleration only — never replaces Skeptic, pre_verdict_gate, methodology
+# audit, or tribunal. Uncertain / invalid answers fail closed to the large model.
+_register("content_screen", input_contracts=("fetch-result",),
+          output_contracts=("screen-verdict",), deterministic_local=False,
+          scientific_gate="experimental Tier-0; advisory screen before context; "
+                          "RULE 2 snippet != evidence",
+          experimental=True)
+_register("semantic_rerank", input_contracts=("source",),
+          output_contracts=("ranked-candidates",), deterministic_local=False,
+          scientific_gate="experimental Tier-0; ranking is discovery order only",
+          experimental=True)
+_register("field_extract", input_contracts=("source",),
+          output_contracts=("field-values",), deterministic_local=False,
+          scientific_gate="experimental Tier-0; verbatim substring values only, "
+                          "never model-written fields",
+          experimental=True)
+_register("classify_check", input_contracts=("source",),
+          output_contracts=("classification",), deterministic_local=False,
+          scientific_gate="experimental Tier-0; labels never replace "
+                          "methodology_appraisal or outcome separation",
+          experimental=True)
+_register("claim_verify", input_contracts=("claim", "source"),
+          output_contracts=("claim-verdict",), deterministic_local=False,
+          scientific_gate="experimental Tier-0; does NOT replace Skeptic nine "
+                          "checks or Pre-Verdict Gate",
+          experimental=True)
+
+
+def capability_registry(*, include_experimental: bool = False) -> dict[str, CapabilitySpec]:
+    """Return the capability registry.
+
+    Experimental Jev Tier-0 IDs are omitted by default (they are acceleration
+    helpers, not role-owned scientific stages). Opt in with
+    ``include_experimental=True`` when resolving --experimental modes 1/2/3.
+    """
+    if include_experimental:
+        return dict(_REGISTRY)
+    return {cid: spec for cid, spec in _REGISTRY.items() if not spec.experimental}
+
+
+def experimental_capability_registry() -> dict[str, CapabilitySpec]:
+    """Return only the Jev/SemDecide Tier-0 experimental capabilities."""
+    return {cid: spec for cid, spec in _REGISTRY.items() if spec.experimental}
 
 
 def capability(capability_id: str) -> CapabilitySpec | None:

package/engine/decision_policy.py

@@ -1,4 +1,4 @@
-"""Single source of truth for the ADOPT direct-evidence gate.
+"""Single source of truth for the four-state decision gate.
 
 The four-state decision (ADOPT / PILOT / REJECT / INSUFFICIENT EVIDENCE) is
 computed by two layers that must never drift apart:
@@ -12,6 +12,15 @@ liked. Both layers now resolve the 'which outcome categories count for this
 domain' question here, and the V1 gate additionally re-derives direct-evidence
 presence from the pack's own evidence records.
 
+Decision matrix (gate-enforced):
+
+  conflict/mixed/support+oppose        -> INSUFFICIENT_EVIDENCE (unresolved_conflict)
+  oppose_adoption (oppose-only)        -> REJECT
+  High    + support + direct_primary   -> ADOPT
+  High    + support + missing direct   -> PILOT   (downgrade_reason=missing_direct_primary)
+  Moderate+ support (+/- direct)    -> PILOT   (confidence_band | missing_direct_primary)
+  Low / 无支持 / 间接无 Moderate    -> INSUFFICIENT_EVIDENCE
+
 Stdlib only, consistent with the "Native Core" policy of engine/.
 """
 from __future__ import annotations
@@ -33,6 +42,18 @@ ADOPT_DIRECTNESS = 2
 #: Confidence band required before ADOPT is possible at all.
 ADOPT_REQUIRED_LABEL = "High"
 
+#: Confidence band that can still reach PILOT given a verifiable primary path.
+PILOT_REQUIRED_LABEL = "Moderate"
+
+#: Auditable reasons a decision was stepped down from a higher action.
+DOWNGRADE_MISSING_DIRECT_PRIMARY = "missing_direct_primary"
+DOWNGRADE_CONFIDENCE_BAND = "confidence_band"
+DOWNGRADE_GATE_CRITICAL = "gate_critical"
+DOWNGRADE_UNRESOLVED_CONFLICT = "unresolved_conflict"
+
+#: Relation values that mark unresolved conflict rather than a clean vote.
+CONFLICT_RELATIONS = frozenset({"conflict", "mixed"})
+
 
 def primary_effect_categories(domain: str) -> tuple[str, ...]:
     """Categories that satisfy the ADOPT direct-evidence gate for a domain."""
@@ -72,25 +93,75 @@ def outcome_category(domain: str, value: str, primary: tuple[str, ...]) -> str |
         return None
 
 
-def decision_action(*, confidence_label: str, decisive_relations: dict[str, str],
-                    has_direct_primary_evidence: bool) -> str:
-    """Gate-enforced decision action (uppercase four-state).
-
-    REJECT requires usable direct opposition evidence (an independent Study
-    folded to oppose_adoption). Low/Insufficient can never yield ADOPT.
-    ADOPT additionally requires direct evidence on the domain's PRIMARY
-    outcome category: High + decisive support WITHOUT such evidence downgrades
-    to PILOT - task performance and procedural efficiency are not
-    decision-grade effects. Moderate + decisive support -> PILOT; otherwise
-    INSUFFICIENT_EVIDENCE.
+def decision_outcome(*, confidence_label: str, decisive_relations: dict[str, str],
+                     has_direct_primary_evidence: bool) -> dict[str, str | None]:
+    """Gate-enforced decision action plus auditable downgrade_reason.
+
+    Returns ``{"action": <uppercase four-state>, "downgrade_reason": <str|None>}``.
+
+    REJECT requires usable direct opposition evidence and no decisive support
+    (oppose-only). Mixed support+oppose or any conflict/conditional/mixed
+    relation is an unresolved conflict → INSUFFICIENT_EVIDENCE (never ADOPT,
+    never a bare REJECT). ADOPT requires High confidence, decisive support,
+    AND direct evidence on the domain's PRIMARY outcome category. PILOT is
+    the bounded-trial exit: High + support missing that direct primary
+    evidence (``missing_direct_primary``), or Moderate + support on a
+    verifiable primary-outcome path (``confidence_band``).
     """
     has_oppose = any(r == "oppose_adoption" for r in decisive_relations.values())
     has_support = any(r == "support_adoption" for r in decisive_relations.values())
+    has_conflict = any(r in CONFLICT_RELATIONS for r in decisive_relations.values())
+
+    # Unresolved conflict (mixed votes or explicit conflict tags) fails closed
+    # before any ADOPT/PILOT/REJECT claim.
+    if has_conflict or (has_oppose and has_support):
+        return {"action": "INSUFFICIENT_EVIDENCE",
+                "downgrade_reason": DOWNGRADE_UNRESOLVED_CONFLICT}
+
     if has_oppose:
-        return "REJECT"
+        return {"action": "REJECT", "downgrade_reason": None}
+
     if (confidence_label == ADOPT_REQUIRED_LABEL and has_support
             and has_direct_primary_evidence):
-        return "ADOPT"
-    if confidence_label in ("High", "Moderate") and has_support:
-        return "PILOT"
-    return "INSUFFICIENT_EVIDENCE"
+        return {"action": "ADOPT", "downgrade_reason": None}
+
+    if confidence_label == ADOPT_REQUIRED_LABEL and has_support:
+        return {"action": "PILOT", "downgrade_reason": DOWNGRADE_MISSING_DIRECT_PRIMARY}
+
+    # Bounded-trial exit: Moderate + support reaches PILOT even without a
+    # verified primary path (Gate C allows PILOT / INSUFFICIENT / REJECT).
+    if confidence_label == PILOT_REQUIRED_LABEL and has_support:
+        reason = (None if has_direct_primary_evidence
+                  else DOWNGRADE_MISSING_DIRECT_PRIMARY)
+        if has_direct_primary_evidence:
+            reason = DOWNGRADE_CONFIDENCE_BAND
+        return {"action": "PILOT", "downgrade_reason": reason}
+
+    # Low / 间接 (Moderate without a verifiable primary path) -> INSUFFICIENT.
+    if confidence_label in ("Low", "Insufficient"):
+        return {"action": "INSUFFICIENT_EVIDENCE",
+                "downgrade_reason": DOWNGRADE_CONFIDENCE_BAND}
+    if confidence_label == PILOT_REQUIRED_LABEL and has_support:
+        return {"action": "INSUFFICIENT_EVIDENCE",
+                "downgrade_reason": DOWNGRADE_MISSING_DIRECT_PRIMARY}
+    return {"action": "INSUFFICIENT_EVIDENCE", "downgrade_reason": None}
+
+
+def decision_action(*, confidence_label: str, decisive_relations: dict[str, str],
+                    has_direct_primary_evidence: bool) -> str:
+    """Gate-enforced decision action (uppercase four-state)."""
+    return decision_outcome(
+        confidence_label=confidence_label,
+        decisive_relations=decisive_relations,
+        has_direct_primary_evidence=has_direct_primary_evidence,
+    )["action"]
+
+
+def downgrade_reason(*, confidence_label: str, decisive_relations: dict[str, str],
+                     has_direct_primary_evidence: bool) -> str | None:
+    """Auditable reason the action was stepped down (or None when not downgraded)."""
+    return decision_outcome(
+        confidence_label=confidence_label,
+        decisive_relations=decisive_relations,
+        has_direct_primary_evidence=has_direct_primary_evidence,
+    )["downgrade_reason"]