eduevidence 6.0.0 → 6.3.0

package/schemas/vNext/task-spec.schema.json

@@ -1 +1,156 @@
-{"$schema":"http://json-schema.org/draft-07/schema#","type":"object","required":["task_id","run_id","base_revision","stage","role","role_profile","objective","reason_for_delegation","evidence_axis","inputs","input_artifacts","allowed_capabilities","forbidden_actions","scope","budget","expected_outputs","output_contract","termination","execution_mode","independent","read_only","timeout_seconds","token_budget","metadata"],"properties":{"task_id":{"type":"string","minLength":1},"run_id":{"type":["string","null"]},"base_revision":{"type":["integer","null"],"minimum":0},"stage":{"enum":["frame","retrieve","extract","challenge","audit","adjudicate","applicability","intervene","evaluate"]},"role":{"type":"string"},"role_profile":{"type":["string","null"]},"objective":{"type":"string","minLength":1},"reason_for_delegation":{"type":["string","null"]},"evidence_axis":{"type":"string","minLength":1},"inputs":{"type":"array","items":{"type":"string"}},"input_artifacts":{"type":"array","items":{"type":"string"}},"allowed_capabilities":{"type":"array","items":{"type":"string"}},"forbidden_actions":{"type":"array","items":{"type":"string"}},"scope":{"type":"object"},"budget":{"type":"object"},"expected_outputs":{"type":"array","items":{"type":"string"}},"output_contract":{"type":"object"},"termination":{"type":"object"},"execution_mode":{"enum":["local","delegated"]},"independent":{"type":"boolean"},"read_only":{"type":"boolean"},"timeout_seconds":{"type":"integer","minimum":1},"token_budget":{"type":["integer","null"],"minimum":1},"metadata":{"type":"object"}},"additionalProperties":false}
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "title": "TaskSpec",
+  "description": "A dispatchable unit of work for one role (engine/orchestration.py TaskSpec). Fixes the stage, role, evidence axis, run and base revision, the granted capabilities, forbidden actions, budget, expected outputs and the output contract. Delegated specs are staging-only: canonical-state artifacts are refused by construction.",
+  "type": "object",
+  "required": [
+    "task_id",
+    "run_id",
+    "base_revision",
+    "stage",
+    "role",
+    "role_profile",
+    "objective",
+    "reason_for_delegation",
+    "evidence_axis",
+    "inputs",
+    "input_artifacts",
+    "allowed_capabilities",
+    "forbidden_actions",
+    "scope",
+    "budget",
+    "expected_outputs",
+    "output_contract",
+    "termination",
+    "execution_mode",
+    "independent",
+    "read_only",
+    "timeout_seconds",
+    "token_budget",
+    "metadata"
+  ],
+  "properties": {
+    "task_id": {
+      "type": "string",
+      "minLength": 1
+    },
+    "run_id": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "base_revision": {
+      "type": [
+        "integer",
+        "null"
+      ],
+      "minimum": 0
+    },
+    "stage": {
+      "enum": [
+        "frame",
+        "retrieve",
+        "extract",
+        "challenge",
+        "audit",
+        "adjudicate",
+        "applicability",
+        "intervene",
+        "evaluate"
+      ]
+    },
+    "role": {
+      "type": "string"
+    },
+    "role_profile": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "objective": {
+      "type": "string",
+      "minLength": 1
+    },
+    "reason_for_delegation": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "evidence_axis": {
+      "type": "string",
+      "minLength": 1
+    },
+    "inputs": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "input_artifacts": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "allowed_capabilities": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "forbidden_actions": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "scope": {
+      "type": "object"
+    },
+    "budget": {
+      "type": "object"
+    },
+    "expected_outputs": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "output_contract": {
+      "type": "object"
+    },
+    "termination": {
+      "type": "object"
+    },
+    "execution_mode": {
+      "enum": [
+        "local",
+        "delegated"
+      ]
+    },
+    "independent": {
+      "type": "boolean"
+    },
+    "read_only": {
+      "type": "boolean"
+    },
+    "timeout_seconds": {
+      "type": "integer",
+      "minimum": 1
+    },
+    "token_budget": {
+      "type": [
+        "integer",
+        "null"
+      ],
+      "minimum": 1
+    },
+    "metadata": {
+      "type": "object"
+    }
+  },
+  "additionalProperties": false
+}

package/schemas/vNext/worker-result.schema.json

@@ -1 +1,60 @@
-{"$schema":"http://json-schema.org/draft-07/schema#","type":"object","required":["task_id","status","staging_artifacts","validated","validation_issues","metrics","summary"],"properties":{"task_id":{"type":"string","minLength":1},"status":{"enum":["completed","failed","blocked"]},"staging_artifacts":{"type":"array","items":{"type":"object","required":["artifact_type"],"properties":{"artifact_type":{"type":"string","minLength":1}},"additionalProperties":true}},"validated":{"type":"boolean"},"validation_issues":{"type":"array","items":{"type":"string"}},"metrics":{"type":"object"},"summary":{"type":"string"}},"additionalProperties":false}
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "title": "WorkerResult",
+  "description": "The lead process's verdict on one worker's submission (engine/worker_result.py). validate_worker_output() reconstructs this from untrusted worker output: any worker-supplied validated flag is ignored, canonical-state artifacts are rejected, and only results that pass here may enter a Judge context.",
+  "type": "object",
+  "required": [
+    "task_id",
+    "status",
+    "staging_artifacts",
+    "validated",
+    "validation_issues",
+    "metrics",
+    "summary"
+  ],
+  "properties": {
+    "task_id": {
+      "type": "string",
+      "minLength": 1
+    },
+    "status": {
+      "enum": [
+        "completed",
+        "failed",
+        "blocked"
+      ]
+    },
+    "staging_artifacts": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "required": [
+          "artifact_type"
+        ],
+        "properties": {
+          "artifact_type": {
+            "type": "string",
+            "minLength": 1
+          }
+        },
+        "additionalProperties": true
+      }
+    },
+    "validated": {
+      "type": "boolean"
+    },
+    "validation_issues": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "metrics": {
+      "type": "object"
+    },
+    "summary": {
+      "type": "string"
+    }
+  },
+  "additionalProperties": false
+}

package/schemas/verdict.schema.json

@@ -5,52 +5,188 @@
   "description": "Output of the Evidence Tribunal: what the evidence supports, cannot support, and the recommended action. Decision is one of ADOPT / PILOT / REJECT / INSUFFICIENT EVIDENCE (expressed here as adopt|pilot|reject|insufficient_evidence). confidence 由 scripts/compute_confidence.py 确定性计算并覆盖模型值；confidence_score 是规则化指数（0-1），不是概率。扩展字段一律放在 extensions 内。",
   "type": "object",
   "additionalProperties": false,
-  "required": ["decision_question", "recommended_action", "confidence"],
+  "required": [
+    "decision_question",
+    "recommended_action",
+    "confidence"
+  ],
   "properties": {
-    "decision_question": { "type": "string" },
-    "target_population": { "type": "string" },
-    "target_context": { "type": "string" },
-    "supported_claims": { "type": "array", "items": { "type": "string" } },
-    "uncertain_claims": { "type": "array", "items": { "type": "string" } },
-    "contradicted_claims": { "type": "array", "items": { "type": "string" } },
-    "reason_for_disagreement": { "type": "string" },
-    "methodology_summary": { "type": "string" },
-    "outcome_specific_findings": { "type": "object", "additionalProperties": true },
-    "short_term_effect": { "type": ["string", "null"] },
-    "long_term_effect": { "type": ["string", "null"] },
-    "transfer_effect": { "type": ["string", "null"] },
-    "risk_effect": { "type": ["string", "null"] },
-    "applicability": { "type": "object", "additionalProperties": true },
-    "confidence": { "type": "string", "enum": ["High", "Moderate", "Low", "Insufficient"] },
+    "decision_question": {
+      "type": "string"
+    },
+    "target_population": {
+      "type": "string"
+    },
+    "target_context": {
+      "type": "string"
+    },
+    "supported_claims": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "uncertain_claims": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "contradicted_claims": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "reason_for_disagreement": {
+      "type": "string"
+    },
+    "methodology_summary": {
+      "type": "string"
+    },
+    "outcome_specific_findings": {
+      "type": "object",
+      "additionalProperties": true
+    },
+    "short_term_effect": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "long_term_effect": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "transfer_effect": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "risk_effect": {
+      "type": [
+        "string",
+        "null"
+      ]
+    },
+    "applicability": {
+      "type": "object",
+      "additionalProperties": true
+    },
+    "confidence": {
+      "type": "string",
+      "enum": [
+        "High",
+        "Moderate",
+        "Low",
+        "Insufficient"
+      ]
+    },
     "confidence_score": {
-      "type": ["number", "null"],
+      "type": [
+        "number",
+        "null"
+      ],
       "minimum": 0,
       "maximum": 1,
       "description": "规则化置信度指数（0-1），由 compute_confidence.py 覆盖模型值。不是概率，禁止宣传为百分比。"
     },
-    "confidence_policy_version": { "type": "string", "description": "确定性置信度策略版本号（如 2026-08-12.v1）。" },
-    "independent_studies": { "type": ["integer", "null"], "minimum": 0, "description": "独立研究数（按 study_id/source_id 去重）。" },
-    "independent_samples": { "type": ["integer", "null"], "minimum": 0, "description": "独立样本数（按 sample_id 去重）。" },
+    "confidence_policy_version": {
+      "type": "string",
+      "description": "确定性置信度策略版本号（如 2026-08-12.v1）。"
+    },
+    "independent_studies": {
+      "type": [
+        "integer",
+        "null"
+      ],
+      "minimum": 0,
+      "description": "独立研究数（按 study_id/source_id 去重）。"
+    },
+    "independent_samples": {
+      "type": [
+        "integer",
+        "null"
+      ],
+      "minimum": 0,
+      "description": "独立样本数（按 sample_id 去重）。"
+    },
     "confidence_breakdown": {
       "type": "object",
       "additionalProperties": true,
       "description": "Rule-based components: evidence_quality, consistency, directness, evidence_count, independent_studies, independent_samples, conflict_penalty, unsupported_penalty."
     },
     "raw_model_confidence": {
-      "type": ["string", "null"],
+      "type": [
+        "string",
+        "null"
+      ],
       "description": "模型原始 confidence 输出（被确定性值覆盖前的值，仅供审计比对）。"
     },
-    "raw_model_confidence_breakdown": { "type": "object", "additionalProperties": true },
-    "what_can_be_claimed": { "type": "array", "items": { "type": "string" } },
-    "what_cannot_be_claimed": { "type": "array", "items": { "type": "string" } },
-    "missing_evidence": { "type": "array", "items": { "type": "string" } },
-    "recommended_action": { "type": "string", "enum": ["adopt", "pilot", "reject", "insufficient_evidence"] },
-    "decision_rationale": { "type": "string" },
-    "exceeds_evidence_boundary": { "type": "array", "items": { "type": "string" }, "description": "Conclusions that currently go beyond the evidence boundary." },
+    "raw_model_confidence_breakdown": {
+      "type": "object",
+      "additionalProperties": true
+    },
+    "what_can_be_claimed": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "what_cannot_be_claimed": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "missing_evidence": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      }
+    },
+    "recommended_action": {
+      "type": "string",
+      "enum": [
+        "adopt",
+        "pilot",
+        "reject",
+        "insufficient_evidence"
+      ]
+    },
+    "decision_rationale": {
+      "type": "string"
+    },
+    "exceeds_evidence_boundary": {
+      "type": "array",
+      "items": {
+        "type": "string"
+      },
+      "description": "Conclusions that currently go beyond the evidence boundary."
+    },
     "extensions": {
       "type": "object",
       "description": "结构化扩展字段的统一容器（P1-01）。未列入本 schema 的字段必须放在这里，禁止在顶层新增属性。",
       "additionalProperties": true
+    },
+    "strongest_support": {
+      "type": "string",
+      "description": "The single strongest conclusion the evidence supports, as a complete reader-facing sentence (<=60 chars zh / ~15 words en). Written by the adjudicator, not assembled by the renderer."
+    },
+    "key_uncertainty": {
+      "type": "string",
+      "description": "The decision-relevant uncertainty or counter-evidence, as a complete reader-facing sentence (<=70 chars zh / ~18 words en)."
+    },
+    "main_risk": {
+      "type": "string",
+      "description": "The principal risk of acting, as a complete reader-facing sentence (<=60 chars zh / ~15 words en)."
+    },
+    "next_action": {
+      "type": "string",
+      "description": "The recommended next step, as a complete reader-facing sentence (<=80 chars zh / ~20 words en)."
     }
   }
 }

package/scripts/build_evidence_library.py

@@ -17,9 +17,15 @@ schemas/v4/evidence-library.schema.json using the repo's zero-dependency
 validator (scripts/validate_schema.py).
 
 Direction semantics (adoption-relevant, conservative):
-  support    -> evidence favors adopting the intervention  (=> pilot)
-  contradict -> evidence opposes adopting the intervention (=> reject)
+  support    -> evidence favors adopting the intervention
+                (offline screening cap => pilot; never adopt)
+  contradict -> evidence opposes adopting the intervention
+                (oppose-only => reject; mixed with support is an unresolved
+                conflict and must fall to INSUFFICIENT_EVIDENCE via
+                engine.decision_policy.decision_outcome)
   neutral    -> inconclusive
+The library never emits adopt: full ADOPT is decided only by
+engine.decision_policy.decision_outcome (High + support + direct primary).
 For gold units the coarse rule is: if a question's expected decision range is
 purely reject-oriented ("reject" present and "pilot" absent), its
 key_claims/key_supporting_sources are harmful evidence => contradict, and its
@@ -266,10 +272,14 @@ def build(generated_at: str | None = None) -> tuple[dict[str, Any], int]:
             "key_claims/key_supporting_sources/known_contradictions/correct_outcome_types）"
             "+ 3 个示例工作流 evidence.jsonl（ai-coding-assistant / ai-tutor / ai-writing-assistant）"
             "抽取生成；按 (source_id, outcome_token, claim_text) 去重合并。"
-            "direction 语义为采纳方向：support=支持采纳（初步裁决=>pilot），"
-            "contradict=反对采纳（=>reject），neutral=中性；金标准条目按 expected_decision_range "
+            "direction 语义为采纳方向：support=支持采纳（离线初筛上限 => pilot，永不输出 adopt），"
+            "contradict=反对采纳（oppose-only 时 => reject；与 support 并存属未决冲突，应交 "
+            "engine.decision_policy.decision_outcome 判为 INSUFFICIENT_EVIDENCE），neutral=中性；"
+            "金标准条目按 expected_decision_range "
             "粗粒度映射方向（纯 reject 问题反向映射），conflict 与混合方向问题的单条断言方向可能不精确。"
-            "仅用于离线初步裁决（preliminary，保守），从不直接给出 adopt。"
+            "仅用于离线初步裁决（preliminary，保守），从不直接给出 adopt；"
+            "完整 ADOPT 只能由 engine.decision_policy.decision_outcome 给出"
+            "（High + support + 主要结果 directness 2）。"
         ),
     }
     _validate(library)

package/scripts/build_report_variants.py

@@ -40,7 +40,9 @@ def bake(examples: Path, *, force: bool = False) -> list[dict]:
         if not force and manifest.is_file():
             try:
                 prior = json.loads(manifest.read_text(encoding='utf-8'))
-                valid = prior.get('cache_key') == cache_key and all(
+                main = prior.get('main_report') or {}
+                main_ok = bool(main.get('file')) and (directory / main['file']).is_file() and                     hashlib.sha256((directory / main['file']).read_bytes()).hexdigest() == main.get('sha256')
+                valid = prior.get('cache_key') == cache_key and main_ok and all(
                     (out_dir / record['file']).is_file() and hashlib.sha256((out_dir / record['file']).read_bytes()).hexdigest() == record['sha256']
                     for record in prior.get('reports', [])) and len(prior.get('reports', [])) == len(THEMES)
                 if valid:
@@ -64,8 +66,22 @@ def bake(examples: Path, *, force: bool = False) -> list[dict]:
                 raise RuntimeError(f'{directory.name}/{theme}: renderer rejected input\n{completed.stdout}\n{completed.stderr}')
             os.replace(temporary, target)
             records.append({'theme': theme, 'file': target.name, 'sha256': hashlib.sha256(target.read_bytes()).hexdigest()})
+        # Also refresh the pack-root report. This used to write only the themed
+        # variants, so examples/*/EduEvidence_Report.html kept whatever bytes it
+        # was first rendered with - the packaged example shipped a report the
+        # current renderer would not produce.
+        from_default = next((r for r in records if r['theme'] == 'claude'), records[0])
+        main_target = directory / 'EduEvidence_Report.html'
+        main_temp = directory / '.EduEvidence_Report.pending.html'
+        main_temp.write_bytes((out_dir / from_default['file']).read_bytes())
+        os.replace(main_temp, main_target)
+        main_sha = hashlib.sha256(main_target.read_bytes()).hexdigest()
+
         value = {'schema_version': 1, 'project': directory.name, 'cache_key': cache_key,
-                 'result_sha256': result_hash, 'renderer_sha256': engine_hash, 'reports': records}
+                 'result_sha256': result_hash, 'renderer_sha256': engine_hash,
+                 'main_report': {'file': main_target.name, 'sha256': main_sha,
+                                 'theme': from_default['theme']},
+                 'reports': records}
         manifest.write_text(json.dumps(value, indent=2) + '\n', encoding='utf-8')
         reports.append(value)
         print(f'{directory.name}: {len(records)} verified report variants')

package/scripts/build_result.py

@@ -30,14 +30,14 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
 from evidence_semantics import effect_direction
 from engine.versions import ENGINE_VERSION
 
-OUTCOME_ORDER = [
-    "knowledge_gain", "concept_understanding", "retention", "transfer",
-    "independent_problem_solving", "completion_time", "accuracy",
-    "code_quality", "assignment_score", "engagement", "motivation",
-    "cognitive_load", "help_seeking", "metacognition", "ai_dependency",
-    "over_reliance", "reduced_effort", "reduced_transfer",
-    "academic_integrity_risk", "false_confidence",
-]
+def _outcome_order() -> list[str]:
+    """Registered outcome tokens in registry order (was a hard-coded list)."""
+    from engine.taxonomy import all_tokens_ordered
+
+    return list(all_tokens_ordered())
+
+
+OUTCOME_ORDER = _outcome_order()
 
 
 def _load_json(path: Path) -> dict[str, Any] | None:
@@ -221,12 +221,74 @@ def build_claims(evidence: list[dict[str, Any]]) -> list[dict[str, Any]]:
     return list(claims.values())
 
 
+def _applicability(pack_dir: Path, verdict: dict) -> dict:
+    """Stage-7 applicability assessment, falling back to the verdict boundary.
+
+    applicability.json is the dedicated deliverable of the Applicability stage.
+    It used to be written and then ignored, because the renderer only looked at
+    the verdict; this is where it re-enters the result.
+    """
+    import json as _json
+
+    path = pack_dir / "applicability.json"
+    if path.is_file():
+        try:
+            data = _json.loads(path.read_text(encoding="utf-8"))
+        except (OSError, _json.JSONDecodeError):
+            data = None
+        if isinstance(data, dict) and data and data.get("status") != "NOT_CAPTURED":
+            return data
+    value = verdict.get("applicability") if isinstance(verdict, dict) else None
+    return value if isinstance(value, dict) else {}
+
+def _derive_study_audits(methodology: list[dict], evidence: list[dict]) -> list[dict]:
+    """Per-study audit rows derived from the audits and evidence present.
+
+    Each row names the study and reports the audit verdict that covers it,
+    so the per-study axis the schema advertises actually exists downstream.
+    Rows are only emitted for studies the audits or evidence actually name.
+    """
+    by_study: dict[str, dict] = {}
+    for audit in methodology:
+        if not isinstance(audit, dict):
+            continue
+        target = audit.get("target") or "overall"
+        if target == "overall":
+            # The aggregate audit is not a study row; label it as the
+            # body-of-evidence review so it cannot be mistaken for one.
+            target = "body_of_evidence"
+        entry = by_study.setdefault(target, {
+            "study_id": target,
+            "verdict": audit.get("verdict"),
+            "audit_items": audit.get("audit_items") or {},
+            "limitations": list(audit.get("limitations") or []),
+            "task_vs_learning_guard": audit.get("task_vs_learning_guard"),
+        })
+        entry.setdefault("evidence_ids", [])
+    known = {e.get("study_id") for e in evidence if e.get("study_id")}
+    for study_id in sorted(known):
+        by_study.setdefault(study_id, {
+            "study_id": study_id,
+            "verdict": None,
+            "audit_items": {},
+            "limitations": [],
+            "task_vs_learning_guard": None,
+            "evidence_ids": [e.get("evidence_id") for e in evidence
+                             if e.get("study_id") == study_id],
+        })
+    return list(by_study.values())
+
+
 def build_result(pack_dir: Path, *, mode: str = "platform_native") -> dict[str, Any]:
     frame = _load_json(pack_dir / "frame.json") or {}
     evidence = _load_jsonl(pack_dir / "evidence.jsonl")
     # methodology.json is a single MethodologyAudit object (or a JSONL list)
     methodology_single = _load_json(pack_dir / "methodology.json")
     methodology = [methodology_single] if methodology_single else _load_jsonl(pack_dir / "methodology.jsonl")
+    # Per-study audits: the contract advertised them but nothing produced
+    # them, so a multi-study review silently shipped a single audit object.
+    # Derive the per-study axis from the audits actually present.
+    study_audits = _derive_study_audits(methodology, evidence)
     verdict = _load_json(pack_dir / "verdict.json") or {}
     intervention = _load_json(pack_dir / "intervention.json") or {}
     evaluation = _load_json(pack_dir / "evaluation.json") or {}
@@ -268,9 +330,12 @@ def build_result(pack_dir: Path, *, mode: str = "platform_native") -> dict[str,
         "sources": sources,
         "evidence": evidence,
         "methodology_reviews": methodology,
+        "study_audits": study_audits,
         "conflicts": [{"reason_for_disagreement": verdict.get("reason_for_disagreement", "")}]
         if verdict.get("reason_for_disagreement") else [],
-        "applicability": verdict.get("applicability", {}),
+        # Prefer the dedicated stage-7 assessment; fall back to the verdict-embedded
+        # boundary when the run carries no separate applicability.json.
+        "applicability": _applicability(pack_dir, verdict),
         "intervention": intervention,
         "evaluation": evaluation,
         "benchmark": {},