feat(oracle): compare verified template evidence

2026-06-26 17:14:19 +08:00
parent e6710e1e2a
commit 5578ba6e29
3 changed files with 171 additions and 0 deletions
@@ -194,6 +194,15 @@ python3 scripts/oracle_evidence_validator.py \

 该验证器输出 `external_oracle_evidence_validation`，会检查 `evidence_packet` 必填元数据、`target_placeholders` 是否已填、是否覆盖 `target_fields`、是否包含外部 artifact，以及是否错误使用本仓库本地引擎输出。当前 draft 队列会保持 `valid_packets: 0` / `ready_for_calibration: 0`；只有状态为 `external_verified` 且证据完整的包才会进入可复核状态。

+证据包通过 validator 之后，再运行边界差异审计，比较本地引擎与外部 Dasha/Shadbala 目标值：
+
+```bash
+python3 scripts/oracle_boundary_audit.py \
+  --oracle-file references/oracle/dasha_shadbala_oracle_cases.json
+```
+
+审计报告中的 `template_comparisons` 会列出 external-verified template 的 Dasha 起点差异、Shadbala 七曜分量/总分差异，并继续保持 `production_tuning_recommended: false`，防止用单个样本调生产常数。
+
 `full-reading` 也会输出 `ai_prompt_pack`：这是给网页/app、skill 或后端 AI 代理使用的结构化 Prompt/RAG 上下文包。它不会硬编码断语，而是携带 D1/D9/Dasha/Shadbala/Ashtakavarga 的证据快照、推荐检索文档和边界提示，要求大模型基于计算证据交叉验证，避免单一配置下结论。

 ### Prerequisites
@@ -232,6 +232,99 @@ def _audit_template_case(case: dict[str, Any]) -> dict[str, Any]:
    }


+def _namespace_from_template(case: dict[str, Any]) -> argparse.Namespace:
+    birth = dict(case.get("birth", {}))
+    settings = case.get("settings", {})
+    if "node_mode" not in birth and settings.get("node_mode"):
+        birth["node_mode"] = settings["node_mode"]
+    return _namespace_from_birth(
+        birth,
+        ayanamsa=settings.get("ayanamsa", "lahiri"),
+        moon_lon=None,
+        nakshatra=None,
+        pada=None,
+        birthdate=None,
+        today=None,
+        years=120,
+        table=False,
+    )
+
+
+def _first_dasha_start(result: dict[str, Any]) -> str | None:
+    timeline = result.get("timeline")
+    if isinstance(timeline, list) and timeline:
+        return timeline[0].get("start")
+    return None
+
+
+def _date_delta_days(engine_date: str | None, target_date: str | None) -> int | None:
+    if not engine_date or not target_date:
+        return None
+    from datetime import date
+
+    try:
+        engine_parts = [int(part) for part in engine_date.split("-")[:3]]
+        target_parts = [int(part) for part in target_date.split("-")[:3]]
+        return (date(*engine_parts) - date(*target_parts)).days
+    except Exception:
+        return None
+
+
+def _template_shadbala_comparison(case: dict[str, Any]) -> dict[str, Any]:
+    target_components = case.get("target", {}).get("shadbala_components")
+    if not isinstance(target_components, dict):
+        return {"status": "missing_external_shadbala_components", "planets": {}}
+
+    result = engine.cmd_shadbala(_namespace_from_template(case))
+    if "error" in result:
+        return {"status": "engine_error", "error": result["error"], "planets": {}}
+
+    rows: dict[str, Any] = {}
+    for planet in SHADBALA_PLANETS:
+        external_row = target_components.get(planet, {})
+        engine_row = result.get("planets", {}).get(planet, {})
+        engine_components = _component_totals(engine_row) if engine_row else {}
+        external_total = external_row.get("total_rupa") if isinstance(external_row, dict) else None
+        engine_total = round(float(engine_row.get("total_rupas", 0.0)), 4) if engine_row else None
+        rows[planet] = {
+            "engine_total_rupa": engine_total,
+            "external_total_rupa": external_total,
+            "total_rupa_delta": (
+                round(engine_total - float(external_total), 4)
+                if engine_total is not None and isinstance(external_total, (int, float))
+                else None
+            ),
+            "engine_components": engine_components,
+            "external_components": external_row,
+        }
+    return {"status": "compared", "planets": rows}
+
+
+def _audit_external_verified_template_case(case: dict[str, Any]) -> dict[str, Any]:
+    target = case.get("target", {})
+    dasha_result = engine.cmd_dasha(_namespace_from_template(case))
+    engine_start = _first_dasha_start(dasha_result) if "error" not in dasha_result else None
+    target_start = target.get("vimshottari_start_date")
+    return {
+        "case_id": case.get("id") or case.get("case_id"),
+        "status": case.get("status"),
+        "source": case.get("source"),
+        "metadata": case.get("evidence_packet", {}).get("metadata", {}),
+        "dasha": {
+            "status": "compared" if engine_start and target_start else "missing_dasha_target",
+            "engine_start_date": engine_start,
+            "target_start_date": target_start,
+            "date_delta_days": _date_delta_days(engine_start, target_start),
+        },
+        "shadbala": _template_shadbala_comparison(case),
+        "calibration_decision": "do_not_tune_single_template",
+        "finding": (
+            "External-verified template rows are comparison evidence. Production constants require "
+            "a multi-source sample matrix and must not be tuned to a single packet."
+        ),
+    }
+
+
 def _status_counts(rows: list[dict[str, Any]]) -> dict[str, int]:
    counts: dict[str, int] = {}
    for row in rows:
@@ -242,6 +335,11 @@ def _status_counts(rows: list[dict[str, Any]]) -> dict[str, int]:

 def build_report(oracle: dict[str, Any]) -> dict[str, Any]:
    template_rows = [_audit_template_case(case) for case in oracle.get("template_cases", [])]
+    template_comparisons = [
+        _audit_external_verified_template_case(case)
+        for case in oracle.get("template_cases", [])
+        if case.get("status") in ORACLE_TEMPLATE_READY_STATUSES and not _missing_target_fields(case.get("target", {}))
+    ]
    dasha_rows = [_audit_dasha_case(case) for case in oracle.get("dasha_cases", [])]
    longitude_rows = [_audit_longitude_case(case) for case in oracle.get("longitude_cases", [])]
    shadbala_rows = [_audit_shadbala_case(case) for case in oracle.get("shadbala_cases", [])]
@@ -251,6 +349,7 @@ def build_report(oracle: dict[str, Any]) -> dict[str, Any]:
        "summary": {
            "template_cases": len(template_rows),
            "template_status_counts": _status_counts(template_rows),
+            "external_verified_template_cases": len(template_comparisons),
            "dasha_cases": len(dasha_rows),
            "longitude_cases": len(longitude_rows),
            "shadbala_cases": len(shadbala_rows),
@@ -262,6 +361,7 @@ def build_report(oracle: dict[str, Any]) -> dict[str, Any]:
            ],
        },
        "template_cases": template_rows,
+        "template_comparisons": template_comparisons,
        "dasha_cases": dasha_rows,
        "longitude_cases": longitude_rows,
        "shadbala_cases": shadbala_rows,
@@ -72,3 +72,65 @@ def test_oracle_boundary_audit_reports_dasha_and_shadbala_boundaries() -> None:
    assert template["status"] == "template_only"
    assert template["ready_for_calibration"] is False
    assert template["missing_target_fields"]
+
+
+def test_oracle_boundary_audit_compares_external_verified_template_rows(tmp_path: Path) -> None:
+    oracle = json.loads((ROOT / "references/oracle/dasha_shadbala_oracle_cases.json").read_text(encoding="utf-8"))
+    case = oracle["template_cases"][1]
+    case["status"] = "external_verified"
+    case["target"]["vimshottari_start_date"] = "1951-11-01"
+    case["target"]["shadbala_components"] = {
+        planet: {
+            "sthana": 1.0,
+            "dig": 2.0,
+            "kala": 3.0,
+            "chesta": 4.0,
+            "naisargika": 5.0,
+            "drik": 6.0,
+            "total_rupa": 21.0,
+        }
+        for planet in ["Sun", "Moon", "Mars", "Mercury", "Jupiter", "Venus", "Saturn"]
+    }
+    case["evidence_packet"] = {
+        "capture_id": "external_template_steve_jobs_dasha_lahiri",
+        "status": "external_verified",
+        "metadata": {
+            "tool_name": "JHora",
+            "tool_version_or_url": "manual-jhora-8.0",
+            "capture_date": "2026-06-26",
+            "source_artifact": "references/oracle/artifacts/steve_jobs_jhora_redacted.png",
+            "ayanamsa": "lahiri",
+            "node_mode": "true",
+            "timezone": "UTC-08:00",
+            "operator_note": "Typed from redacted external JHora screenshot.",
+        },
+    }
+    oracle_path = tmp_path / "oracle.json"
+    oracle_path.write_text(json.dumps(oracle, ensure_ascii=False), encoding="utf-8")
+
+    completed = subprocess.run(
+        [
+            sys.executable,
+            "scripts/oracle_boundary_audit.py",
+            "--oracle-file",
+            str(oracle_path),
+        ],
+        cwd=ROOT,
+        text=True,
+        capture_output=True,
+        timeout=60,
+        check=False,
+    )
+
+    assert completed.returncode == 0, completed.stderr or completed.stdout
+    report = json.loads(completed.stdout)
+    assert report["summary"]["external_verified_template_cases"] == 1
+    assert report["summary"]["production_tuning_recommended"] is False
+    comparison = report["template_comparisons"][0]
+    assert comparison["case_id"] == "template_steve_jobs_dasha_lahiri"
+    assert comparison["status"] == "external_verified"
+    assert comparison["dasha"]["target_start_date"] == "1951-11-01"
+    assert comparison["dasha"]["date_delta_days"] is not None
+    assert comparison["shadbala"]["planets"]["Sun"]["external_total_rupa"] == 21.0
+    assert comparison["shadbala"]["planets"]["Sun"]["total_rupa_delta"] is not None
+    assert comparison["calibration_decision"] == "do_not_tune_single_template"