eval_results.json

{
  "protocol": {
    "honesty": "100% synthetic bilingual Permit-to-Work corpus. No real permits, incidents, PII, or named facilities. Zones/equipment tags are fictional and partly reused from the PetroSafe RAG demo unit for portfolio consistency. About 20% of permits deliberately mismatch free text vs structured flags to simulate documentation gaps.",
    "n_permits": 620,
    "n_train": 495,
    "n_test": 125,
    "text_flag_mismatch_rate": 0.19193548387096773,
    "rule_noncompliant_fraction": 0.3419354838709677,
    "n_simops_pairs": 260
  },
  "risk_classifier": {
    "accuracy": 0.888,
    "macro_f1": 0.8660594730892965,
    "confusion_matrix": {
      "labels": [
        "low",
        "medium",
        "high",
        "critical"
      ],
      "matrix": [
        [
          30,
          0,
          0,
          0
        ],
        [
          1,
          47,
          4,
          4
        ],
        [
          1,
          1,
          11,
          1
        ],
        [
          0,
          1,
          1,
          23
        ]
      ]
    },
    "per_class": {
      "low": {
        "precision": 0.9375,
        "recall": 1.0,
        "n": 30
      },
      "medium": {
        "precision": 0.9591836734693877,
        "recall": 0.8392857142857143,
        "n": 56
      },
      "high": {
        "precision": 0.6875,
        "recall": 0.7857142857142857,
        "n": 14
      },
      "critical": {
        "precision": 0.8214285714285714,
        "recall": 0.92,
        "n": 25
      }
    },
    "critical_recall": 0.92
  },
  "worktype_classifier": {
    "accuracy": 1.0,
    "macro_f1": 1.0
  },
  "simops_conflict_detector": {
    "n_pairs": 260,
    "n_conflict": 130,
    "n_no_conflict": 130,
    "conflict_rate": 0.5,
    "rule_breakdown": {
      "SIMOPS-01": 26,
      "SIMOPS-02": 26,
      "SIMOPS-03": 26,
      "SIMOPS-04": 26,
      "SIMOPS-05": 26
    },
    "note": "SIMOPS conflict checking is a deterministic rule engine (src/permitguard/rules.py), not a learned model -- there is no separate ground truth to score it against, so precision/recall would be a circular 100%/100%. The numbers above (pair balance + which of the 5 rules fired) demonstrate the engine covers a realistic, balanced mix of conflict and non-conflict scenarios, including 'mitigated near-miss' cases."
  },
  "published_baselines": [
    {
      "method": "ContractGuard hybrid TF-IDF + Logistic Regression (Aria AI sibling product)",
      "finding": "Site claim: 92% clause-classification accuracy and 100% recall on high-risk clauses, on a synthetic bilingual contract-clause corpus.",
      "citation": "Aria AI ContractGuard product page / Hugging Face Space contractguard-clause-analyzer. https://aria-ai.ir"
    },
    {
      "method": "NLP hazard identification from construction/safety text (public literature)",
      "finding": "Text classifiers are used to flag missing PPE/hazard statements in permits and JSAs; reported performance is corpus-specific and not a universal PTW KPI.",
      "citation": "Goh, Y.M. & Ubeynarayana, C.U. (2017). Construction accident narrative classification: An automatic text mining approach. Safety Science 99:70-80. https://doi.org/10.1016/j.ssci.2017.04.001"
    }
  ],
  "pipeline_seconds": 2.7
}