File size: 3,385 Bytes
aa80965 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 | {
"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
} |