Spaces:
Paused
Paused
File size: 7,062 Bytes
5d10875 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 | """AgenticSafetyScorecard — leadership-facing retrospective scorecard.
Every field is a ScorecardField carrying an OBSERVED / PREDICTED /
HYPOTHETICAL / REQUIRED_EVIDENCE tag plus an evidence reference, so a reader
can always tell a measured number from a model's guess (an improvement over a
plain key/value scorecard).
"""
from __future__ import annotations
from typing import Dict, Optional
from sensorflow.agentic import store as store_mod
from sensorflow.agentic.models import (AgenticSafetyScorecard,
ConcentrationAnalysis, FailureEvent,
ScorecardField, StatisticalAssessment,
new_id)
def _f(value, tag: str, ref: str = "") -> ScorecardField:
return ScorecardField(value=value, tag=tag, evidence_ref=ref)
def build_scorecard(failure: FailureEvent,
statistical: Optional[StatisticalAssessment],
concentration: Optional[ConcentrationAnalysis],
policy_evaluation: Optional[Dict],
agent_results: Dict) -> AgenticSafetyScorecard:
basis = failure.detection_basis
if statistical is not None:
freq = _f({"rate": statistical.candidate.rate,
"wilson_ci": statistical.candidate.wilson_ci,
"events": statistical.candidate.events,
"denominator": statistical.candidate.denominator,
"baseline_rate": statistical.baseline.rate,
"significant": statistical.significant},
"OBSERVED", "statistical_assessment")
else:
freq = _f("statistical assessment not yet run", "REQUIRED_EVIDENCE",
"FAILURE_ANALYSIS stage")
if concentration is not None:
conc = _f({"determination": concentration.determination,
"concentrated_dimensions": concentration.concentrated_dimensions,
"top_strata": [s.model_dump() for s in concentration.strata
if s.relative_risk and s.relative_risk >= 3.0][:3]},
"OBSERVED", "concentration_analysis")
else:
conc = _f("concentration analysis not yet run", "REQUIRED_EVIDENCE",
"FAILURE_ANALYSIS stage")
# "escalated" is a successful analysis that requests human review (e.g.
# unsafe replay behavior) — only "failed" means the evidence is missing.
safety = agent_results.get("safety_impact")
if safety is not None and safety.status in ("ok", "escalated"):
chain = safety.output.get("chain", [])
if safety.output.get("behavioral_evidence") == "observed":
downstream = _f(safety.output, "OBSERVED", "safety_impact_agent (replay)")
else:
downstream = _f({"assessment": "safety assessment UNCERTAIN — no "
"behavioral evidence",
"chain": chain},
"REQUIRED_EVIDENCE", "safety_impact_agent")
else:
downstream = _f("safety impact not assessed", "REQUIRED_EVIDENCE",
"safety_impact_agent")
mining = agent_results.get("scenario_mining")
novelty_val = (mining.output.get("novelty", "unknown")
if mining is not None and mining.status in ("ok", "escalated")
else "unknown")
pol = policy_evaluation or {}
severity_field = _f(
{"severity": pol.get("severity"),
"criteria": (pol.get("severity_assignment") or {}).get("taxonomy_description")},
"OBSERVED" if pol else "REQUIRED_EVIDENCE",
"policy_engine.assign_severity")
resid = None
for row in pol.get("expected_loss_table", []):
if row["option"] == pol.get("recommended_option"):
resid = row
break
residual = (_f({"option": resid["option"],
"residual_failure_rate": resid["residual_failure_rate"],
"expected_loss": resid["expected_loss"],
"feasible": resid["feasible"]},
"PREDICTED", "policy_engine.expected_loss_table")
if resid else _f("no policy evaluation yet", "REQUIRED_EVIDENCE",
"LAUNCH_DECISION stage"))
ec = "INSUFFICIENT_EVIDENCE"
if statistical is not None:
if statistical.significant and not statistical.small_sample_flags:
ec = "CONFIRMED"
elif statistical.significant:
ec = "LIKELY"
elif statistical.candidate.events > 0:
ec = "POSSIBLE"
card = AgenticSafetyScorecard(
scorecard_id=new_id("card"),
failure_id=failure.failure_id,
title=failure.title,
failure_summary=_f(
{"kind": failure.kind, "gt_class": failure.gt_class,
"predicted_class": failure.predicted_class,
"candidate_model": failure.candidate_model,
"baseline_model": failure.baseline_model,
"instances_captured": len(failure.instances)},
"OBSERVED", "failure_detection_agent"),
frequency=freq,
exposure=_f({"denominator": basis.denominator,
"population": failure.population_id,
"construction_zone_exposure_share":
next((s.exposure_share for s in
(concentration.strata if concentration else [])
if s.stratum == "construction"), None)},
"OBSERVED", "rate population telemetry"),
severity=severity_field,
confidence=_f(ec, "OBSERVED", "statistical_assessment quality rules"),
novelty=_f(novelty_val,
"OBSERVED" if novelty_val != "unknown" else "REQUIRED_EVIDENCE",
"scenario_mining_agent (deterministic retrieval)"),
concentration=conc,
downstream_impact=downstream,
mitigations=_f(
{"validated_mitigation": bool((pol.get("input") or {}).get("mitigation_validated")),
"description": (pol.get("input") or {}).get("mitigation_description") or "none"},
"OBSERVED", "policy input"),
residual_risk=residual,
evidence_quality=ec,
policy_outcome=pol.get("outcome"),
recommended_option=pol.get("recommended_option"),
policy_version=pol.get("policy_version", ""),
notes=[
"Field tags: OBSERVED = measured on the synthetic campaign; "
"PREDICTED = deterministic model output (e.g. expected loss); "
"HYPOTHETICAL = agent hypothesis; REQUIRED_EVIDENCE = missing.",
],
)
store_mod.write_json(card, "scorecards", f"{card.scorecard_id}.json")
store_mod.audit("scorecard_generated", failure.failure_id, "scorecard",
detail=f"scorecard {card.scorecard_id}")
return card
def load_scorecard(scorecard_id: str) -> Optional[Dict]:
return store_mod.read_json("scorecards", f"{scorecard_id}.json")
|