Spaces:
Paused
Paused
File size: 9,339 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 | """Worked example: pedestrian -> construction cone at ~0.01%.
Runs the full five-layer pipeline on the seeded synthetic population,
through evidence aggregation, analysis, the deterministic launch decision, a
recorded (clearly demo-labeled) human review and the learning flywheel, and
assembles a walkthrough where EVERY value is labeled:
OBSERVED measured on the seeded synthetic run
HYPOTHETICAL an agent hypothesis, unverified
REQUIRED-EVIDENCE evidence that does not exist and would be required
The walkthrough is deterministic for a fixed seed.
"""
from __future__ import annotations
from typing import Dict, Optional
from sensorflow.agentic import METHODOLOGY_PROVENANCE
from sensorflow.agentic import data as data_mod
from sensorflow.agentic import flywheel as flywheel_mod
from sensorflow.agentic import pipeline as pipeline_mod
from sensorflow.agentic import review as review_mod
from sensorflow.agentic import store as store_mod
WORKED_EXAMPLE_SEED = data_mod.DEFAULT_SEED
DEMO_REVIEWER = "safety-reviewer-demo"
def _find_ped_cone_failure() -> Optional[Dict]:
for doc in pipeline_mod.list_failures():
if (doc.get("gt_class") == "pedestrian"
and doc.get("predicted_class") == "construction_cone"):
return doc
return None
def run_worked_example(seed: int = WORKED_EXAMPLE_SEED,
use_llm: bool = False) -> Dict:
# Layer 1 — detection (idempotent: reuse an existing detection if present)
existing = _find_ped_cone_failure()
if existing is None:
pipeline_mod.detect_failures(seed=seed, use_llm=use_llm)
existing = _find_ped_cone_failure()
if existing is None:
raise RuntimeError("worked example: pedestrian->cone pattern not "
"detected on the seeded population")
fid = existing["failure_id"]
failure = pipeline_mod.get_failure(fid)
state = pipeline_mod.get_state(fid)
# Layers 2-4
for stage in ("EVIDENCE_AGGREGATION", "FAILURE_ANALYSIS", "LAUNCH_DECISION"):
if state.stage_record(stage).status == "pending":
state = pipeline_mod.run_stage(fid, stage, seed=seed, use_llm=use_llm)
failure = pipeline_mod.get_failure(fid)
evaluation = state.policy_evaluation or {}
# Human review (demo decision, clearly labeled as such)
if not failure.validated:
review_mod.record_decision(
failure, reviewer=DEMO_REVIEWER, decision="confirm_failure",
rationale=("DEMO WALKTHROUGH DECISION: replay evidence shows "
"unsafe delayed braking; statistics significant; "
"confirming the failure as real for the worked example"),
evidence_reviewed=["evidence_graph", "statistical_assessment",
"safety_replays", "concentration_analysis"],
policy_version=evaluation.get("policy_version", ""))
failure.validated = True
pipeline_mod.save_failure(failure)
store_mod.audit("failure_validated", fid, DEMO_REVIEWER,
detail="confirm_failure (worked-example demo reviewer)")
# Layer 5 — flywheel
if state.stage_record("LEARNING_FLYWHEEL").status in ("pending", "blocked"):
state.stage_record("LEARNING_FLYWHEEL").status = "pending"
pipeline_mod.save_state(state)
pipeline_mod.run_stage(fid, "LEARNING_FLYWHEEL", seed=seed,
use_llm=use_llm)
state = pipeline_mod.get_state(fid)
regression = flywheel_mod.regression_evaluate(seed=seed)
# ---- assemble the labeled walkthrough ------------------------------
stat = state.statistical
conc = state.concentration
safety = state.agent_results.get("safety_impact")
fusion = state.agent_results.get("sensor_fusion_verification")
mining = state.agent_results.get("scenario_mining")
vlm = state.agent_results.get("vlm_scene_analysis")
def L(label: str, value) -> Dict:
return {"label": label, "value": value}
walkthrough = {
"methodology_provenance": METHODOLOGY_PROVENANCE,
"seed": seed,
"deterministic": True,
"failure_id": fid,
"title": failure.title,
"layers": {
"1_failure_detection": {
"candidate_events": L("OBSERVED",
failure.detection_basis.candidate_events),
"baseline_events": L("OBSERVED",
failure.detection_basis.baseline_events),
"denominator": L("OBSERVED", failure.detection_basis.denominator),
"observed_rate": L("OBSERVED",
failure.detection_basis.candidate_rate),
"headline": L("OBSERVED",
f"~{failure.detection_basis.candidate_rate:.4%} "
"pedestrian->construction_cone rate on the "
"candidate model"),
"scene_instances_captured": L("OBSERVED", len(failure.instances)),
},
"2_evidence_aggregation": {
"evidence_graph_nodes": L(
"OBSERVED",
[{"type": n["node_type"], "status": n["status"]}
for n in (store_mod.read_json(
"evidence", f"{fid}.json") or {}).get("nodes", [])]),
"snippets_built": L("OBSERVED", len(pipeline_mod.get_snippets(fid))),
"temporal_context": L("OBSERVED",
"before/failure/after frames captured "
"per snippet"),
},
"3_failure_analysis": {
"statistics": L("OBSERVED", stat.model_dump() if stat else None),
"concentration": L("OBSERVED", {
"determination": conc.determination if conc else None,
"concentrated_dimensions": (conc.concentrated_dimensions
if conc else []),
"construction_stratum": next(
(s.model_dump() for s in (conc.strata if conc else [])
if s.stratum == "construction"), None),
}),
"fusion_verdict": L("OBSERVED",
fusion.output.get("overall_verdict")
if fusion else None),
"novelty": L("OBSERVED",
mining.output.get("novelty") if mining else None),
"scene_hypotheses": L("HYPOTHETICAL",
(vlm.output.get("hypotheses", [])[:3]
if vlm else [])),
"safety_chain": L("OBSERVED",
safety.output.get("chain") if safety else None),
"replay_worst_case": L("OBSERVED",
safety.output.get("worst_case")
if safety else None),
"replay_coverage": L("REQUIRED-EVIDENCE",
safety.output.get("coverage_note")
if safety else "no safety agent output"),
},
"4_launch_decision": {
"policy_version": L("OBSERVED", evaluation.get("policy_version")),
"severity": L("OBSERVED", evaluation.get("severity")),
"outcome": L("OBSERVED", evaluation.get("outcome")),
"recommended_option": L("OBSERVED",
evaluation.get("recommended_option")),
"matrix_row_fired": L("OBSERVED", evaluation.get("matrix_row_fired")),
"option_c_rejection": L(
"REQUIRED-EVIDENCE",
[c for c in (evaluation.get("option_c_evaluation") or {})
.get("checks", []) if not c.get("passed")]),
"expected_loss_table": L("HYPOTHETICAL",
evaluation.get("expected_loss_table")),
"mandatory_review_triggers": L(
"OBSERVED",
[t for t in evaluation.get("mandatory_review_triggers", [])
if t["fired"]]),
"human_review": L("OBSERVED", review_mod.decisions_for(fid)),
},
"5_learning_flywheel": {
"suites_created": L("OBSERVED", state.suite_ids),
"regression_hook": L("OBSERVED", regression.get("suites")),
},
},
"scorecard_id": state.scorecard_id,
"audit_records": L("OBSERVED", len(store_mod.audit_trail(fid))),
"audit_chain_valid": L("OBSERVED",
store_mod.verify_audit_chain(fid).get("valid")),
"separation_of_powers": (
"agents produced hypotheses and advisory summaries only; rates, "
"CIs, significance, severity, the policy outcome and the option "
"recommendation above all came from deterministic code, and the "
"final validation was a recorded human decision"),
}
store_mod.write_json(walkthrough, "worked_example", "latest.json")
return walkthrough
|