Analyzing a Pipeline

analyze_pipeline() is the package's flagship API. One call, full report.

Inputs

from minimal_oversight import analyze_pipeline

report = analyze_pipeline(
    pipeline,                   # PipelineGraph, LangGraph, ADK Agent, or dict
    p_min=0.80,                 # quality target
    traces=workflow_traces,     # optional: calibrate from logs
    governance_gap=0.02,        # λ — complexity sensitivity (~0.02/bit)
    process_entropy=0.0,        # H(W) — estimated from routing traces
    eta=10.0,                   # observation rate
    delta=2.0,                  # decay rate
)

What's in the report

The PipelineReport contains:

Section	What it tells you
report.feasibility	Can the pipeline hit p_min? C_op, B_eff, bottleneck node
report.node_estimates	Per-node σ_raw, σ_corr, M*, catch rate (from traces)
report.node_capacities	Per-node capacity from theory
report.motifs	Detected topology patterns (chain, fan-out, diamond, merge)
report.node_risks	Nodes ranked by delegation centrality and SOTA score
report.intervention_schedule	Per-node T*_auto and review frequency
report.alerts	Masking warnings, buffer alerts, capacity threshold breaches
report.recommendations	Ordered list of actionable governance changes
report.failure_explanation	Human-readable failure surface analysis

Reading the output

print(report)  # formatted summary

# Or access sections directly
if not report.feasibility.feasible:
    print(f"Target {report.feasibility.p_min} exceeds capacity {report.feasibility.c_op}")
    print(f"Bottleneck: {report.feasibility.bottleneck_node}")

for r in report.recommendations[:3]:
    print(f"{r.priority}. [{r.target_node}] {r.action}")
    print(f"   {r.rationale}")

With and without traces

Without traces: Parameters come from Node attributes or role-based defaults. Good for design-time analysis ("can this architecture work?").

With traces: estimation.py computes σ_raw, σ_corr, M*, and catch rate from real outcomes. The report includes confidence intervals. This is production monitoring.

# Design time — use estimated parameters
report = analyze_pipeline(pipeline, p_min=0.80)

# Production — use real logs
report = analyze_pipeline(pipeline, p_min=0.80, traces=production_traces)