Reading the Report

A guide to interpreting analyze_pipeline() output and acting on it.

The quick version

If you read nothing else, check three things:

Is it feasible? report.feasibility.feasible — if False, stop. Redesign first.
Where's the masking? report.alerts — any CRITICAL masking alerts need attention now.
What to change first? report.recommendations[0] — the highest-priority action.

Feasibility

f = report.feasibility

Field	Meaning
`f.feasible`	`True` if \(p_\text{min} \leq C_\text{op}\)
`f.c_op`	Pipeline's quality ceiling
`f.p_min`	Your quality target
`f.b_eff`	Autonomy buffer (positive = safe, zero = cliff, negative = impossible)
`f.bottleneck_node`	The node that limits capacity
`f.explanation`	Human-readable verdict

If infeasible: No routing rule, review allocation, or monitoring strategy will help. You need better agents, better correctors, or a different topology.

If feasible but b_eff is small (< 0.05): The pipeline is near the autonomy cliff. Any increase in workflow complexity or drift could push it over. Simplify routing or increase review capacity.

Alerts

for alert in report.alerts:
    print(f"[{alert.level.value}] {alert.node_name}: {alert.message}")

Three categories:

Alert	Trigger	Action
Masking (\(M^* > \tau\))	Corrector hiding weakness	Track \(\sigma_\text{raw}\) separately
Buffer (\(B_\text{eff} < 0.05\))	Near autonomy cliff	Simplify or increase review
Capacity (\(K/N < \text{threshold}\))	Corrector overloaded	Increase review capacity or reduce scope

Recommendations

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

Recommendations are sorted by the SOTA priority score: \(S(v) = \text{DC}(v) \times M^*(v) \times \kappa(v)\). The node where governance investment produces the largest improvement appears first.

Intervention schedule

for s in report.intervention_schedule:
    print(f"{s.node_name}: review every {s.t_auto:.0f} steps")

Nodes with short \(T^*_\text{auto}\) need frequent human review. Nodes with long \(T^*_\text{auto}\) can operate autonomously. The schedule is a linear program that minimizes total review cost subject to meeting each node's minimum frequency.

If you observe...

Observation	Diagnosis	Paper reference
High \(\sigma_\text{corr}\) but falling \(\sigma_\text{raw}\)	Masking deterioration	Table 9, row 1
\(p_\text{min} > C_\text{op}\)	Infeasible — redesign	Table 9, row 2
Large \(\lambda H(W)\) eating the buffer	Process overload	Table 9, row 3
One node with very short \(T^*_\text{auto}\)	Operational bottleneck	Table 9, row 4
Quality drops only when upstream A fails	Diamond fragility	Table 9, row 5
Scope collapses to easy tasks only	Missing coverage constraint	Table 9, row 6