Topology gate checks calibration-phase metrics instead of final network metrics

[DeveshParagiri]

Bug

The topology gate rejects networks based on metrics from the calibration phase (similarity-only edges), not the final network (after structural edges, triadic closure, and rewiring). This causes valid networks to be quarantined.

Evidence

ASI scenario, 5,000 agents, avg_degree=10:

Metric	Gate Bounds	Calibration (checked)	Final (ignored)
avg_degree	8.5–11.5	1.53 FAIL	10.4 PASS
clustering	≥0.14	0.01 FAIL	0.26 PASS
modularity	0.30–0.50	0.23 FAIL	0.83 FAIL*
LCC	≥0.95	0.58 FAIL	0.999 PASS
edge_count	≥16,250	3,827 FAIL	25,989 PASS

Calibration only produces ~3,800 edges via similarity threshold. Then ~22,000 structural edges (neighbor, congregation, partner, political_community, etc.) are merged post-calibration — the gate never sees them.

*Modularity is a separate issue — see below.

Root Cause

In generator.py, best_metrics and accepted are set during the calibration loop (after _sample_edges + _triadic_closure). Post-calibration steps (structural edge merge, _apply_rewiring) change the network substantially but the gate never re-evaluates.

The compute_network_metrics() call in generate_network_with_metrics() computes correct final metrics, but they're only informational — never fed back into the gate decision.

Fix

Re-evaluate the topology gate against the final network metrics (after structural edges + rewiring), not the calibration-phase metrics.

Related Issue: LLM-generated targets don't match population structure

The LLM set target_modularity: 0.55 but designed edge rules (state + urban/rural + race blocking) that naturally produce modularity ~0.83. For 5,000 agents across 50 states with strong geographic sorting, high modularity is realistic — the LLM's edge rules are correct but its target numbers are textbook defaults rather than population-specific reasoning. The config generation prompt should instruct the LLM to reason about expected metric ranges for the specific population structure.

[DeveshParagiri]

Deeper issue: calibration is irrelevant when structural edges dominate

The gate bug is a symptom. The recurring structural problem:

Calibration tunes similarity threshold → produces ~15% of final edges. Structural edge rules (LLM-generated) produce ~85%. Calibration is optimizing a minor knob while the network's real topology is determined by rules it doesn't see.

In the ASI run:

• Calibration edges: 3,800 (15%)
• Structural edges: 22,162 (85%)
• Calibration spent 3+ minutes searching for a threshold that barely matters

This will recur for any population with strong geographic/demographic blocking (i.e., most realistic populations), because the LLM correctly generates edge rules based on state, religion, race, politics — and those rules produce far more edges than similarity threshold matching.

The calibration architecture assumes similarity-based edge selection is the primary mechanism. When it isn't, calibration wastes time, produces misleading metrics, and the gate rejects good networks.

Possible directions:

1. Calibration should be aware of structural edges — run them first, then calibrate similarity threshold to fill remaining degree budget
2. Or skip calibration entirely when structural edge rules are present and sufficient
3. Gate should always evaluate the final network regardless

[RandomOscillations]

Closing as fixed. Topology gate evaluation now uses final-network metrics ( with pre/post-rewire gate handling) and strict gate reporting uses those final metrics. See and .