🛡️ Autonomous Agent Rule Synthesis & Governance Auditor

The saved audit fixture declares forty two trace steps. Only five are explicit source log entries; the remaining thirty seven are synthesized by linear token and cost growth. At seven pixels per count, explicit entries measure thirty five, generated entries two hundred fifty nine and total steps two hundred ninety four. The final fixture assigns one million four hundred fifty thousand tokens and seven hundred dollars. This is a scripted reconstruction, not independent evidence of a real agent session or actual invoice. One human baseline rule requests completing a refactor within its token budget. At step forty the fixture has one million three hundred eighty thousand tokens and six hundred sixty six point two one dollars. Step forty two has one million four hundred fifty thousand tokens and seven hundred dollars, adding seventy thousand tokens and thirty three point seven nine dollars. At one pixel per five thousand tokens the earlier, final and difference bars measure two hundred seventy six, two hundred ninety and fourteen. At the final step an invented rule demands three-level human consensus. The audit can identify that fixture text; it does not prove that every refusal or request for review in real work is unauthorized. The counterfactual depth option displays a cost cap of two hundred forty one dollars and savings four hundred fifty nine against seven hundred. At zero point four pixels per dollar the bars measure ninety six point four, one hundred eighty three point six and two hundred eighty. But the code only swaps a fixed result string; it does not recompute the trace or validate prevented costs. Blocking unauthored constraints takes priority when both options are selected and claims all seven hundred saved, another unverified counterfactual. Cytoscape initialization safely returns when unavailable, so native simulation text, scrubber and JSON export still work without the graph. Export retains all forty two fixture entries and the original final totals even after a counterfactual text change. Preserve that distinction and do not count a result label as an actual protected execution.