Case Study

Two end-to-end runs on the humor book prompt, two months apart. Same four-sentence input. Different demonstrations.

Run	When	Demonstrates	Pages
v2 — Hierarchical Dispatch	May 2026	Multi-level dispatch tree (78 dispatches across 3 levels), ACT-R memory in active use, both escalation modes from §7, skill-candidate reuse from a prior run	Overview · Dispatch Tree · Proxy & ACT-R · Artifacts · Takeaways
v1 — End-to-End Manuscript	March 2026	The CfA state machine driving Intent → Planning → Execution end-to-end, proxy participation at gates, a complete 55,000-word manuscript	Overview · Intent & Planning · Execution · Manuscript · Obstacles · Learnings · Takeaways

Why two. v1 was the first end-to-end run that produced a real manuscript. Its scope statement was honest about what it did not cleanly demonstrate — most prominently the Hierarchical Teams pillar, where execution was driven primarily by the project lead rather than by independent sub-agents running their own CfA cycles in isolated worktrees. v2, run two months later on the same prompt, closes that gap: 78 dispatches across a 3-level tree, with concrete evidence of memory, skill reuse, and both escalation modes. v1 is preserved as the end-to-end intent-satisfaction baseline; v2 is where the architectural claims now ground.

What's the same. Same prompt (a four-sentence ask for a book on universal humor), same project lead role, same proxy concept, same CfA state machine. The protocol didn't change between runs.

What's different. What the system did with that protocol. v2's run is the one to read for evidence of the four pillars composed.