Methodology

Phase A — Seed corpus (11 frameworks)

Phase A identified 11 canonical reference frameworks as the research baseline: spec-kit, openspec, superpowers, BMAD-METHOD, agent-os, claude-flow/ruflo, taskmaster-ai, kiro, ccmemory (+ ccmemory-plain delta variant), claude-conductor, spec-driver. These were analyzed against RUBRIC-v3 to establish the initial 5-archetype taxonomy.

Phase B — Tier A frameworks (319 frameworks)

33 thematic batches, each containing ~10 frameworks discovered via recursive web search and GitHub corpus analysis. Each framework received the full 10-file treatment: 00-summary through 09-uniqueness, plus METRICS.yaml. Batch synthesis notes capture intra-batch patterns and highlight the most interesting finds. Phase B expanded the taxonomy from 5 to 15 archetypes, uncovering an operational layer (orchestrators, PR-lifecycle daemons, governance, UI passthroughs) the seeds had no vocabulary for.

Phase C — First synthesis (Phase B only)

Cross-cutting synthesis of the Phase B corpus (330 frameworks total). Produced: 15 archetypes, 25 engineering dimensions, 10 one-of-a-kind primitives (first catalogue), and the Phase C 5-framework reference stack. Superseded by Phase E but retained as a baseline checkpoint.

Phase D — Tier B-from-D frameworks (323 frameworks)

Another 33 thematic batches across the Tier-B candidate list. Discovered three new archetypes (cross-vendor model routing A16, compiled enforcement A17, self-evolving plugins A18) and three more (token-efficient encoding A19, OS-distro substrates A20, LLM-evaluated hooks A21). Phase D batch notes are in _index/phase-d-batch-NN-notes.md.

Phase E — Full corpus synthesis (654 frameworks)

Cross-corpus synthesis across all 654 frameworks. Produced the final deliverables: 21-archetype taxonomy, 45 one-of-a-kind primitives, 5-framework synthesis baseline, 7-framework defensible set, top-15 deep-dive list, and the headline finding that the community has split into five concurrent races.