Architect Squad

Nine specialized AI agents that maintain a living Hierarchical Concept Tree (HCT) — a graph-based knowledge representation built from meetings, chats, documents, and voice transcriptions. The system produces whitepapers and delivery plans on demand.

Two Pipelines

Merge Pipeline (per-input)

                  ┌─ Shannon (conversations) ─┐       ┌─ same agent ─┐                ┌─────────┐
Input → router ──┤                            ├─ R1 ─┤              ├─ R2 ─→ Hawking ─┤ R1 → R2 ├─→ Dijkstra → Noether ↔ Dijkstra → Feynman ↔ Dijkstra → HCT v(N+1)
                  └─ Faraday  (documents)    ─┘       └──────────────┘                └─────────┘                                                         ↑
                                                  normalization two-pass            extraction two-pass                                            Scribe (always-on)

Each normalizer and Hawking run two passes — round 1 produces a provisional artifact, round 2 refines it with full context.

Synthesis Pipeline (on-demand)

"Produce whitepaper" → Turing → Von Braun

The Agents

Agent	Role	Pipeline	Named For
🧬 Shannon	Conversational Normalizer	Merge	Claude Shannon — proved noise can be separated from signal mathematically
📖 Faraday	Document Normalizer	Merge	Michael Faraday — whose structured notebooks were mined for insights decades later
🌌 Hawking	Concept Cartographer	Merge	Stephen Hawking — mapped invisible structure: black holes, radiation, spacetime
⚛️ Dijkstra	Integrator	Merge	Edsger Dijkstra — proved correct systems are built through disciplined composition
🔭 Noether	Coherence Guardian	Merge	Emmy Noether — proved every symmetry implies a conservation law
🔬 Feynman	Challenger	Merge	Richard Feynman — "you must not fool yourself, and you are the easiest person to fool"
📜 Turing	Whitepaper Author	Synthesis	Alan Turing — formalized the abstract into the concrete
🚀 Von Braun	Delivery Planner	Synthesis	Wernher von Braun — turned theoretical physics into phased engineering programs
📋 Scribe	Decision Logger	Always-on	The unnamed scribes — they predate individual fame; they are why we know history

How It Works

Merge Pipeline

1. Shannon or Faraday normalizes raw input into chunked, attributed representations with information-theoretic metrics (signal ratio, entropy, redundancy). Shannon handles conversations; Faraday handles documents.
2. Hawking extracts a local concept graph — a labeled property graph with typed relationships and evidence links. He never reads the global HCT to ensure unbiased extraction.
3. Dijkstra merges the local graph into the global HCT through three deterministic phases: identity resolution via optimal bipartite matching (≥0.80 match, ≤0.60 reject, between = CONFLICT), edit operation computation with correctness proofs, and invariant validation. He never guesses — ambiguity produces CONFLICTs, not resolutions.
4. Noether independently verifies Dijkstra's proofs and validates coherence across five dimensions. She holds veto power — overall coherence must be non-decreasing across merges. If she vetoes, Dijkstra must revise.
5. Feynman audits traceability from transcript chunk → evidence → concept → identity resolution → edit operation → final tree placement. If traceability breaks, Dijkstra must address challenges (max 3 iterations).

Synthesis Pipeline

6. Turing generates a formal whitepaper from the current HCT snapshot — includes concept stability indicators, maturity assessment, and ADRs.
7. Von Braun translates the whitepaper into a phased delivery plan. Stable concepts get firm work items; volatile concepts get only decision gates.

Scribe observes every stage and maintains an append-only decision log with coherence snapshots — the institutional memory across all runs.

Graph as Source of Truth

The HCT is a labeled property graph, not a tree. Hierarchy (contains edges) is one edge type among many. The contains-edge subgraph must form a valid tree (connected, acyclic, single-parent). If graph and tree views disagree, the graph wins.

Concepts get permanent IDs (C-{8-char-hex}) on first insertion. Hawking produces local IDs (LC-{kebab}); Dijkstra resolves them against the identity registry using optimal bipartite matching with explicit thresholds.

Persistent Shared State

File	Purpose
global-hct.json	The living concept graph
identity-registry.json	Concept identity across sessions
tree-grammar.json	Structural rules for valid tree shapes
coherence-config.json	Deterministic coherence formula parameters
decision-log.json	Append-only institutional memory

Repo Structure

.squad/
  agents/{name}/charter.md    Agent charters — role, boundaries, voice
  routing.md                  Pipeline routing and handoff rules
  team.md                     Canonical team roster
schemas/                      JSON Schema contracts between agents (draft 2020-12)
  common.schema.json            Shared types: envelope, evidenceRef, conceptId
  normalized-input.schema.json  Shannon/Faraday output
  local-concept-graph.schema.json  Hawking output (graph — source of truth)
  concept-tree.schema.json      Hawking output (derived tree view)
  global-hct.schema.json        The living concept graph schema
  merge-result.schema.json      Dijkstra output
  coherence-report.schema.json  Noether output
  identity-registry.schema.json Concept identity mappings
  tree-grammar.schema.json      Structural validation rules
  challenge-report.schema.json  Feynman output
  whitepaper.schema.json        Turing output
  delivery-plan.schema.json     Von Braun output
  decision-log.schema.json      Scribe's append-only log
  pipeline-manifest.schema.json Orchestrator configuration
OVERVIEW.md       End-to-end pipeline description
DESIGN.md         Schema design decisions and extension points
INPUT-GUIDE.md    How to prepare inputs for the pipeline
routing.md        Legacy routing reference
charter.md        Legacy charter reference

Key Design Decisions

Deterministic coherence — Coherence is computed from a deterministic formula configured in coherence-config.json. Both Dijkstra and Noether must produce identical scores from identical inputs — any delta is a bug, not a tolerance.

Two-pass processing — Normalizers and Hawking each run two passes. Round 1 is provisional; round 2 refines with full context. This catches what single-pass extraction misses.

Graph-first extraction — Hawking produces a labeled property graph (source of truth) and derives a tree view from it. The graph captures relationships that hierarchy alone cannot represent.

Artifact envelope — Every output wraps in a standard envelope carrying artifact_id, input_artifact_ids (provenance chain), and session_id. Full traceability from delivery plan back to source transcript.

Evidence references — Every concept links to specific chunks in the normalized input via evidenceRef. No concept floats without grounding.

Identity matching — Optimal bipartite matching with explicit thresholds (≥0.80 match, ≤0.60 reject, ambiguous zone emits CONFLICT). No guessing.

Boundary Rules

These are strict — violating them breaks the architecture:

• Shannon handles conversations only; Faraday handles documents only
• Hawking never reads the global HCT (extraction must be unbiased)
• Dijkstra never guesses — ambiguity produces CONFLICTs, not resolutions
• Feynman never fills gaps — he audits traceability only
• Scribe never modifies artifacts — append-only to the decision log

Current Status

• Schemas: Defined for all nine agents
• Agent charters: Complete at .squad/agents/{name}/charter.md
• Pipeline execution: Manual — sequential calls
• First pipeline run: Completed — 114 nodes, coherence 0.982