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.
┌─ 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.
"Produce whitepaper" → Turing → Von Braun
| 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 |
- Shannon or Faraday normalizes raw input into chunked, attributed representations with information-theoretic metrics (signal ratio, entropy, redundancy). Shannon handles conversations; Faraday handles documents.
- 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.
- 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.
- 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.
- 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).
- Turing generates a formal whitepaper from the current HCT snapshot — includes concept stability indicators, maturity assessment, and ADRs.
- 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.
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.
| 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 |
.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
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.
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
- 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