Grand Diomande Research · Full HTML Reader

Stage 0: Research Brief -- Comp-Core Production Map

Comp-Core is a monorepo at `Desktop/Comp-Core/` containing 73 components across 8 domain layers, 13 packages, ~30 applications, and 4 backend services. It started as a motion-intelligence system for computational choreography and has grown into Mohamed's full infrastructure backbone: retrieval (RAG++), semantic graph (Graph Kernel), agent orchestration, audio synthesis (Echelon), gateway connectors, and ML pipelines.

Language as Infrastructure research note experiment writeup candidate score 22 .md

Full Public Reader

Stage 0: Research Brief -- Comp-Core Production Map

What Is Comp-Core

Comp-Core is a monorepo at `Desktop/Comp-Core/` containing 73 components across 8 domain layers, 13 packages, ~30 applications, and 4 backend services. It started as a motion-intelligence system for computational choreography and has grown into Mohamed's full infrastructure backbone: retrieval (RAG++), semantic graph (Graph Kernel), agent orchestration, audio synthesis (Echelon), gateway connectors, and ML pipelines.

Version: 0.1.0. Last INDEX.md update: 2025-02-02. Actual development has continued aggressively through Q1 2026.

---

The 8 Core Layers -- Real State

Layer 1: Runtime (core/runtime/) -- ~87K lines

ProjectLangClaimed StatusReal State
cc-corePythonCompletePython LIM-RPS algorithms. Has pyproject.toml + requirements.txt. Foundational. Works.
cc-core-rsRust/PyCompleteDSP kernels, lock-free buffers. Has Cargo.toml + pyproject.toml (PyO3).
cc-kernelRustSubstantialSMART execution kernel. Has Cargo.toml. No release binary found.
cc-protocolRustCompleteUnified protocol definitions. Has Cargo.toml.
cc-migration-architectTSPresentHas package.json. Schema migration tooling.

Honest assessment: Runtime layer is the most stable. cc-core and cc-core-rs are genuinely complete. cc-kernel has code but "substantial" means it compiles, not that it's battle-tested under load.

Layer 2: Motion (core/motion/) -- ~51K lines

ProjectLangClaimed StatusReal State
cc-typesRustFROZENCore motion types. Intentionally locked.
cc-anticipationRust/PyCompleteAnticipation kernel with 7 scalars. Has target/ dir (was compiled).
cc-collectionRust/PyCompleteExtended Kalman Filter fusion. Has target/ dir.
cc-gestureRustPartialHas target/ dir but gesture recognition is incomplete.
cc-window-alignerRust99
cc-motion-utilsTSPlannedDoes not exist as real code.
sensor-pipeline--PresentListed in directory, not in INDEX.md.

Honest assessment: The motion math works. The integration story between motion and audio (the whole point of computational choreography) has gaps, specifically cc-gesture and the window-aligner integration hole.

Layer 3: Semantic (core/semantic/) -- ~70K lines

ProjectLangClaimed StatusReal State
cc-semanticRustFROZENReplay-stable semantic frames. Locked.
cc-semantic-languageRust/PyCompleteN'Ko operator-based processing.
cc-cognitivetwin-bridgeRustCompleteConversation DAG bridge with FFI.
cc-graph-kernelRustCompleteDeterministic context slicing. Has release binary (graph_kernel_service). Actually deployed as LaunchAgent on cloud-vm. Port :8001.
cc-inscriptionRustCompleteN'Ko inscription compiler. Has target/ dir.

Honest assessment: This is the strongest layer. Graph Kernel is genuinely production -- it has a compiled binary, a LaunchAgent plist, health check endpoints, and runs against Supabase PostgreSQL. cc-inscription powers the anticipation geometry research (validated with 40K records in Supabase).

Layer 4: Retrieval (core/retrieval/) -- ~6.7M lines (inflated by data/generated files)

ProjectLangClaimed StatusReal State
cc-rag-plus-plusRust/PyCompleteFastAPI service at :8000. Has .venv. Actually deployed as LaunchAgent. 332K rows re-embedded with OpenAI text-embedding-3-small. pgvector search live.
cc-retrievalRustCompletePhrase selection contracts, ~2200 lines.
cc-context-recovery-mcpPythonCompleteMCP server for context recovery.
cc-idea-vaultRust/PySubstantialResearch idea management.
cc-context-intelligence--PresentNot in INDEX.md. Exists in directory.
cc-retrieval-indexer--PresentNot in INDEX.md. Exists in directory.
cc-vault-mcp--PresentNot in INDEX.md. Exists in directory.

Honest assessment: RAG++ is genuinely production. It has a running service, LaunchAgent, health endpoint, 332K embedded rows. The line count is inflated by data files in the directory tree. The MCP servers (context-recovery, vault) are functional but not load-tested. Three projects in the directory aren't even listed in INDEX.md -- the index is stale.

Layer 5: Audio-Media (core/audio-media/) -- ~107K lines

ProjectLangClaimed StatusReal State
cc-echelonRustPartial19 nested crates. Has release binary (libmotion_bridge). Biggest codebase in the monorepo. The canonical gap: Layer 4 Temporal Intelligence not built (phase extraction, soft tempo, adaptive bar boundaries).
cc-djRust/PySubstantialDJ control with Gemini integration. Has target/ dir.
cc-speakRust/PyCompleteAudio capture engine. Has target/ dir. SpeakFlow V2 evolved from this.
cc-streamRustPartialStreaming infrastructure. Has target/ dir.
cc-cinematographerRustSubstantialVideo processing (beat-sync, Ken Burns). Has target/ dir. FFmpeg/HLS export stubs only.
cc-event-busRustCompleteEvent routing, pub/sub. Has target/ dir.

Honest assessment: This is the most ambitious and most incomplete layer. Echelon alone is 19 crates, but the critical gap (temporal intelligence -- what makes "performer = clock" true) is explicitly unbuilt. cc-cinematographer has stubs where real FFmpeg calls should be. cc-stream is partial. The event-bus works. cc-speak works. But the integrative vision of "motion drives audio in real-time" is not production-ready.

Layer 6: Agents (core/agents/) -- ~44K lines

ProjectLangClaimed StatusReal State
cc-agentRustCompleteMulti-provider agent client. Has Cargo.toml.
cc-agent-daemonTSCompleteTask polling daemon. Has node_modules, actually installed.
cc-agent-sdkTSCompleteTypeScript agent presets.
cc-conductorRustCompleteControl-theoretic policy layer. Has target/ dir (no release binary).
cc-orchestrator-agentRustCompleteMulti-agent task orchestrator.
cc-training-orchestratorPythonCompleteML training pipeline. Has requirements.txt.
cc-project-spawnerTSPartialHas node_modules. Project scaffolding.
cc-spawn-testTSPartialAgent spawn testing.

Honest assessment: The agent layer is well-coded but redundant with external systems. meshd (Rust PTY daemon at :9451) and the pane orchestrator have largely replaced what cc-conductor and cc-orchestrator-agent were designed to do. The agent-daemon polls Supabase for tasks -- it works, but the multi-agent mesh has moved to tmux injection + NATS, not this polling model. There's overlap that needs resolution.

Layer 7: Gateways (core/gateways/) -- ~28K lines

ProjectLangClaimed StatusReal State
cc-relayRustCompleteMocopi UDP relay with TLV parsing.
cc-geminiRustCompleteGemini API client. Has target/ dir.
cc-sms-gatewayRustSubstantialTwilio SMS webhooks.
cc-telegram-gatewayRustStubNot implemented.
cc-claude-code-gatewayRustStubNot implemented.
cc-bridgedRustDeployed4MB arm64 binary. 3 adapters (OSC :9462, WS :9463, Supabase). Actually deployed.
cc-discord-gateway--PresentIn directory but not in INDEX.md.
clawdbot-integration--PresentIn directory but not in INDEX.md.

Honest assessment: cc-bridged is the winner here -- it's compiled, deployed, and replaces 15+ Python/Node bridges. cc-relay works for Mocopi. The SMS gateway is substantial. Two stubs (telegram, claude-code) are dead weight. Two projects aren't in the INDEX. The INDEX is stale again.

Layer 8: ML (core/ml/) -- ~96K lines

ProjectLangClaimed StatusReal State
cc-mlPythonSubstantialMotion synthesis, diffusion models.

Honest assessment: One project, ~96K lines of Python. This is the training pipeline for motion synthesis. Separate from KARL (which lives in Desktop/karl/) and Thunder-Train (distributed MLX). The ML work has fragmented across the mesh -- Mac4/Mac5 do the actual compute, this directory has the pipeline definitions. Whether it runs end-to-end without manual intervention is unclear.

Additional Layers (Not in CLAUDE.md's 8-layer table)

LayerPathReal State
orchestratorscore/orchestrators/cc-pulse-orchestrator (Rust, partial), beyond (Rust, has release binary)
visioncore/vision/cc-perception-mesh (0 lines currently)
mesh-computecore/mesh-compute/Rust workspace, 543 lines. Cargo.toml present.

---

The Packages Layer (packages/) -- 13 packages

PackageLangLinesState
admissibility-kernel-pyRust/Py205KPyO3 bindings. The line count includes generated/compiled artifacts.
cognitive-twinPython32KUser pattern learning, DPO. V3.0.0. Substantial.
cc-voice-unifiedTS5.3KVoice command grammar. Has vitest tests.
cc-event-sourcingTS2.9KCQRS/DDD patterns.
cc-cognitive-pulseTS2.4KCognitiveTwin for Pulse.
anticipation-geometryPython2.2KResearch package. Validated with Supabase data.
cc-evolution-frameworkTS827Creative evolution.
cc-voice-pipelineTS636Voice processing. package.json present.
admissibility-kernelRust--DAG context slicing.
rag-plusplus-coreRust--SIMD vector search core.
rag-plusplus-pyRust/Py--PyO3 bindings for RAG++.
shared--61Shared utilities.

---

Applications -- ~30 apps across 5 platforms

Trajectory Suite: 11 apps (search, orbit, analytics, desktop, mobile, walk, ios, latent-browser, novel, dream-studio, kinetic-theater, research-browser). trajectory-search and trajectory-orbit have compiled Rust binaries.

Web Apps: 9 Next.js apps. cc-dashboard is the main one. learnnko, skills-graph, cc-rag-viz are active.

iOS: 3 apps in Comp-Core (EchelonCapture, CinemaWalk, cc-handguard). Plus 43+ additional iOS apps elsewhere in the fleet.

Desktop: Echelon (Tauri/Rust).

HEF Evolutions: 2 experimental web apps (aesthetic-dna, failure-museum).

---

Infrastructure That Actually Runs

ServicePortDeploymentStatus
Graph Kernel:8001LaunchAgent on cloud-vmLIVE
RAG++:8000LaunchAgent on Mac1LIVE (was offline in March, recovered)
OrbitcloudCloud RunLIVE
cc-bridged:9462/:9463Binary on Mac1LIVE
meshd:9451LaunchAgent all machinesLIVE
NATS:4222Mac1LIVE
OPA:8181Mac1LIVE

---

The Gap Between Built and Production

### What "production" means here
Mohamed's mesh is not a SaaS startup. "Production" means: the service starts automatically (LaunchAgent/systemd), survives reboot, has health checks, logs to a known location, and other services depend on it without manual intervention.

### Actually production (4 projects)
1. cc-graph-kernel -- binary, LaunchAgent, health check, Supabase backend
2. cc-rag-plus-plus -- venv, LaunchAgent, health check, 332K rows
3. cc-bridged -- binary, 3 adapters, deployed
4. cc-event-bus -- works as internal pub/sub

### Could be production with 1-2 days work (6 projects)
1. cc-core / cc-core-rs -- stable but no service layer
2. cc-anticipation -- works but needs deployment wrapper
3. cc-inscription -- works, used in research
4. cc-relay -- works for Mocopi
5. cc-speak -- works for audio capture
6. cc-agent-daemon -- polls tasks, needs LaunchAgent

### Prototype / substantial but not deployable (12 projects)
cc-echelon, cc-dj, cc-cinematographer, cc-stream, cc-conductor, cc-orchestrator-agent, cc-gesture, cc-window-aligner, cc-sms-gateway, cc-ml, cc-kernel, cc-pulse-orchestrator

### Stubs / dead weight (4 projects)
cc-telegram-gateway, cc-claude-code-gateway, cc-motion-utils, cc-perception-mesh

### Experimental / may abandon (4 projects)
latent-browser, novel, aesthetic-dna, failure-museum

---

Build System -- Current State

There is no unified build system. The current state:

  • Rust projects: Individual Cargo.toml files. No workspace Cargo.toml at root. Each project compiles independently. Some have release binaries, most have only debug builds or no builds at all.
  • Python projects: Mix of pyproject.toml, requirements.txt, and setup.py (root level). No unified virtual environment. RAG++ has its own .venv.
  • TypeScript projects: Individual package.json files. No monorepo tool (no turborepo, no nx, no pnpm workspace). Each has its own node_modules.
  • Docker: docker-compose.yml exists but references an old inference service. Not used for current production services.
  • CI/CD: 8 GitHub Actions workflows exist (ci.yml, rust-tests.yml, etc.) but their actual run state is unknown.
  • Makefile: Exists but targets are for Episode 1 (the original motion capture project), not the full monorepo.

---

Source Signals

1. INDEX.md (2025-02-02) -- comprehensive but stale. Missing 5+ projects that exist in directories.
2. STRUCTURE.md -- matches reality for top-level dirs.
3. CLAUDE.md (in Comp-Core) -- accurate for service ports and running state.
4. HEALTH-REPORT-20260301.md -- honest about RAG++ and GK being down. RAG++ has since recovered.
5. *CP- completion files -- 10 completed project plans. Evidence of real execution.
6.
Agent memory -- comp-core-layers.md reference exists but file was missing from pulse-runner memory. Echelon canonical gap documented separately.
7.
Filesystem scan** -- 73+ components verified. Release binaries confirmed for: graph-kernel, bridged, beyond, echelon (libmotion_bridge only).

---

Key Tensions

1. Monorepo without monorepo tooling. 73 components, no workspace build, no dependency graph enforcement.
2. Three languages, zero unified CI. Rust/Python/TypeScript each have their own world. No single `make build-all` that works.
3. INDEX.md drift. The authoritative index hasn't been updated in 14 months. Reality has outgrown documentation.
4. Agent layer redundancy. cc-conductor, cc-orchestrator-agent, meshd, pane-orchestrator, and NATS all do overlapping work.
5. Audio-media ambition vs. reality. Echelon is 19 crates and 107K lines but temporal intelligence (the core differentiator) isn't built.
6. Production services run but aren't reproducible. Graph Kernel and RAG++ work, but deploying them on a new machine requires manual .env setup, manual binary compilation, and manual LaunchAgent installation.

Promotion Decision

Attach run IDs, datasets, metrics, and reproduction commands.

Source Anchor

evo-cube-output/comp-core-production-map/stage0-research.md

Detected Structure

Method · Evaluation · References · Code Anchors · Architecture · is Stage Research