Elixir / OTP / BEAM  ·  Open Source  ·  governed execution stack

Governed AI execution
as a typed platform stack.

The system is split by ownership. app_kit is the product boundary and northbound API. mezzanine owns reusable operational truth: packs, lifecycle, execution, decisions, evidence, projections, reviews, archival posture, and Temporal handoff. outer_brain owns semantic context and memory recall. citadel owns authority, policy, and governance compilation. jido_integration is the durable Spine for Brain ingress, lower facts, leases, review truth, and connector admission. execution_plane is the hazmat layer for raw runtime mechanics.

Technical deep dive

The stack is not a loose brand diagram. It is a set of typed handoffs, durability boundaries, and proof surfaces. Product code does not stitch lower services together directly. It enters through AppKit, which exposes stable DTOs and governed surfaces. Mezzanine turns product intent into substrate records. Brain-side packages shape semantic meaning and authority. The Spine accepts or rejects lower work, records facts, and mediates reads. Execution Plane performs raw effects behind explicit receipts. Stack Lab, AITrace, release manifests, schema registries, and no-bypass scans provide the proof surface.

AppKit boundary product-facing entry, stable DTOs, no-bypass enforcement

app_kit is the workspace product code is allowed to touch for governed platform behavior. It carries chat, typed domain, operator, work-control, runtime gateway, review, installation, semantic assist, trace lookup, and leased lower-read surfaces. It is also the enforcement point: product and hazmat scanner profiles reject direct imports into Mezzanine runtime services, Citadel internals, Jido Integration, and Execution Plane unless the product is authoring a pure Mezzanine.Pack model contract.

Mezzanine substrate neutral operational truth above the lower stack

mezzanine is the neutral high-level monorepo behind AppKit. It owns pack registration, compiled installation config, source admission, subject and object lifecycle, execution ledgers, decision and review state, evidence, projections, audit, archival restore posture, operator actions, runtime scheduling, and Temporal workflow handoff. Oban remains a local delivery and garbage-collection tool where explicitly retained; Temporal owns active workflow lifecycle, while Postgres stores substrate truth and projection facts.

Brain chain semantic runtime, policy compilation, authority

outer_brain owns semantic session work: context assembly, clarification, memory-shaped recall, normalized semantic outcomes, semantic failure journaling, and AITrace-linked provenance. citadel owns authority and policy compilation before mutation: action requests, capability posture, governance packets, topology, and the authorization shape that lower layers must receive. The split is deliberate: cognition can be rich, but durable mutation still needs deterministic authority.

Spine durable intake, auth lifecycle, review, lower facts

jido_integration is the Spine. It accepts Brain-origin work, validates ingress shape, owns auth and lease lifecycle, records review and control-plane truth, admits connectors, publishes connector effects, and exposes lower facts through tenant-scoped substrate reads. It is not just a connector workspace; it is the durable middle layer between Brain governance and raw execution mechanics.

Execution substrate transport, placement, process lanes, raw fact emission

execution_plane owns low-level effect mechanics: HTTP, process, JSON-RPC, placement, sandbox and operator-terminal lanes, plus future cell-style execution. Provider and runtime-family kits such as Pristine, Prismatic, CLI Subprocess Core, and Agent Session Manager normalize transport and session behavior. They emit receipts and raw facts; they do not own product truth, semantic meaning, or review state.

Governed memory access graph, immutable provenance, tiered recall

The memory architecture extends the same contract discipline into recall. Platform.AccessGraph.v1 makes user, agent, resource, scope, policy, and review-authority relations epoch-stamped. Private, shared, and governed memory live in separate stores with different writers and constraints. Share-up and promotion create new fragments, preserve parent links, apply transform policy, and emit proof tokens so a future audit can answer which memories influenced a decision, under which policy, and whether those fragments would still be admissible now.

Operator, trace, proof switchyard stack_lab AITrace ElixirScope
145
Repository nodes
516
GitHub stars
6
Explicit stack layers
3
Truth planes: semantic, durable, raw

Typed ingress before side effects

Product calls enter through app_kit with request context, tenant and installation refs, idempotency, trace, and typed DTOs. Mezzanine turns that into substrate records. Citadel compiles authority. Jido Integration accepts or rejects lower ingress. Execution Plane only sees shaped runtime intent, never product shortcuts.

Durability belongs to the owner

mezzanine records operational ledgers, workflow handoff, reviews, projections, audit, and archival evidence. jido_integration records lower facts, leases, connector admission, and review-adjacent Spine truth. execution_plane records raw runtime receipts. Recovery logic is attached to the layer that can actually replay or repair the fact.

Proof is part of the contract

stack_lab scenarios, AITrace spans, schema registries, release manifests, no-bypass scans, projection hashes, proof tokens, and retrospective audits are not documentation after the fact. They are how the platform proves recall, promotion, dispatch, operator review, tenant scoping, version skew, and boundary discipline under failure.

All Repositories

View on GitHub →
AI Infrastructure 27
json_remedy A practical, multi-layered JSON repair library for … ★ 31 snakepit High-performance, generalized process pooler and … ★ 11 rag_ex Elixir RAG library with multi-LLM routing (Gemini, … ★ 9 snakebridge Compile-time Elixir code generator for Python library … ★ 8 gepa_ex Elixir implementation of GEPA: LLM-driven evolutionary … ★ 3 tinkex_cookbook Elixir port of tinker-cookbook: training and evaluation … ★ 3 portfolio_core Hexagonal architecture core for Elixir RAG systems. … ★ 2 slither Lightweight Elixir runtime for composing and executing … ★ 2 tinkex Elixir SDK for the Tinker ML platform—LoRA training, … ★ 2 command Core Elixir library for AI agent orchestration - … ★ 1 execution_plane Execution Plane is an Elixir/OTP runtime substrate for … ★ 1 nsai_gateway Unified API gateway for the NSAI … ★ 1 nsai_registry Service discovery and registry for the NSAI … ★ 1 nsai_work NSAI.Work - Unified job scheduler for North-Shore-AI … ★ 1 pilot Interactive CLI and REPL for the NSAI ecosystem—unified … ★ 1 skill_ex Claude Skill Aggregator ★ 1 tiktoken_ex Pure Elixir TikToken-style byte-level BPE tokenizer … ★ 1 app_kit Shared app-facing surface monorepo for the nshkr … ground_plane Shared lower infrastructure monorepo for the nshkr … hf_hub_ex Elixir client for HuggingFace Hub—dataset/model … hf_peft_ex Elixir port of HuggingFace's PEFT (Parameter-Efficient … inference Reusable Elixir semantic inference contracts, adapters, … outer_brain Semantic runtime above Citadel for raw language intake, … portfolio_index Production adapters and pipelines for PortfolioCore. … portfolio_manager AI-native personal project intelligence system - … stack_lab Local distributed-development harness and proving … tinkerer Chiral Narrative Synthesis workspace for Thinker/Tinker …
Crucible Stack 27
LlmGuard AI Firewall and guardrails for LLM-based Elixir … ★ 9 ExFairness Fairness and bias detection library for Elixir AI/ML … ★ 1 crucible_examples Interactive Phoenix LiveView demonstrations of the … ★ 1 ExDataCheck Data validation and quality library for ML pipelines in … cns_crucible crucible_adversary Adversarial testing and robustness evaluation for the … crucible_bench Statistical testing and analysis framework for AI … crucible_datasets Dataset management and caching for AI research … crucible_deployment ML model deployment for the Crucible ecosystem. vLLM … crucible_ensemble Multi-model ensemble voting strategies for LLM … crucible_feedback ML feedback loop management for the Crucible ecosystem. … crucible_framework CrucibleFramework: A scientific platform for LLM … crucible_harness Experimental research framework for running AI … crucible_hedging Request hedging for tail latency reduction in … crucible_ir Intermediate Representation for the Crucible ML … crucible_kitchen Industrial ML training orchestration - backend-agnostic … crucible_model_registry ML model registry for the Crucible ecosystem. Artifact … crucible_telemetry Advanced telemetry collection and analysis for AI … crucible_trace Structured causal reasoning chain logging for LLM … crucible_train ML training orchestration for the Crucible ecosystem. … crucible_ui Phoenix LiveView dashboard for the Crucible ML … crucible_xai Explainable AI (XAI) tools for the Crucible framework datasets_ex Dataset management library for ML experiments—loaders … eval_ex Model evaluation harness for standardized … hf_datasets_ex HuggingFace Datasets for Elixir - A native Elixir port … metrics_ex Metrics aggregation and alerting for ML … training_ir Training IR for reproducible ML jobs across Crucible …