BEAM-native AI infrastructure
and evidence you can check.
180 repositories and 673 stars of Elixir/OTP libraries for AI systems: LLM routing, multi-agent orchestration, provider SDKs, and data orchestration. Behind them sits NSHKR, a governed execution stack where AI-proposed actions become authorized, receipted, replayable operations, and a mechanistic interpretability research program run with frozen protocols, matched controls, and reports that preserve negative results.
Systems Engineering: Governed AI Execution
The Elixir/OTP side of the portfolio builds the write-path substrate an
enterprise needs after an AI system proposes an action. The important
artifact is not a prompt transcript; it is a durable chain from request
to authority, workflow, external effect, receipt, evidence, review, and
replayable proof. nshkr
is the production composition and release backbone; its
Nshkr.Runtime application boots the bounded owner services
as one fail-closed platform.
Actor, tenant, installation, request context.
Capability, policy, scope, review gate.
Operation context, causation, idempotency.
Operation class, manifest, lower instruction.
Effect receipt, status, failure class.
Evidence refs, attached artifacts, proof tokens.
Review case, decision, override, escalation.
AITrace DAG, predecessor refs, merge semantics.
Spatial topology, target host ID, BEAM node ref.
nshkr is the production backbone rather than another
capability package. Nshkr.Runtime composes persistence,
workflows, context, authority, provider accounts, execution, cluster
reconciliation, Synapse, and Extravaganza into reproducible monolith
and distributed deployments. Production profiles reject memory,
fixture, no-op, and static-success substitutes.
app_kit is the northbound surface product code is allowed
to touch for governed platform behavior. It accepts product-level
commands and stable DTOs, exposes operator reads and review controls,
and keeps products from stitching lower execution paths together by
hand.
mezzanine owns reusable operational truth: binding
registry, compiled run snapshots, workflow handoff, execution ledgers,
operation receipts, evidence, projections, review state, audit, and
operator actions. It records the facts that make the write path
replayable.
outer_brain owns semantic context, recall, normalized AI
outcomes, and semantic failure carriers. citadel
authorizes after resolution, when the operation class, manifest,
side-effect class, required scope, and credential constraints are
known.
jido_integration resolves connector manifests, operation
descriptors, and credential leases into governed lower invocation. It
lets provider mechanics be specific without leaking provider-shaped
control flow into reusable platform surfaces.
execution_plane performs raw mechanics across HTTP, CLI,
process, JSON-RPC, sandbox, terminal, filesystem, and future execution
lanes. It emits receipts and raw facts. Product meaning, review state,
and operational projections stay above it.
chassis is the self-replicating spatial substrate that maps virtual stack components to physical BEAM nodes. It coordinates host inventory, provisions remote nodes via native Elixir SSH bootstrap or pluggable adapters, and injects memory-decrypted age/SOPS credentials directly into Systemd service environments.
AITrace records causal execution events with predecessor
references so replay is not just emission order. stack_lab
adds scanners, acceptance gates, negative controls, failure drills,
and second-product validation so generality is tested instead of
assumed.
Research Programs
Each program keeps its question, controls, artifacts, and claim boundary together. Results are stated at the level the evidence supports: a decoder is not a mechanism, an internally coherent operator is not automatically useful, and a clean null is still a result.
CGL — Causal Generalization Laboratory
An open experimental program to predict and control how fine-tuning, continued adaptation, and skill composition alter model behavior across contexts. Ten linked experiment families connect paired model organisms, presentation controls, independent persona and style subspaces, causal interventions, prospective forecasts, continued-training stress tests, composition, behavioral breadth, and cross-model transfer. Python 3.14 and PyTorch executions preserve data and model revisions, source snapshots, checkpoints, raw scores, failures, and aggregate evidence. Both 7,049-example paired training runs and four GPU acceptance campaigns are complete; behavioral evaluation and the wider research program continue. Read the research plan.
Geometry of Conditional Truth
A cross-model Qwen3-4B and Phi-4-mini replication. Across eight preregistered endpoints, Phi supported one and Qwen supported none; both remained Level 1 of 6. Cross-model report.
Architecture Mechanics
Tiny trained models and known synthetic features make transport, overwrite, packing, and intervention outcomes directly measurable across mixing mechanisms.
Attention Lab
A matched GPT-pretraining harness for alternative attention architectures. Two confirmatory pretraining runs reached full-depth analysis; the current mechanism verdict remains insufficient evidence.
Superposition Zoo
Synthetic retrieval comparisons across attention, routing, linear attention, delta memory, and state-space mixing. Retrieval findings are real; the central feature-isolation question remains open.
Authority follows resolution
A broad product request is not enough. Citadel authorizes the resolved operation plan after Mezzanine knows the operation class, manifest ref, binding ref, side-effect class, credential scope, and review constraints.
Provider specificity is data
A pack can bind issue tracking to Linear, code hosting to GitHub, or runtime work to Codex. Reusable surfaces still speak product roles, manifests, operation classes, receipts, and projections.
Proof is executable
stack_lab scenarios, AITrace DAGs, schema registries, release manifests, no-bypass scans, projection hashes, and proof tokens are not documentation after the fact. They are acceptance gates.
Repository Atlas
Generated from live repository metadata across the NSHKR and North Shore AI ecosystem.