SuperInstance
SuperInstance is a collection of open-source projects for running AI agents inside sandboxed environments, and for the supporting math used to analyse them. The central project is OpenConstruct — a fork of NVIDIA OpenShell that wraps each agent sandbox in a self-contained room whose layout and configuration shape how the agent behaves, without weight-tuning. Around it sit a Python SDK for spectral graph conservation analysis, a Rust workspace for room-local model distillation, and a federated model-evaluation system.
Concepts
Rooms
In OpenConstruct, a room is a sandboxed agent workspace with its own context files, controls, help documents, and configuration. The room's layout acts as the prompt: it teaches the agent what to do without explicit prompting. Each room is isolated through OpenShell's control plane — Landlock, seccomp, and a local proxy enforce policy, credential, identity, and network boundaries. Lightweight monitor agents called ensigns watch each room for anomalies and stay in a cautious "yellow alert" state by default, on the principle that over-preparing is cheaper than under-preparing.
Single-number model control (JEPA gravity)
Each room carries one f64 value — its gravity — from which temperature, prompt style, max tokens, and sampling strategy are derived. This adapts a model's behaviour to the room without fine-tuning or weight changes. (JEPA = Joint Embedding Predictive Architecture, a learning approach that predicts how representations change rather than memorising what they are.)
The signal chain (L0–L4)
Most decisions a room makes do not need a large model. plato-nervous distils each room's intelligence through a five-layer signal chain, escalating only when lower layers cannot resolve a reading:
| Layer | Role | Model | Latency |
|---|---|---|---|
| L0 | Deadband — algorithmic threshold, no model | — | <1 ms |
| L1 | Nano completion model | phi4-mini (2.5 GB) | ~0.3 s |
| L2 | Room-local fine-tune (LoRA) | liquid-1.2b (698 MB) | ~0.3 s |
| L3 | Fleet / cross-room routing | gemma4:e2b (7.2 GB) | ~5 s |
| L4 | Cloud LLM for irreducible cases | any large LLM | 2–5 s |
Per the benchmark, the L0 deadband alone catches 76% of readings, so the majority of decisions never reach a model at all.
Source: plato-nervous/BENCHMARK.mdConservation spectral analysis
A separate line of work models structure as a weighted graph and analyses it through its Laplacian spectrum. The conservation ratio CR = λ₂ / λₙ measures how much spectral structure a transformation preserves; the spectral gap and Cheeger constant bound the quality of that structure by the graph's geometry. This is packaged as a Python SDK, conservation-spectral-python (published on PyPI as cocapn), with ports in TypeScript, Ada, and Rust.
Architecture & usage
OpenConstruct is written in Rust and runs each agent in an isolated room supervised by the OpenShell control plane (Docker / Podman / Kubernetes / VM compute drivers, with Landlock + seccomp policy). See the repository's architecture/ directory for the full design.
The conservation SDK has a small, stable surface. A minimal analysis, copied from the package README:
from conservation_spectral import ( TensionGraph, build_laplacian, eigendecompose, conservation_ratio, analyze, ) # Build a weighted directed (tension) graph g = TensionGraph() g.add_edge("A", "B", tension=0.8) g.add_edge("B", "C", tension=0.5) g.add_edge("C", "A", tension=0.3) # Spectral analysis: CR = λ₂ / λₙ L = build_laplacian(g) eigen = eigendecompose(L) cr = conservation_ratio(eigen.eigenvalues) print(f"Conservation ratio: {cr:.4f}") report = analyze(g) # ConservationReport(spectral_gap=..., cheeger=..., anomalous=False, ...)
Demos
Browser-based demos that ship in this repository. All run client-side.
- Demo Gallery →Index of every demo below, plus hosted music/constraint-theory demos.
- The NarrowsThree boats navigate a channel; Eisenstein (E12) integers stay exact where floats drift.
- Eisenstein 2DExplore the hexagonal lattice; points stay on-grid through 60° rotations.
- Eisenstein 3DRotate the hex lattice in three dimensions.
- Voxel Engine v2Browser voxel engine on Eisenstein coordinates.
- Voxel Engine v1Original voxel demo, educational.
- Live PlaygroundREPL for Eisenstein expressions and norms.
- Divergence-Aware ToleranceThresholds where exact arithmetic beats floating point.
- ErrataPublished negative results for the Eisenstein work.
Repositories
The projects that make up SuperInstance. Each link is a live repository.
| Repository | What it is | Language |
|---|---|---|
| OpenConstruct | Agent onboarding platform — agents in sandboxed rooms. Fork of NVIDIA OpenShell (Apache 2.0). | Rust |
| conservation-spectral-python | Spectral graph conservation SDK — tension graphs, Laplacian eigenvalues, conservation ratios, anomaly detection. On PyPI as cocapn. | Python |
| plato-nervous | Room-specific model distillation — the L0–L4 signal chain. | Rust |
| casting-call | Federated model-evaluation system — evaluation schema, trust engine, CLI. | Python |
| casting-call-gpu | CUDA kernel for pairwise Hamming-distance signature matching. | CUDA |
| eisenstein | Eisenstein-integer arithmetic crate, used by the math demos above. | Rust |
A live status view of the fleet telemetry service is on the fleet status dashboard.