Sphere is a research program focused on the invariant infrastructure needed to make it safe for any type of stateless intelligence to take autonomous stateful action.
The name Sphere comes from the geometric property that a sphere has a finite and fully known surface area while containing an infinite and uncountable number of radii.
Sphere is currently focused on engineering with LLMs and is investigating two claims:
- Agentic action can be bounded, observable, and verifiable without trusting the agent or forfeiting its expected utility.
Sphere hypothesizes that agents operating within its engineering produce lower out-of-bounds action rates and higher verifiability than the same agent without it, with no drop in work quality. This is tested by ablating components across fixed agent tasks, isolating the engineering as the variable. The hypothesis is weakened if ablation produces no measurable change, if the engineering reduces work quality, if prompts alone produce the same bounds, or if a placebo version produces comparable results.
- Agent performance can be stabilized across open and SOTA models with the use of a closed codebase vocabulary, deterministic tools, cross-model evaluations, and interpretability-informed training.
Sphere hypothesizes that an open model within its engineering performs comparably to state-of-the-art models from one or two generations ago, on the same work, with less variance across runs. This is tested by running the same work items on the open model and prior-generation SOTA models with engineering active and ablated, isolating the engineering as the variable. The hypothesis is weakened if the open model does not approach prior-generation SOTA performance, if one component does all the work, if results don't generalize across models, or if a placebo version produces the same results.
Research is executed across ten tracks: Axis, Inclination, Tangent, Radii, Curve, Arc, Surface, Shell, Sector, Limit. Each track enables, produces, or furthers Sphere's work.
Orientation:
FINDINGS.md— what Sphere has learned, with links to sources.INCLINATION.json— the live operator working surface; intents, hypotheses, experiments.experiments/— per-experiment design, runs, artifacts, papers.architecture/— design documents spanning experiments.
Axis is the infrastructure the research relies on.
A single Mac mini was chosen to host the research due to the unified memory it offers and its wide consumer availability. Where possible, infrastructure choices favor open source and enterprise-grade solutions to strengthen scalability and reproducibility of research findings in both enterprise and independent contexts.
| Term | Description |
|---|---|
| Identity | Domain, Entra IAM, Microsoft Exchange, YubiKey |
| Vault | Bitwarden |
| Compute | Mac mini (M4 Pro, 64GB RAM, 512GB SSD), GCP, AWS |
| Network | Tailscale, Spectrum |
| Container | Podman |
| Engine | MLX LM, Ollama, llama.cpp |
Inclination is the intent, context, expected outputs, and desired outcomes provided to the agent via the Work directory.
The human originates, the agent executes, and then the human reviews and approves. Evaluation, verification, and enforcement are ensured independently of both.
| Term | Description |
|---|---|
| Human | Operator; originates intent, reviews and approves outputs |
| Work | Work directory; intent, context, outputs, outcomes |
Tangent provides the human-computer interface.
A custom CLI provides both a unified command and observability surface. The Mac mini and CLI are configured to enable remote access via SSH over Tailscale using Blink Shell.
| Term | Description |
|---|---|
| CLI | Sphere CLI |
| Remote | Blink Shell, herdr (open-source terminal multiplexer for agents) |
Runtime environment for the main builder agent.
The agent's attention is focused on executing work while deterministic code enforces authorization boundaries and workflow requirements.
| Term | Description |
|---|---|
| Runtime | Vendor/OSS |
| Sandbox | macOS |
| Codebase | Closed code vocabulary, documentation |
| Stop Hooks | Per-turn batch commits |
| Tools | just-bash (Vercel Labs) |
| LLM1 | Builder model |
Curve is the study of agent trajectories as new capabilities are established.
Where Arc records the execution that occurred, Curve examines the executions that could occur, the edge cases, adversarial chains, and failure paths.
| Term | Description |
|---|---|
| Red-teaming | Adversarial probing of reachable agent trajectories |
| Capability analysis | Mapping the paths a new Radii capability makes reachable |
Mediation and evidence layer for agent invocation and machine execution.
The agent invokes tools via a deterministic intermediary that initiates execution, returns outputs, and writes each step to an append-only log; tools requiring network connectivity route through an APISIX gateway.
Each work cycle is captured in Git, enforced against policy as code (OPA, AST walkers), and bundled into the Debrief, which is an evidence packet in HTML format that the operator reads to approve or reject the work for push.
The append-only log and APISIX gateway provide real-time observability via the Sphere CLI.
| Term | Description |
|---|---|
| Intermediary | Deterministic (Rust) |
| Log | Append-only |
| Gateway | APISIX |
| Git | Version control, commit signing, pre-push hooks |
| Policy | OPA, AST walkers |
| Debrief | Evidence packet in HTML format |
Surface is the machine-visible record of execution. Surface makes agent action visible through OS execution artifacts.
| Term | Description |
|---|---|
| Witness node | A Surface Laptop 5 running Linux, observing execution at the machine layer |
| Telemetry | Process, file, and network traces (osquery, auditd, journald, OpenTelemetry) |
Enhancement layer that adds evaluation and adversarial cross-model review to Arc.
Builder model's (LLM1) outputs are scored by the evaluations harness, which is served via the intermediary like any other tool. A reviewer model (LLM2), running on different weights and intelligence than the builder (LLM1), reviews the work and produces a justification that folds into the Debrief.
| Term | Description |
|---|---|
| Evals | Output performance and behavior measurements |
| LLM2 | Adversarial cross-model reviewer |
Interpretability research and experimentation inside LLM1 and LLM2.
Everything to this point has been engineering around the model. Sector goes inside the model, applying interpretability tooling to open instrumentable models and using what's learned to inform training pipelines.
Sector will run various experiments to explore the effects that various combinations of variables from across Radii, Arc, and Shell have on agent performance.
| Term | Description |
|---|---|
| Interpretability | TransformerLens, Circuit Tracer, SAELens, Neuronpedia, HeadVis |
| Weights | Open instrumentable models (Qwen, Gemma, GPT-OSS), Hugging Face |
| Training | Interpretability-informed training, RLHF (reinforcement learning from human feedback) |
Empirical limit of open-model performance against state-of-the-art.
Limit takes what's been built across Radii, Arc, and Shell, combines it with what's been learned in Sector, and then measures how close open models can come to state-of-the-art models on the same work, and where gaps remain.
| Term | Description |
|---|---|
| Benchmarks | Open vs state-of-the-art performance measurements |