# Nexus-PRIME Labs

> Nexus-PRIME Labs is a working AI lab for custom, review-heavy workflows. It defines what AI may prepare, what people must decide, and what each handoff in scope must record.

Official site: https://labs.nexus-prime.cfd/
Contact: pilots@nexus-prime.cfd

## What the Lab works on

- Agent continuity: shared task memory, bounded scope, pending checks, and handoff state across coding-agent tools and sessions.
- Context efficiency: token-aware reads, focused unfolding, fixture-measured reduction, and visible fallback.
- Bounded workflow execution: an operating record with its owner, checkpoints, exception path, approval, and next review.

The working method is Observe, Bound, Prove: identify the repeated failure and owner; separate stable steps from accountable decisions; keep a record another person can review.

## Custom workflow creation

Workflow review defines the trigger, failure point, accountable owner, safe AI-run steps, human decision, exception path, review evidence, and one measure before architecture or automation expands.

The site's WorkflowBench deterministically assembles visitor inputs into a Draft Workflow Brief. It does not run a model, assess feasibility, score readiness, or validate a workflow. Product delivery, regulated review, and case operations are editable templates, not customer deployments.

The pilot intake opens a draft in the visitor's email client. Nothing is sent by the form or sent until the visitor reviews and sends the email. Visitors are told not to include credentials or sensitive records.

## Systems

- Nexus Prime: An MCP control plane and orchestrator for coding agents. It provides shared memory, bounded execution, handoffs, and a reviewable operating record.
  Product site: https://nexus-prime.cfd/
  Install: npm i -g nexus-prime

- tetris.codes: Context compression and token-aware inspection for coding agents, with visible fallback when input should stay whole.
  Product site: https://tetris.codes/
  Install: curl -fsSL https://get.tetris.codes | sh

- Grain: A private system for defining the information, checks, and handoff record a workflow requires.

- NXL: A private system for making repeatable agent work observable and replayable.

- Phantom: A private system for keeping multi-step work tied to its goal, owner, checkpoints, and review history.

Grain, NXL, and Phantom are in private development. Fit and availability depend on the workflow and must be confirmed with Nexus-PRIME Labs.

## Benchmark and evidence status

- tetris.codes fixture result, verified 27 July 2026: on a deterministic generated large-Rust fixture, Atlas reduced estimated input for a one-function task from 6,003 tokens for a full read to 999 for an outline. A small file remained whole, a generated bundle was denied in strict mode, and the tested security-sensitive input was kept verbatim. Fourteen fixture checks passed. These are fixture states, not an end-to-end task-quality, provider-latency, or universal security claim.
- Nexus Prime current-release result: not published yet. The existing benchmark artifact targets an older release and lacks enough environment and repeated-run detail to present as current. The validation protocol covers token planning, memory store and recall, retrieval quality, bootstrap context, environment, repeated runs, variance, and failure cases.
- AdaptiveKVMerge: research into how models can carry less context without losing the information a task needs.

A performance result is published only when its baseline, fixture or dataset, version, method, controls, and limits are stated. Hardware- or provider-dependent claims also require environment details and repeated-run variance.

## Working with the lab

- Install Nexus Prime or tetris.codes through their open product paths.
- Review one recurring workflow with one measure and agreed operating boundaries. Timing follows the workflow, access, and approval requirements.
- Build a custom workflow system with the memory, context selection, routing, checkpoints, approvals, and review evidence the operation requires.

## Data and site terms

WorkflowBench runs in the visitor's browser and does not send its free-text workflow description to Nexus-PRIME Labs. The pilot intake prepares an email draft; nothing is submitted by the site. Visitors should not enter credentials, personal data, medical records, financial records, or other sensitive information.

- Privacy policy: https://labs.nexus-prime.cfd/privacy/
- Terms and conditions: https://labs.nexus-prime.cfd/terms/
- Nexus Prime product privacy policy: https://nexus-prime.cfd/privacy
- Nexus Prime product terms: https://nexus-prime.cfd/terms

## Domain workbench

These are application patterns, not claims of customer deployments.

- Product delivery: engineering and incident workflows; AI may carry context, prepare bounded changes, and run agreed checks while people retain architecture, code review, incident judgment, and release approval.
- Regulated review: healthcare administration, financial services, and public systems; AI may assemble evidence and route exceptions while qualified people retain clinical, regulated, eligibility, and policy decisions.
- Case operations: customer, vendor, and exception workflows; AI may assemble context and route ownership while people retain policy, tone, commitments, exceptions, and escalations.

## Canonical pages

- Home: https://labs.nexus-prime.cfd/
- Systems catalog: https://labs.nexus-prime.cfd/experiments/
- Privacy policy: https://labs.nexus-prime.cfd/privacy/
- Terms and conditions: https://labs.nexus-prime.cfd/terms/