boundary.labs

Boundary Labs / Collaborate

A concrete question.
A useful public result.

Partner on persistent memory, practical inference, or operational knowledge. Start with a scoped experiment, agree on the evidence, and publish what it teaches us.

Discuss a research partnership →See the proposals

What is already here

An existing research foundation.

Explore the dated benchmark record, persistent-memory architecture, and operational-memory prototype. The Zenodo paper is a preprint, not a peer-reviewed publication.

Production backend: GPT-5.6 Luna (OpenAI API)
Verified 2026-09-06. Agent orchestration and memory run locally; production inference uses an external API through a shared gateway. This is a dated verification, not a live status feed.

The former dual-GPU tower was retired July 20, 2026. Its results are historical evidence, not a claim about hardware currently available to this lab.

Proposed collaborations

Choose the question to investigate.

These are starting scopes. Timing begins after access, resources, and the evaluation plan are agreed. No result or performance gain is promised in advance.

01 / Memory · proposed 30 days

Does memory survive a model change?

Question: How does recall change when the same memory is used by different models?

Resource requested: Partner-provided model access or evaluation credits for an agreed test matrix.

Deliverable: A comparison report, reproducible configuration, and failure analysis with sample sizes and test conditions.

Publication: Public aggregate results and permitted reproduction artifacts; credentials and private evaluation items remain excluded.

Discuss the memory study →
02 / Inference · proposed 60 days

Which deployment fits the workload?

Question: How do accessible hardware and API inference compare on task quality, latency, and cost?

Resource requested: A hardware loan or compute allocation with documented specifications and usage terms.

Deliverable: A matched-workload evaluation, configuration notes, and a deployment guide covering failures as well as wins.

Publication: Public methodology and results with sponsorship disclosed. No endorsement or favorable outcome is guaranteed.

Discuss hardware or compute support →
03 / Context Farm · proposed 30-day pilot

Can a team retrieve the right exception?

Question: Can one bounded workflow produce useful answers with the governing rule, exception, and source?

Resource requested: A design partner, a small approved or synthetic document set, and scheduled feedback from a domain expert.

Deliverable: A worked prototype, a reviewed question set, and a report on retrieval errors and review effort.

Publication: Only synthetic or explicitly approved examples. Agree on storage and inference boundaries before sharing operational data.

Discuss a design partnership →

How we work

Scope first.
Evidence throughout.

  1. Define the question. Agree on the workload, success criteria, resources, and data boundaries.
  2. Run the study. Record conditions and inspect unexpected results.
  3. Deliver the evidence. Share the findings, limitations, and agreed artifacts.

Start a conversation

What are you trying to find out?

Send the problem, available resources, and intended timeline. For consulting, include the decision you need the work to support. Scope, fees, and commitments are agreed before work begins.

Email Dino →