Research / Safety & Alignment

We work on AI safety
at the runtime.

The model is one ingredient. The runtime around it is the product. Our safety work targets the layer where an AI agent meets your filesystem, your shell, and your network.

All research Trust Center

Approach

Three commitments that govern every release.

Approval-gated runtime

The model is treated as untrusted text until a human authorises an action. This is the security boundary that ships in every Dropstone tier and the foundation we build the rest of our safety work on.

Honest evaluation

We publish the Joule Index as an open benchmark for normalising coding-agent capability per dollar. The methodology is reproducible against published provider quotas.

Provenance disclosure

We say which open-weight model serves each tier and refresh the choice monthly. The runtime is the safety claim; the model is an implementation detail we replace when something stronger ships.

Current research

What we are working on now.

n = 14

Active

The Dictionary at the End of the Wire

On the gap between knowledge access and knowledge creation in the LLM era

Read

Active

Dropstone 1.5: Technical Report

Inference, Pricing, and Runtime Safety Architecture for a Versioned, Model-Agnostic Coding Agent

Read

Active

The Joule Index: a benchmark for what intelligence costs

Today we are releasing the Joule Index, the first AI benchmark that publishes the dollar cost, joule cost, and human merge-readiness of frontier coding agents on the same chart. In our preview release, eight agent tiers from four vendors produce the same merged diff on three real open-source bugs at a twelve-and-a-half-fold cost spread.

Read

Active

Introducing Hope

A post-transformer research initiative on general intelligence at low compute. Hope-1 combines discrete-latent program codes, verifier-driven search, and a pre-registered seven-rung research framework.

Read

Active

A Universal Bimodal Drift-Rate Ratio in Repeating Fast Radio Bursts

A pre-registered cross-source drift-rate ratio in repeating FRBs. CoV 3.8% across 4 sources. F4 null Monte Carlo p ≤ 5 × 10⁻⁴.

Read

Active

Discovery of Bimodal Drift Rate Structure in Fast Radio Bursts

We analysed 978 cosmic radio signals from a single source. Within it, the upward-drifting population splits into two statistically distinct groups.

Read

Active

Hunting for Real Wormholes: We Searched Every Available Dataset

We fed every publicly available astronomical dataset to Primus and asked one question: is there a wormhole hiding in plain sight?

Read

Active

Introducing Primus

A reasoning layer on top of frontier large language models that carries scientific investigations from a posed question to a verified result. Version 0.2 conducts investigations end-to-end. It is not autonomous, but it is closer to autonomy than any prior system we are aware of.

Read

Active

The 50-Year-Old Algorithm We're Trying to Beat

For 50 years, no one has improved 3×3 matrix multiplication. Blankline Research reveals the mathematical barriers blocking progress on this open problem.

Read

Active

Project CELSIUS: Rethinking AI Infrastructure from the Ocean Floor Up

A Bankline research initiative to overcome the thermodynamic limits of AI scaling through deep-ocean infrastructure, developed in collaboration with domain experts and advanced AI modeling.

Read

Active

Beyond Retrieval-Augmented Generation: How We Solved the Temporal Event Horizon Problem

This research introduces the D3 Adaptive Memory Architecture, which solves the Temporal Event Horizon failure mode in standard RAG systems where data older than six months becomes statistically unretrievable.

Read

Active

Dropstone: Overcoming the Linearity Barrier via Recursive Swarm Intelligence

The Dropstone D3 Engine utilizes a recursive swarm architecture to replace linear prompting with distributed, self-correcting agents, solving context saturation in long-horizon engineering tasks.

Read