Arc Intelligence

Arc's intellectual output — research papers, technical notes, white papers, and essays. An archive, not a marketing feed.

All items — 10

Research PaperMay 2026

Recoverable Provenance: Making Retrieved Context Usable

A formal treatment of when a retrieved fragment can be acted on without its surrounding document — and a provenance scheme that keeps recovery usable, not merely findable.

hunter-ragRetrievalProvenance

Read

Research PaperApr 2026

Evaluation-First: Measuring Model and Agent Behaviour

An argument for building the measurement system before the product — defensible evaluation of model and agent behaviour as the precondition for trustworthy deployment.

lancetEvaluationAgents

Read

Thoughts

Rethinking Agent Skills, Tools, and Capabilities

When agents underperform, the instinct is to add more tools. Most of the time, that is the wrong diagnosis — and the real gap is in skills and governance, not capability.

June 2026

White Paper

Why Arc's Forward-Deployed Engineering Differs

Most AI deployment is implementation. The unsolved problem is engineering the boundary between the deterministic and the non-deterministic — and why the firms built to ship a model, wire a stack, or hand over advice cannot close it.

lancet

June 2026

Engineering Note

Found but Unusable: The Recoverable Chunk Problem

Retrieval often surfaces the right source and still fails the task, because the recovered fragment is not usable on its own. A note on recoverable provenance.

hunter-rag

May 2026

Thoughts

Why Arc Exists

A short reflection on building a private R&D house at the boundary, instead of chasing the nearest market signal.

May 2026

White Paper

Deep Technical Options: An Operating Model for Frontier R&D

A framework for treating R&D output as a portfolio of options rather than a pipeline of products.

hunter-rag

Apr 2026

White Paper

Local-First, Trust-Bound Systems

A white paper on a system class that keeps data and computation local by default, with trust boundaries as first-class architecture — for organisations that cannot externalise sensitive context.

cawsal

Mar 2026

Thoughts

The Arc Category System

Why Arc works through eight product and technology categories rather than a list of projects — an essay on structuring a portfolio of deep technical options.

Mar 2026

Engineering Note

Trust-Bound Generation on Sensitive Data

A note on governing generative systems that operate on sensitive organisational data — boundaries, gates, and controlled state transfer instead of post-hoc filtering.

cawsal

Feb 2026

Papers, notes, and technical artifacts.

Recoverable Provenance: Making Retrieved Context Usable

Evaluation-First: Measuring Model and Agent Behaviour

Rethinking Agent Skills, Tools, and Capabilities

Why Arc's Forward-Deployed Engineering Differs

Found but Unusable: The Recoverable Chunk Problem

Why Arc Exists

Deep Technical Options: An Operating Model for Frontier R&D

Local-First, Trust-Bound Systems

The Arc Category System

Trust-Bound Generation on Sensitive Data