Context systems
Preserve intent, trust-boundary instructions, and live evidence while compacting replaceable history.
Research directions
We study the systems around capable models: what preserves intent, constrains delegation, verifies behavior, and makes a completion claim credible.
Research areas
What we are testing
right now.
Working directions shaped by implementation pressure, not polished claims.
Verification
Connect completion claims to real checks, rendered behavior, diff review, and inspectable evidence.
Orchestration
Give specialist agents scope, budgets, durable state, and clear return contracts from one accountable run.
Model interfaces
Adapt prompt structure, edit dialect, and tools without changing the standard used to call work complete.
Current thesis
A capable model is not an accountable system
Capability can generate edits. Accountability requires one system to own context, tools, delegation, checks, browser state, and the decision to stop.
Explore the notesFig. 01 — context pressure across a long-running task
Working notes
Claims, methods,
open questions.
6 active entries. Status stays visible until the work earns a stronger label.
Systems
A capable model is not an accountable system
Capability can generate edits. Accountability requires one system to own context, tools, delegation, checks, browser state, and the decision to stop.
Context
Compaction across trust boundaries
A context engine should preserve instructions and current evidence verbatim while replacing only history that can safely be reconstructed.
Evaluation
Proof-carrying completion for coding agents
Completion should be a claim backed by declared checks, rendered behavior when relevant, and an inspectable evidence chain.
Orchestration
Bounded delegation with accountable parent runs
Specialist agents can fan out independent work while a parent retains budgets, role restrictions, durable state, and trace ownership.
Interfaces
Model-aware execution, model-independent proof
Execution should adapt to each model family's prompt and tool dialect without weakening the standard used to call work complete.
Verification
When does browser evidence change the answer?
For product work, source inspection and unit checks may be insufficient. We are mapping when rendered interaction becomes required evidence.