Flowers
An experimental personal-agent harness. A testbed for reliability ideas: it plans before acting, asks before anything irreversible, and checks its own work.
01
Guidance & memoryResearch
On SWE-bench Verified, repository guidance that is iteratively refined outperforms static context, and the accumulated guidance transfers across repositories.
PreprintarXiv
Jan–May 2026Probe-and-Refine Tuning of Repository Guidance for Coding Agents
A methodology for refining repo-level guidance via synthetic bug-fix probes; raises SWE-bench Verified resolve rate to 33.0% from a 25.5% unguided baseline over 4 trials, with gains driven by coverage rather than precision, plus evidence of cross-model degradation.
Read →Preprint
May 2026Cross-Repository Guidance for Coding Agents: A Single-Trial Replication Addendum
A companion to probe-and-refine: can comparable guidance be assembled without ever tuning on the target repo? Mining transferable edits from 50 external repositories yields guidance that resolves 45.4% of SWE-bench Verified vs. 39.6% baseline.
Read →02
Tools & systemsSystems
Aedos
Verified symbolic memory for LLMs. Extracts claims, routes them to verifiers, and accumulates a verified cache over time — a study in the soundness–coverage tradeoff.
npm
Aspect Code
Open-source CLI that generates and maintains repository guidance for AI coding agents — the tooling counterpart to our probe-and-refine research.
03
Autonomous researchExperiments
Active
Market Autoresearch
A multi-agent system that researches financial markets. It generates hypotheses, runs quantitative experiments against market data, and surfaces the findings that survive independent meta-review.
With an independent collaborator (markets).
Voynich Autoexploration
An autonomous multi-agent loop that probes the structure of the Voynich Manuscript end to end. It ran over 600 corpus-statistical experiments and over 5400 cipher-mechanism simulations, and it surfaced novel findings.
04
ExternalWriting
- A Model’s A Priori Is Frozen A PosterioriJun 27, 2026
- Emergent Misalignment is Co-occurrence Density Made VisibleApr 22, 2026
- What 40 Hours of Autonomous Research Taught Me About Long-Running AgentsApr 15, 2026
- Autonomous Agents on the Voynich ManuscriptApr 9, 2026
- Aspect Code is now on npmMar 31, 2026
- Oracle Tuning Results and the Path to Adaptive Context SelectionMar 16, 2026
- A New Direction: CLI-First and Iterative Context RefinementFeb 24, 2026
- Making AI Reliable on Real Codebases: The KB + Feedback LoopJan 3, 2026
- Benchmarking AI Codebase Context — From −24 to +25 Net TestsDec 10, 2025
- Aspect Code: Giving AI Assistants the Context They're MissingNov 30, 2025
05
Lab & collaboratorsPeople
Markets
Independent collaborator
Unnamed. Works on the markets-information side — not the AI autoresearch or the HFT implementation.
Aspect Research also draws on informal advisors across adjacent fields.
Contact
Get in touch: asa@aspectresearch.org
Our community Slack is invite-only — email us to join. Code and releases live on GitHub.