AI execution layer
Produces bounded work products such as searches, scripts, tables, checks, figures, summaries, and draft analyses.
Methodological project
AlphaScience: Claim Calibration for AI-Assisted Scientific Exploration
AlphaScience is personal methodological research on how AI-assisted scientific work can move through auditable evidence states before it becomes a manuscript-facing claim.
Summary
The object of study is not AI output, but claim state.
AlphaScience addresses a bottleneck that appears after AI systems generate hypotheses, code, figures, analyses, and manuscript-like arguments: how can we decide what those outputs are responsibly allowed to claim?
The project treats AI-assisted science as a routing problem. Work products may be useful, weak, ordinary, incomplete, negative, or misleading. The methodological question is how to keep those states visible instead of allowing a smooth narrative to promote them into stronger conclusions.
Problem
Premature promotion is a scientific failure mode.
A weak signal can become a candidate; a candidate can become a result; a result can be written as a discovery. This is not only a hallucination problem. It is an evidence-state problem: the route from output to claim can become too fast, too opaque, or too flattering.
AlphaScience therefore focuses on claim calibration before publication language: what the evidence supports, what it does not support, what remains missing, and which apparent stories should be stopped rather than refined.
Framework
Three roles are separated deliberately.
AI scientific-guidance layer
Reviews interpretation, claim boundaries, missing validation, paired non-claims, and reasons to reject a proposed story.
Human workflow supervisor
Controls routing, completion state, release decisions, authorship boundaries, and the final responsibility for scientific claims.
Evidence-state model
The durable unit is a manuscript-facing evidence object.
The core object is not a chat transcript. It is a bounded evidence object with state: draft ledger row, reviewed ledger row, and approved record. Each bounded claim is paired with a non-claim, making explicit what the evidence does not establish.
This structure is meant to preserve scientific judgement when AI-assisted workflows become fast enough to produce plausible narratives before the evidentiary boundary has been checked.
Current status
The public story is intentionally narrower than the internal work.
A manuscript-only draft has been tested through a private presubmission-style exchange with a selective scientific journal. The response was useful enough to justify preparing a full manuscript, but it is not a review decision and should not be read as external validation.
That distinction matters. The public version of AlphaScience is presented as a release-safe methodological story, not as a disclosure of unpublished case studies or a public archive of project materials.
Manuscript-only draft available upon request.
Release boundary
The page is designed to be informative without exposing unpublished work.
Public materials describe the claim-calibration problem, the evidence-ledger framing, and collaboration interests. They do not disclose:
- institutional project materials or internal communications;
- raw data, active repositories, source-proof packages, or runtime logs;
- AI transcripts, prompt trails, or private review records;
- case-specific biological claims or externally unvalidated discovery claims.
Collaboration interest
Useful pressure is more valuable than broad agreement.
I am interested in collaboration with researchers working on scientific agents, AI-for-science infrastructure, hypothesis-generation systems, AI evaluation, reproducible scientific workflows, research integrity, and human-AI collaboration.
The most useful conversations are bounded: what would count as a serious external check, what should be removed from a claim, and what evidence would make a story worth pursuing.