English | 中文

Med Auto Science

An AI research agent for real medical studies, built to turn data, evidence, and drafts into manuscript-ready work

Disease Studies · Evidence Building · Analysis Support · Manuscript Delivery

Who It Serves Clinicians, PIs, and medical research teams working with disease-specific data and moving studies toward manuscripts

What It Organizes Study questions, data assets, analysis progress, evidence gaps, and manuscript-facing files inside one governed workspace

How To Start Tell it the disease area, the dataset you have, the question you want to answer, and the paper outcome you want

Med Auto Science is for teams already doing real medical research. It keeps study questions, data, analyses, evidence, drafts, and delivery files connected on one governed study line so the work can keep moving and stay reviewable.

Why Med Auto Science

The hard part of medical research is rarely a single paragraph. It is moving a study from data and ideas toward a paper that can be reviewed, revised, and submitted.

Medical teams often run into the same problems:

• You have data, but it is not clear which question is worth pursuing.
• You have preliminary results, but they do not yet form one manuscript line.
• Figures, drafts, analysis notes, and validation results drift across folders and conversations.
• Review, extra analysis, claim revision, and delivery files overlap until progress is hard to explain.
• Multiple disease studies run in parallel, and key evidence or decisions can be lost.

Med Auto Science is built around those research problems.

It organizes a medical study as a research line: frame the valuable question, prepare the data and evidence, advance analysis and validation, build the manuscript story, and keep paper-facing files ready for review and delivery.

It does not treat medical research as a rigid pipeline. A study can start with several possible directions, return to data and evidence for comparison, and gradually converge into a clearer manuscript line. AI keeps the work moving, organized, and revised; researchers keep clinical framing, claim acceptance, and final submission judgment.

One-Sentence Quick Start

You can start with prompts like:

• "Help me find a paper-worthy question from this colorectal cancer dataset, tell me what evidence is still missing, and propose the next step."
• "I already have preliminary results. Turn them into one manuscript line and tell me what validation to do next."
• "Keep pushing this disease study toward a publishable paper, and keep the progress plus files organized as we go."

Core Highlights

Identify paper-worthy questions from data Start with a disease cohort, registry, or real-world dataset, then identify questions with clinical value, evidence support, and manuscript potential before piling up analyses.

Turn scattered results into a manuscript line Existing analyses, early findings, figures, and drafts are organized into a clearer research story with explicit next evidence steps.

Keep progress and delivery files together Tasks, files, figures, drafts, validation notes, and deliverables stay tied to the same study workspace so the line remains reviewable and easy to continue.

Let AI do the heavy lifting while researchers keep judgment AI can help prepare data, run analyses, organize evidence, and report progress. Clinical framing, claim acceptance, and final submission decisions stay with researchers and PIs. Compare, revise, and review repeatedly Medical papers do not finish in one generation. The system can keep multiple claims, evidence gaps, analysis routes, and review findings on the same research line, then keep producing the next more-reviewable manuscript and evidence package.

What It Helps With

• Finding a study question worth continuing from a disease-specific dataset, registry, or cohort.
• Turning existing analyses and early results into one manuscript line.
• Managing validation, subgroup analysis, calibration, clinical utility analysis, and other supporting evidence.
• Keeping multiple related studies organized in one workspace.
• Keeping paper-facing results, figures, drafts, and delivery files tied to their study.
• Comparing study claims, evidence gaps, analysis routes, and review findings inside the same study stage, then producing the next manuscript and evidence package.

Current Position And Boundary

• Med Auto Science is the medical research Foundry Agent for turning disease data, study questions, evidence, and manuscript work into one governed research line.
• In the OPL family, MAS is an OPL Package(kind=agent): MAS retains medical-domain authority, while OPL owns generic runtime and hosted surfaces. Package identity, capabilities, dependencies, research work items, and typed views remain independent of any one carrier or executor.
• It can be used as the Research Foundry inside One Person Lab, and it can also be called directly by Codex or another agent through stable capability entries.
• Codex is the current first-class path because it offers the best implementation and maintenance economics. Its Plugin is a carrier projection and its CLI is an executor; neither is MAS package identity or the complete installed package.
• MAS owns the medical work itself: study questions, evidence organization, manuscript direction, manuscript quality, and delivery materials. One Person Lab handles hosted runtime, progress display, recovery/retry, and the cross-agent product entry.
• Manuscript quality is governed by study charters, evidence ledgers, review records, AI reviewer workflow, publication gates, and controller records. Status panels and script checks provide supporting evidence.
• Clinical framing, claim acceptance, and final submission decisions stay with researchers and PIs.
• Journal submission and external system interaction stay under human supervision.

Technical boundary for operators

• Med Auto Science is a medical research domain agent and Foundry Agent. It can be called directly by Codex, and it can also be discovered and hosted as an OPL-compatible package under OPL Framework.
• MAS owns the medical work itself: study intake, workspace context, evidence progression, progress explanation, manuscript quality judgment, runtime-facing owner receipts/projections, artifact authority, and manuscript-facing delivery.
• OPL Framework is the upper stage-led framework. It owns the generic runtime platform: stage attempts, queues, wakeups, recovery, approvals, receipts, state-machine execution, and cross-domain projection. MAS keeps medical conclusions, manuscript quality, domain transition semantics, artifact authority, and submission-facing judgment.
• MAS-local schedulers, runners, session stores, status shells, installers, and workbench wrappers are retired or tombstone/provenance only. New MAS program surfaces must justify a medical authority role and return owner receipts, typed blockers, domain refs, or safe action refs.
• In the OPL framework, a Stage is a large task step such as scouting, analysis, writing, reviewer repair, or delivery. An Agent executor is the minimum execution unit inside a stage; Codex CLI is the current first-class executor.
• A MAS stage pack gives the executor a goal, context, authority boundary, available affordances, knowledge refs, and quality gate. During the attempt, the executor decides what to read first, which tools to call, whether to run in parallel, whether to generate multiple candidates, and when to route back or request a reviewer; OPL route orchestration does not pre-script that reasoning.
• Candidate generation, reflection, review, and meta-review are Stage-internal execution strategies and evidence refs. They can guide exploration and review pressure, but they do not become a hardcoded workflow and cannot close a Stage without a MAS owner receipt, stable typed blocker, or independent reviewer/auditor receipt where quality is at stake.
• MAS tool declarations follow a Tool Affordance Boundary: they declare capability, permission, credential boundary, write scope, side effects, forbidden authority, and evidence entry points. They do not freeze the executor's literature reading, statistical checks, candidate generation, route comparison, or question-asking order into an out-of-prompt workflow script.
• MAS has completed monolith closeout. MedDeepScientist / DeepScientist remains available as provenance, explicit archive import, backend audit, upstream learning, and parity reference.
• Long-running OPL-hosted production execution is Temporal-backed. Temporal is the required production provider for OPL durable stage attempts, signal/query, retry/dead-letter, and workflow history. Hermes-Agent is not the target session/wakeup substrate, but it remains available as an explicit Agent executor adapter / proof lane that promises connectivity and auditability, not behavior or quality equivalence with Codex CLI.

How To Read This Repository

1. Potential users and medical experts should start here, then continue to the Docs Guide.
2. Technical readers and planners should read Project, Status, Architecture, Invariants, and Decisions.
3. Developers and maintainers should continue from the Docs Guide into docs/active/, docs/runtime/, docs/delivery/, docs/references/, and docs/policies/.

For Codex / Agents

This repository includes a Codex Plugin marketplace for the existing med-autoscience carrier. From a clone of this repository, add the marketplace, inspect the available entry, install it, and verify the installed snapshot:

repo_root="$(git rev-parse --show-toplevel)"
codex plugin marketplace add "${repo_root}"
codex plugin marketplace list
codex plugin list --marketplace med-autoscience --available --json
codex plugin add med-autoscience@med-autoscience
codex plugin list --marketplace med-autoscience --json

In the Codex App, open Plugins, select Med Auto Science, and install it. Start a new task after installation so the App or CLI loads the installed plugin snapshot. To remove the carrier and its marketplace configuration:

codex plugin remove med-autoscience@med-autoscience
codex plugin marketplace remove med-autoscience

These commands prove only that Codex can discover, install, and load the Plugin carrier. The OPL Package id remains mas; the Codex Plugin id remains med-autoscience. A carrier install does not prove that the complete OPL Package, managed runtime, study workspace, or publication surface is ready, and it does not make the required mas-scholar-skills dependency optional.

Use OPL-owned readbacks for the complete Package and hosted work-item state:

opl packages status --package-id mas --json
opl packages status --package-id mas --scope workspace --target-workspace <absolute_path> --json
opl workspace work-item readback --workspace <absolute_path> --work-item <id> --agent mas --profile full

Those readbacks report current Package, required-dependency, carrier, runtime, and hosted work-item evidence. They do not create a MAS owner receipt, accept a medical claim, authorize publication, or turn legacy transaction/receipt fields into current MAS authority.

Agent And Operator Quick Start

Start here if you are handing this repo to Codex or another agent

• Install, update, or remove the MAS package with opl packages install mas, opl packages update mas, or opl packages uninstall mas. MAS requires mas-scholar-skills; ordinary readiness checks that its identity exists and required capabilities are callable. Missing capability blocks MAS and routes to managed install/repair, but does not block unrelated packages and cannot be treated as optional.
• The target distribution is independently owner-published complete MAS bytes in its own GHCR latest-stable; Codex Plugin materialization is only one carrier projection. Current machine contracts and readbacks may still expose legacy version-range, ABI, lock, payload, digest, atomic-closure, receipt, or shared Release Set fields while migration is in progress; these fields do not prove the target model is implemented.
• The current compatibility label "required mas-scholar-skills dependency closure" means the required identity/callability edge during migration; it does not make version solving, locking, or an atomic cross-package closure part of the target readiness model.
• Cloning this repository provides the MAS declarative pack, not an OPL runtime installation. OPL discovers contracts/domain_descriptor.json, compiles six public Stage actions, and uses the closed registry to host five internal authority actions plus the self-evolution closeout callable. OPL generates the CLI/MCP/Skill/product surfaces and hosts runtime/workbench behavior.
• The canonical domain id is mas. med-autoscience is only the repository, package, and plugin locator.
• The repo-root agent/ pack and action schemas are the interface source. Public execution uses six OPL-hosted Stage actions. Qualification work-item provisioning, lifecycle reactivation, candidate admission, build-dependency currentness, and paper-mission authority evaluation are internal registry-bound MAS callables, not user commands.
• Runtime environments are prepared by OPL from contracts/runtime_environment_requirements.json; MAS does not install Python/R packages or plugins from its import path or workspace bootstrap.
• The only current stage-route authority is Codex CLI selected declared stage. StageOutcome, legacy NextActionEnvelope, queue, attempt, provider, status, and workbench state are nonbinding context or observation; they cannot select the next stage or prove paper progress.
• MAS may be used through its Codex skill or an OPL-hosted interface. Both routes return to the same MAS medical truth, quality, publication, and artifact authority surfaces.
• The V2 public/default command cutover is structurally landed. Legacy internal domain_entry, status/read-model, and queue-oriented callers still require OPL primitive parity, source-closure proof, caller migration, and physical-retirement authorization. Runtime, paper-line, publication, and production readiness remain separate live-evidence claims.
• Read Bootstrap, Architecture, and Status before changing package or runtime boundaries.

Further Reading

• MAS Whitepaper (HTML)
• MAS Whitepaper (PDF)
• Docs Guide
• Project
• Status
• Architecture
• Invariants
• Decisions