Anyone who has inherited a colleague's analysis knows the problem. The chart is there, the conclusion is there, and the story of how one became the other is somewhere else entirely. AIPOCH, the company behind the open-source Open-Science workbench, has built its answer to that gap around a single idea: the shareable .science package. Depending on the retained materials, a package can bring outputs together with conversation records, data versions and code that help explain the work behind them.
AIPOCH announced the formal public release of Open-Science on 14 October 2026. The workbench covers a lot of ground, but for a working researcher the part that changes daily habits is the package.
Why AI-assisted research gets scattered #
AIPOCH describes the starting problem plainly. In an AI-assisted research effort, task requirements live in the conversation, computations live in notebooks, results land in CSVs, charts or reports, and review comments correspond to some historical version of the work. Each piece sits in a different place, and the connection between them is the part most easily overlooked.
That missing link is context. The company's release lists the consequences: files and conversation histories get transferred separately, so the full research context is hard to review, hand off or build on. A polished answer or figure can also be difficult to judge when its source material, producing code, execution history or intermediate decisions are missing.
A .science package is meant to close that gap. It organizes the recorded content of a session into one portable research session package, so the work can be opened, inspected and continued on another computer.
Walmart Opens Its Store to AI Shopping Agents as Amazon Shuts Them Out Walmart's gateway lets Meta's Muse, ChatGPT, and Gemini buy groceries and goods directly, while a federal judge just ordered Perplexity's Comet to stop shopping on Amazon and delete the data it collected. - walmart ai shopping agents inventory search capability - how walmart agent gateway competes with amazon shopping
The airplane black box comparison #
AIPOCH asks researchers to think of a .science package as an airplane's black box. Opening it reveals the final charts and literature PDFs, and also, depending on what the researcher chooses to retain at export, a layered record of how they were produced.
The first layer contains selected conversation branches and recorded exchanges, showing how the researcher framed questions and how the AI’s responses developed. The second is snapshots of the input data and the version relationships of artifacts, which show which version of the data produced which chart. The third is notebook execution records and environment locks, covering which software versions were used and which lines of Python or R code were run. The fourth is review records with checksum verification, which in effect leave a verifiable digital fingerprint on the files.
For a researcher, each layer answers a question that usually has to be settled by email. What did you ask the model? Which data went into this figure? Which version of the software was running? Has anyone changed this file since it was reviewed?
What a handoff looks like #
When a colleague imports a package, the imported research history remains read-only. The recipient can inspect the retained conversation branches and available review records step by step.
Continuing the research is a deliberate second step. To build on the work, the recipient opens a separate writable copy, known as a fork. The earlier history and the later work are stored separately, so the original record stays intact while the new work grows beside it.
Several safeguards are built into the import. The history is presented read-only, code is not executed automatically, and the original author's access credentials are not restored. The release adds that imported records do not reuse the original external-service authorizations.
Choosing what to share #
Sharing is not all or nothing. At export time, researchers choose what to share. A package can contain selected conversation branches, file and artifact versions, notebook records, available provenance and verification evidence. That matters for anyone handling work in progress, since a package prepared for review need not carry every exploratory detour.
Nvidia and TCV back Higgsfield in a Series B extension as Jensen Huang visits Kazakhstan
Nvidia and TCV extended Higgsfield's Series B without disclosing new terms, building on an August round that valued the company at $5.4 billion. Higgsfield's annualized revenue run rate hit $700 million in August, up from $200 million in January. - Nvidia Series B funding for AI video startup - Jensen Huang Kazakhstan AI Month Higgsfield investment news
The release lists the situations the packages are meant to support: review, collaboration, handoff, continuation and long-term archiving. They can be shared with another researcher, moved to another computer, or imported as a read-only record.
AIPOCH is also careful about what it does not claim. Reproducing an original result still depends on whether the exported inputs, code, dependencies, runtime, execution records and any required external data are complete and still available. The package preserves the context. It cannot restore what was never exported.
Review records tied to the right version #
A .science package preserves available review and verification records alongside the artifact versions they concern. This helps recipients identify what was checked and which version the check applies to.
For collaborative work, that distinction matters: a check on an earlier version should not be mistaken for an assessment of a later revision. The retained records help researchers inspect the scope of previous checks while leaving scientific interpretation to their judgment.
The workbench behind the package #
The packages come from AIPOCH Open-Science, an open-source AI research workbench released under the Apache License 2.0. Project data is stored locally by default. The workbench supports multiple AI model providers, agent frameworks, compatible custom gateways and locally deployed models, so the package is not tied to one vendor.
Inside the workbench, users can describe a task in plain language and work with an AI agent that can run Python, R and shell workflows, subject to user permissions. Reports, tables, figures and notebooks are preserved as immutable versions, with available connections to inputs, producer code, execution history, environment details and conversation branches. Those connections are what a package carries out of the workbench.
The foundation is informed by open standards such as RO-Crate, which describes research data, code, authors, licenses and provenance together as a single exchangeable research object. AIPOCH says this gives the packages a portable, standards-based footing for sharing and archiving across projects and institutions.
Who is behind it #
A preprint describing AIPOCH Open-Science is available on arXiv (arxiv.org/abs/2610.17110). Beyond AIPOCH, the paper lists 76 author-affiliation entries spanning medical centers, universities and companies worldwide. They include medical centers such as Zhongshan Hospital of Fudan University, Amsterdam UMC and Beth Israel Deaconess Medical Center, universities such as Peking University, Shanghai Jiao Tong University and the Technion, and technology and biotechnology companies such as SenseTime and Amoy Diagnostics.
Huimei Wang, CEO of AIPOCH, framed the release in terms of access. "The era of Vibe Research is coming, where more people can turn their curiosity and research ideas into meaningful work," said Wang. "We hope to make scientific exploration more accessible and build an open ecosystem where researchers and developers can contribute to and benefit from one another's work."
The company positions the package as an answer to what it calls agentic data transportability, and suggests the model may become a new paradigm in the agent era. That is AIPOCH's own assessment, and the test will be whether researchers adopt the habit of exporting their sessions.
Why it matters for daily work #
For a researcher, the appeal is modest and practical. A package turns a handoff from a pile of attachments and a long explanatory message into a single object that can be opened, read and forked. A reviewer can check a figure against the data version that produced it. A student can study how an experienced colleague phrased and revised their questions, which is the idea behind the release's language about work that others can follow, learn from and share. It also changes what an archive means. A folder of outputs records what was found. A package records how it was found, including the questions, data versions and checks that sit between the outputs.
More about AIPOCH and Open-Science is available at aipoch.com, and the source code is published on GitHub under the aipoch organization.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.