AIPOCH has released Open-Science v1.0.0, an open-source AI research workbench built around portable “.science” packages. The AI-assisted research tool is local-first and model-agnostic, runs on macOS, Windows and Linux under the Apache 2.0 licence, and keeps a research question, its files, its code execution and its outputs in one inspectable workspace. A.science package carries selected conversation branches, file versions, notebook records and verification evidence, so a colleague receiving a result can see how it came about rather than just what it says. A preprint accompanies the release, and the code is public on GitHub.
The workbench ships with 23 built-in Skills, 525 Skills in a marketplace, 36 scientific connectors, 10 Specialists and four agent backends — Claude Code, OpenCode, Codex and CodeBuddy. It runs Python, R and shell workflows locally or on configured SSH hosts and Slurm clusters, and it keeps immutable artifact versions with available producer code, inputs, execution history and environment details. For artifact-level export it uses RO-Crate, the community standard that describes research data, code, authors and licences as one exchangeable “research object.”
AI-assisted research: Why the problem is real
The tool targets a genuine and well-documented gap. Research in Patterns (Cell Press) argues that reproducing a computational result requires preserving data, code and runtime, and that an LLM placed “on the data path” becomes part of the runtime — a dependency that may change, fall out of support or become inaccessible. The European Commission’s guidance tells researchers they remain accountable for AI-assisted output and should disclose tool use and consider non-determinism. In other words, the polished answer hides the path, and the path is what makes work checkable.
The strongest support comes from a 2026 PNAS study of open-source “AI scientist” systems. It found systematic methodological pitfalls — inappropriate benchmark selection, undocumented data manipulation, post hoc selection bias — that are largely invisible in the final manuscript. An LLM-based auditor given only the manuscript performed barely better than chance (53.7% accuracy); given full execution records, accuracy rose to 81.5%. The authors recommend that journals move beyond manuscript-only evaluation to process auditing. AIPOCH’s.science packages are, in effect, a productised answer to that recommendation.
What the package can and can’t do
AIPOCH is unusually candid about the limits, which is worth noting in a category full of overclaiming. An imported.science package opens read-only; it does not execute code, call a model or restore credentials, and a received verification record reports the sender’s checks rather than proving the recipient reran them. Re-executing a result still requires a complete recipe, the required inputs and a usable runtime. The release notes state plainly that replay “does not guarantee deterministic reproduction of a whole session or establish scientific validity.”
That honesty matters because packaging is not the same as reproduction. A session record improves inspection and handoff, which is exactly what the PNAS evidence says helps auditors catch problems; it does not solve the deeper issue that a vendor may retire a closed model or that the same inputs may yield different outputs. The framing in the release — that.science “may become a new paradigm in the agent era” and that the approach “has been validated by many of the participating institutions” — outruns the evidence. The institutions listed are affiliations of the paper’s authors, not adopters of the tool.
What to watch
Three things will show whether Open-Science matters. First, adoption: whether researchers outside the founding group use.science packages for real handoffs, since a format is only as useful as the community that reads it. Second, whether the preprint survives peer review, given that it shares an author with the tool it describes. Third, whether the broader ecosystem — journals, funders and institutions — adopts process-auditing norms, because a packaging standard only changes practice if reviewers ask for it.
What the launch does and doesn’t establish
It establishes a v1.0.0 open-source release, a portable session-package format, an RO-Crate export, a catalogue of Skills, connectors and agent backends, and clearly stated limits. It does not establish adoption, independent validation of the tool, peer-reviewed support for its claims, or that packaging solves AI-assisted research’s reproducibility problem.
The honest summary: AIPOCH Open-Science is a serious, well-scoped attempt at the provenance half of AI-assisted research, and it arrives alongside independent evidence that process artifacts are exactly what make automated research auditable. Its candour about what a package cannot prove is a point in its favour. But a session record is a record, not a reproduction, and the grander framing outruns a v1.0.0 release from a single vendor. The value will turn on whether researchers use.science packages in practice — and whether the institutions reviewing their work start asking to see them.
Editor’s Note: This article draws on the AIPOCH release and the company’s published materials, the GitHub v1.0.0 release, and the arXiv preprint, plus independent research by Patterns (Cell Press), PNAS, the European Commission and the RO-Crate community. The tool’s features, catalogue and stated limits are as described by its developers; adoption is not independently verified, the preprint is not peer-reviewed, and the “validated by institutions” claim is qualified. No interview was conducted.