Project ProtoStar wants science to be AI-ready. Keep the provenance attached

THE SIGNAL IN ONE SENTENCE

Public science is not automatically reusable science. A dataset can be online and still be nearly impossible to understand. The file may be separated from its collection method. The labels may have been changed without a record. The license may be vague. The code may expect a vanished dependency. The paper may describe one version while the download quietly serves another. Now put an AI system in front of that mess. It will not become cleaner because the model can read quickly. It may simply become faster at combining mistakes. Project ProtoStar is a new effort to make publicly funded scientific datasets, models, code and documentation easier to discover, reuse and build upon. Hugging Face announced the initiative on October 29 with universities, research organizations and open-science partners. The founding university group includes Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University and George Washington University through the Open Forum for AI. Astera Institute, Hugging Face and the Open Research Community Accelerator are also listed as founding or hosting partners. Participating organizations sign non-binding, non-exclusive memoranda supporting the release of suitable public research in AI-ready formats on open infrastructure. The plan has two connected parts. First, help institutions publish research artifacts in machine-readable forms with useful documentation, licensing, versioning and preservation practices. Second, run public challenges that test whether those artifacts can actually be discovered, combined and extended. The inaugural challenge is hosted by Carnegie Mellon and runs November 2 through November 13. It asks participants to use historical weather records for 210 U.S. cities to forecast urban climate indicators and propose a climate adaptation idea. The official project page says teams will use data through 2018 to forecast 2019 through 2025. The launch article says 2019 through 2026 and describes final results after complete 2026 records become available. Those two first-party descriptions disagree about the forecast window. That is not a reason to dismiss the project. It is a tiny demonstration of the exact problem the project needs to solve. When a date range changes, which page is authoritative? When did it change? Which participants saw which version? Which results can be compared fairly? Provenance is how a scientific system answers those questions without relying on somebody's browser history. The useful object is not just the CSV, checkpoint, notebook or PDF. It is the artifact plus a receipt. The receipt begins with identity. Give every release a durable identifier. Record the institution, research group, maintainers, creation date, publication date and version. If the artifact is updated, do not replace the old version without a trail. A link that always points to "latest" is convenient for browsing and terrible for reproducing a result six months later. The receipt should preserve what a researcher actually used. Next comes collection. For a dataset, explain where the observations came from, who or what produced them, when they were collected, which instruments or systems were involved and what populations, locations or conditions are represented. If people contributed information, record the consent basis, privacy protections, access restrictions and intended uses. If a sensor generated the data, record the calibration, units, known failure modes and missing periods. If a model created synthetic examples, say which model, prompt, sampling settings and filtering process were used. "Public" is a distribution status. It is not a collection method. Then document transformation. Scientific data rarely travels straight from instrument to model. It is cleaned, joined, normalized, filtered, imputed, deduplicated, aggregated and relabeled. Outliers may be removed. Categories may be collapsed. Personally identifying details may be stripped. Images may be resized. Text may be translated. Every transformation changes what the artifact can support. Record the code, parameters, input versions and output checksums for each meaningful step. Preserve the rejected records or at least the reason and count for rejection when policy allows. If a team cannot reconstruct the pipeline, the final file is an orphan wearing a lab coat. Licensing belongs inside the receipt, not in a distant institutional policy page. State who owns the underlying material, which license applies, whether commercial use is allowed, whether derivatives must carry the same terms and how people should cite the work. One collection can contain layers with different rights. A public-domain government measurement, a copyrighted annotation guide and a model checkpoint built from restricted data do not become one simple license because they share a folder. Machines need those distinctions in structured fields. Humans need them in plain language. Then add scope and limitations. What scientific question was the artifact built to answer? What question does it not answer? Which geographic regions, species, languages, communities, devices or time periods are missing? What bias was measured? What bias was suspected but not measured? Which labels depend on human judgment? Which values are estimates? Which measurements fall below a detection threshold? An AI-ready artifact should be easier for a model to reject when it is the wrong evidence. That may sound like a strange goal for a reuse project. It is a sign of maturity. The dangerous version of AI readiness means formatting every file so a model can ingest it. The useful version means formatting the evidence and the boundaries so a model can understand when not to use it. ProtoStar's challenge model can help test that difference. A challenge should publish the input artifacts, baseline methods, evaluation metric, submission format, environment and scoring code. It should record when rules change and which submissions were evaluated under which version. For the climate challenge, simple historical averages and linear trends provide useful baselines. A complicated model should not receive applause for losing to a ruler. But leaderboard accuracy is only one layer. Participants also propose a policy idea informed by the forecast. That step moves from prediction to public consequence. The receipt should therefore connect the recommendation back to the exact forecast, uncertainty range, data limitations and assumptions that support it. A city should not receive an adaptation suggestion that has become detached from the weather evidence that produced it. The project page says participating organizations keep control over which materials are appropriate to release and how they are shared. That is important. Open science does not mean every dataset belongs on the open internet. Some artifacts contain sensitive health data, Indigenous knowledge, endangered-species locations, security-relevant details or contractual restrictions. Responsible release may require aggregation, controlled access, community governance, delayed publication or no publication at all. The provenance receipt should record the access decision and its authority. Do not treat a closed artifact as a failure if closing it is part of ethical stewardship. AI readiness also needs maintenance. Name the steward. Publish an issue channel. State how long the artifact will be hosted. Explain what happens when an error is found. Keep a changelog. Mark withdrawn versions without erasing the record that they existed. If a dataset is corrected after a paper or model uses it, notify downstream maintainers when possible. Open infrastructure can make the chain visible, but visibility has to be designed. The same applies to citations. When an AI system retrieves an artifact, the output should retain the durable identifier, version, creators and license. When it combines several sources, each contribution should remain separable. A generated summary that cites only the hosting platform has flattened the work of the scientists who collected and curated the evidence. Make the citation travel with the file. Then test the package with a cold reader. Give the artifact to a researcher outside the original team. Ask them to reproduce a basic result, identify the license, explain the limitations and locate the transformation history without private guidance. Give the same package to an AI system in a controlled evaluation. Ask it which questions the artifact can answer, which it cannot, what version it used and how it would cite the source. Score the failures. If the human and model both guess the units incorrectly, the problem is not their curiosity. It is the package. ProtoStar is still an initiative at launch, not proof that public science has become interoperable. The official page currently lists one public dataset collection for the signed memoranda and no public models. The first challenge has not yet produced results. That is a reasonable beginning. The project is explicitly trying to learn which technical barriers, documentation gaps, institutional incentives and infrastructure choices make reuse succeed or fail. The useful output will be more than a count of uploaded datasets. It will be evidence about whether another team could understand, reproduce, combine and extend the work. Hugging Face reports more than 17,000 public datasets across core science domains on its platform, with the figure approaching 29,000 when medicine and mathematics are included. It also reports that researchers shared more than three times as many new science datasets during the first nine months of 2026 as during the same period in 2025. Those are platform-reported counts. They show activity, not scientific quality. Downloads show interest, not validity. An immaculate model card cannot rescue a bad experiment. The provenance receipt should never become decorative paperwork that gives weak evidence a professional border. Scientific review still matters. Domain repositories still matter. Ethics review, community consent, statistical analysis and independent replication still matter. AI-ready packaging can help those systems travel. It cannot replace them. That is why the strongest version of ProtoStar would publish a reusable artifact standard alongside real failure reports. Show the datasets that could not be combined. Show the license conflict that stopped reuse. Show the missing unit that changed a result. Show the model that ignored a limitation. Show the challenge rule that needed correction. Then update the guidance. Public research should be easier to put to work. The work should still arrive with its memory intact.