Lab notebooks and test records are wanted by AI labs because they show how experimental work is actually done, step by step, including the runs that did not work. Published papers leave most of that out. A lab, test house or R&D group can license these records if it controls them, has the right to share them, and keeps patient data and sponsor-owned work out.
Which records count
The useful records are the ones your team already keeps to do its job. Based on what Paid4Data lists for wet labs, battery and energy companies and medical device design firms, they include:
- Protocols and SOPs with revision history. The method, and each change made to it over time.
- Electronic lab notebook entries. What was planned, what was done, what was observed and what the scientist decided to try next.
- Instrument run logs and assay results. The settings and raw output of each run, tied to the sample and protocol.
- Sample tracking, inventory and ordering records. Where a sample came from and what happened to it.
- Test-to-failure records and teardown notes. Cycle tests that ran until the part failed, with the notes on what was found afterward.
- Formulations and process conditions. What went into a batch or cell and how it was made.
- Verification and validation protocols and reports. The test plan, the acceptance criteria and the result.
These usually sit in an electronic lab notebook such as Benchling or LabArchives, a LIMS, instrument and battery cycler exports, a quality management system and shared drives.
Why failed experiments matter
Published research tilts toward experiments that worked. Researchers call this publication bias.
One study in the journal Science measured it directly. The authors tracked 221 social science survey experiments that had all passed peer review before they were run, so every study was known whether or not it was later published. Strong results were 40 percentage points more likely to be published than null results, and 60 percentage points more likely to be written up at all. The authors concluded that the gap opens early: researchers do not write up and submit null findings.
That study covers survey experiments, not bench science, so the numbers do not transfer to a wet lab or a test house. The pattern is the relevant part. If results that did not work are rarely written up, then the notebook is often the only place they exist.
A notebook records the buffer that did not improve yield, the anneal temperature that cracked the part and the assay that had to be repeated three times before it gave a clean signal. As the wet labs page puts it, step-by-step records of real experiments, including failed runs, are what models need to learn lab work, and they are rarely published.
Why test-to-failure records are wanted
A datasheet gives a rated value. A test record shows the whole path to failure: the conditions, the readings along the way, the point where performance fell off and what the teardown showed.
Those histories are scarce outside the companies that ran them. They are more useful when the test can be tied back to the design, formulation or process conditions behind the part, because the record then links a cause to an outcome. A cycle log with no reference to what was being tested says much less.
Regulated design work has a similar quality. Medical device design firms document each decision and the evidence for it in design history files, risk analyses and verification reports. That level of detail is unusual in most business data.
What has to be left out
Patient and health records. Paid4Data does not accept them. This matters most for clinical labs, diagnostic work and any study that used human samples linked to a person.
For a sense of what counts as health information, the HIPAA Privacy Rule is a useful reference. The US Department of Health and Human Services describes protected health information as individually identifiable information, held or transmitted by a covered entity or its business associate, that relates to a person's physical or mental health or condition, the provision of health care to them, or payment for that care. HHS gives a laboratory report as one of its examples, alongside a medical record and a hospital bill.
The Privacy Rule also sets a de-identification standard with two methods: a formal determination by a qualified expert, or removal of a list of specified identifiers. HHS notes that both methods still leave some risk of identification. Paid4Data does not take patient and health records at all, so the practical step is to keep those systems and folders out of scope.
This is a description of the rule, not legal advice. Whether your lab is covered by HIPAA or holds health information is a question for your lawyer or privacy officer.
Sponsor- and client-owned work. Contract research, sponsored studies and client testing may belong to the party that paid for them, or be covered by confidentiality terms. That work can only be included where you have the right to share it. This is reviewed with you on the scoping call.
Export-controlled or confidential designs. The same rule applies to client-owned designs and work covered by export controls.
The standard exclusions. Payroll and HR records are always excluded. Payment card numbers, bank account numbers, government ID numbers, data about children and biometric data are never accepted.
What happens to names and dates
Lab records are full of people: who ran the experiment, who signed the page, which supplier sent the reagent, which customer sent the part. Before any buyer sees a dataset, names, emails, phone numbers, street addresses, account and invoice numbers, tax IDs and signatures are removed. People and company names are replaced with consistent pseudonyms, so one scientist's entries still read as one person's work.
Dates are shifted by a consistent offset. Free text is scanned with automated tools and samples are then reviewed by hand. Raw exports are deleted within 30 days of the dataset being finalized. The post on how de-identification works covers the limits of the process as well.
How to describe your records in an intake
The advanced intake asks you to describe each system and its data volume. No files or data are uploaded at that stage. A clear description for a lab covers six things:
- The system. Name the ELN, LIMS, cycler software or shared drive.
- The record types in it. Protocols, notebook entries, run logs, test reports, teardown notes.
- The years covered. Note any gaps or system changes.
- The approximate volume. Storage used, in GB or TB.
- Whose work it is. Separate internal R&D from sponsored, client or grant-funded work, and note what those agreements say.
- What you want excluded. Programs, clients, product lines or date ranges.
It also helps to say whether records can be linked, for example whether a notebook entry references the protocol version and the instrument run it used.
Who can apply
The minimum to qualify is 20 or more employees, three or more years in business and 200 GB or more of data. Smaller groups can still submit the form. It does not have to be a whole operating company either: a single business unit or a closing practice can qualify if it controls the records and has the right to license them.
The full sequence, from scoping call to license, is on the how it works page. To see whether your records are in range, request a free data review.
Sources
Checked on November 16, 2026.
- Social science. Publication bias in the social sciences: unlocking the file drawer, Franco, Malhotra and Simonovits, Science (2014), abstract as indexed by PubMed, U.S. National Library of Medicine
- Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule, U.S. Department of Health and Human Services, Office for Civil Rights