AI Transparency Statement and Instructions for Use

Version 2026-08.2 — Effective August 6, 2026

1. What this document is

Ploid AI Biotech, S.L., a Spanish limited liability company with Tax ID (CIF) B19786623, registered office at Avenida Diagonal, 612 - P. 1 PTA. 3, 08021 Barcelona, Spain ("Ploid", "we", "us"), is the provider of the AI system described here.

This document serves two purposes:

1. It is our transparency statement under Article 50 of Regulation (EU) 2024/1689 (the AI Act): it tells you that you are interacting with an AI system, on whose behalf that system acts, and how content it generates is marked.
2. It is our instructions for use: what the system is for, what it can and cannot do on its own, where it is known to fail, and what a competent user has to do to use it safely.

It applies to every Ploid deployment, including deployments Ploid operates exclusively for a single organization. Unlike our other legal documents, this one is not conditional on a Ploid contract governing your use: the obligations it discharges are ours as provider of the system, and they do not change because your organization signed a separate agreement. Where Ploid operates a dedicated deployment, the AI system is provided by Ploid and operated on behalf of that organization, and both are disclosed to you in the application.

This document is referenced from Sections 4A and 9A of our End User License Agreement and forms part of it.

2. Intended purpose

The Services are a research platform for bioinformatics and life-sciences analysis. Their intended purpose is to help a qualified researcher explore data, write and run analysis code, execute computational workflows, and draft research documents — with that researcher retaining responsibility for the scientific conclusions.

The Services are intended for research and informational use only. In particular:

• They are not a medical device and are not intended to be one, in whole or as a component. They are not intended for use in diagnosis, prevention, monitoring, prediction, prognosis, treatment or alleviation of disease.
• They do not provide medical, clinical, or diagnostic advice, and no output is a clinical finding.
• No output may be used as the sole or primary basis for a decision about the care, treatment, or medication of an identifiable person.
• They are not intended for, and must not be deployed in, the high-risk domains listed in Annex III of the AI Act — including employment and worker management, education, access to essential private or public services, creditworthiness, insurance pricing, law enforcement, migration and border control, and the administration of justice.

Section 4 of the EULA makes these prohibitions binding obligations, not merely stated intentions. If you place the Services on the market under your own name, modify them substantially, or change their intended purpose so that the system falls into a high-risk category, you become a provider of that system and assume the corresponding obligations; Section 9A of the EULA sets this out.

3. What the agent does, and what it does without asking

The Services are agentic: on a single request the AI agent plans a course of action and carries it out over many steps, rather than answering once. It works inside a per-study sandbox — an isolated computing environment holding that study's files.

Acts the agent performs on its own, with no confirmation step:

• reading, writing, editing and deleting files inside the study sandbox;
• running code — shell commands, Python and R — inside the sandbox;
• installing software packages into the sandbox at runtime;
• starting computational workflow runs, which consume compute and your quota;
• reading your datasets and the files inside them;
• searching the web and querying public scientific data sources;
• delegating sub-tasks to its own worker agents (Section 4).

Read that list carefully. The agent is not asked to approve each command before it runs it. Oversight of execution is retrospective — you see every command and its result in the conversation, and you can stop a run in progress — rather than a prior approval gate. This is a deliberate design choice for research work, and it is the single most important thing to understand about supervising this system.

Acts that require your explicit confirmation before they take effect:

• creating a dataset, a script, or a skill from sandbox output. The agent stages these and they are shown to you as pending; nothing is persisted to your workspace until you confirm in the interface;
• exporting a report as a PDF or a Word document. Every export is preceded by a review step you have to pass, described in Section 6.

Acts the agent cannot perform at all:

• reaching outside its study sandbox into another study, another workspace, or Ploid's infrastructure;
• changing your workspace settings, your plan, your billing, or the membership and permissions of your organization;
• sending email or messages on your behalf, or publishing anything outside the Services;
• creating or revoking a share link that would expose a study to people outside your workspace;
• deleting your account or your workspace.

Where it stops and asks you. The agent can pause and put a question to you when a request is ambiguous or a decision is yours to make. While it is waiting, the conversation is blocked until you answer. You can also stop any run in progress at any point.

The record of what it did. Every step — each command run, each result returned, each question asked and answered — is kept and shown to you in the conversation. Section 6 describes how to read it as a single record and download it.

4. The agents you may interact with

Requests are handled by an orchestrating agent, which may delegate parts of the work to specialized worker agents. Each is an AI system component, and each is identified as such where its work is shown to you:

You do not choose between them; the orchestrator decides when to delegate. Their work appears in the conversation attributed to the agent that performed it. You do choose the model tier the agent runs on, which trades speed against depth of reasoning.

5. Known limitations, and how outputs fail

Outputs are generated by large language models. They are probabilistic: the same request can produce different results on different runs, so an analysis is not reproducible from the request alone — it is reproducible from the code and the workflow the agent produced, which is why both are kept in the study.

The failure modes below are the ones we consider realistic and material. They are not hypothetical, and they are not exhaustive.

• Fabricated or misattributed citations. A reference may look correct in every respect — plausible authors, plausible journal, plausible year — and not exist, or exist but not support the claim it is attached to. Check every citation against the source before relying on it or publishing it.
• Over-confident interpretation. Variant calls, effect sizes, enrichment results and similar findings are frequently stated in more certain language than the underlying evidence supports, and confidence intervals, multiple- testing correction and power may be omitted unless asked for.
• Statistical tests applied without checking their assumptions. A test may be chosen because it fits the shape of the data as presented rather than the design of the experiment, and normality, independence, variance and batch effects may go unexamined.
• Code that runs successfully on the wrong input. A script that completes without error and produces a well-formed plot is not evidence that the right columns, units, reference genome, or sample groupings were used. Silent mis-specification is the most common way an analysis here goes wrong.
• Silent context loss on long analyses. In a very long conversation, earlier detail may fall out of the model's working context. A constraint you set early may stop being applied later without any notice that it has been dropped.
• Confident answers about your data that the agent did not verify. The agent may describe a dataset's contents from its name, its metadata, or an earlier summary rather than by reading it.
• Hidden dependence on the model tier. Results, and the quality of reasoning, differ between tiers. A finding reproduced on one tier may not reproduce on another.

6. Human oversight

The Services are built on the assumption that a qualified researcher reviews the work rather than accepting the conclusion. These are the points at which that is enforced rather than merely expected:

• Nothing is written to your workspace without you. Datasets, scripts and skills the agent proposes are staged and shown as pending until you confirm.
• Every export is gated. Exporting a report as PDF or Word opens a review step naming the failure modes in Section 5. You confirm you have reviewed the content; nothing is produced until you do, and the confirmation is not remembered between exports — each export is its own decision.
• The sign-off travels with the artifact. Your name and the time of the confirmation are recorded and written into the exported file, alongside the provenance described in Section 8. Where an export is made through a share link by someone with no account, there is no identified person to name, and the file says so rather than implying a review took place.
• You can stop a run. Any run in progress can be cancelled; partial output is kept.
• The run log is a record you can take away. Each conversation has a run log listing every agent run — its flow, model and outcome — and every step beneath it, downloadable as Markdown or JSON. Where the record is incomplete because it exceeded a size limit, both the screen and the downloaded file say so. It shows how a result was produced; it is not a statement that the result is correct.

7. Accuracy, validation, and what we do not claim

• No accuracy figure is claimed for interpretive output. We do not publish, and you should not assume, an error rate for the agent's scientific conclusions, variant interpretations, or literature summaries.
• No clinical validation has been performed. The Services have not been validated for any clinical or diagnostic purpose, have not been assessed under Regulation (EU) 2017/745 or Regulation (EU) 2017/746, and carry no CE marking.
• Benchmark results measure something narrower than fitness for your work. Where we publish results on public bioinformatics task benchmarks, those measure the agent's ability to complete benchmark tasks whose expected answers were fixed in advance. They do not establish accuracy on your data, in your organism, for your assay, or for interpretive and clinical questions, and they are not a substitute for your own validation.
• You are expected to validate. The intended user is a qualified researcher who reviews the code, the intermediate results and the reasoning, not only the conclusion.

8. Data types, and human data in particular

The Services are built for research data, and their design assumptions — including that no output is a clinical finding — follow from that.

Each workspace declares its data type. A workspace is created as holding plant, animal or microbial data, and an owner or administrator can declare that it holds human data instead. Declaring human data requires an explicit confirmation that your organization has a lawful basis under Article 9(2) GDPR, and that confirmation is recorded against the person who made it. While the declaration is in force, the workspace carries a research-use-only notice that cannot be dismissed, and that notice is written into every report exported from it, on every page. Workspaces that hold no human data carry none of this.

Where you process personal data, you are the controller and Ploid is the processor; our Data Processing Agreement governs that processing and our Privacy Policy explains the boundary between the data we hold as controller and the data we hold as processor.

Human genetic, clinical and other special-category data. Data of this kind is special-category personal data under Article 9 GDPR. You must not place it in the Services unless:

• you have a valid legal basis under Article 9(2) GDPR;
• you have applied appropriate safeguards — pseudonymization or de-identification — before upload; and
• your use remains within the intended purpose in Section 2, which excludes every clinical use.

Ploid does not inspect the content you upload and cannot detect a failure to meet these conditions. Section 4 of the EULA, Section 2.3 of the DPA and Section 3.3 of the Privacy Policy state the same requirement as a binding obligation.

Research on plants, animals, microbial and environmental data, where no personal data is involved, does not engage these conditions.

9. How generated content is marked

Reports, figures and documents the Services produce are synthetic content: they are generated by an AI system, not authored by a person.

• In the application, AI-generated content is identified where it appears. The conversation shows each step the agent took, each command it ran and each result it received, so that what you are reading can be traced to how it was produced.
• In files you export — PDF and Word documents — a provenance block records that the content was generated by an AI system, which entity provided and operated that system, the model tier used, the study the content came from, and who confirmed at the review step that they had reviewed it (or, where no identified person did, that nobody is recorded). The same information is written into the file's document metadata, so it can be read by software rather than only by a person.
• On exports from a human-data workspace, the research-use-only notice appears at the foot of every page of an exported PDF as a page rule rather than as body text, so it is not removed by editing the content.

Two deliberate limits on what is marked:

• Intermediate steps and machine-facing calls are not marked. The agent's reasoning, its internal tool calls and the API traffic between Ploid's own services carry no marking. They are not published content, and marking them would serve no reader.
• Marking asserts provenance; it does not certify correctness. A provenance record means "an AI system produced this". It does not mean the content was reviewed, and it does not mean it is right.

If you republish or circulate content the Services generated, the obligation to disclose that it is AI-generated travels with it. Removing or obscuring the provenance record does not remove that obligation, and Section 4 of the EULA prohibits presenting outputs as human-authored or as clinically validated.

10. Third-party models and where inference runs

Producing an output requires sending the relevant part of your content to a model inference provider. Those providers are sub-processors, and they are named with their role and location at Subprocessors.

A workspace is created with EU data residency, under which model calls are routed to inference endpoints inside the European Union. An administrator of your workspace may change it to global residency, which permits routing outside the EU; that is a controller decision, and Sections 5.3 and 7 of the DPA govern the resulting transfer.

Ploid selects inference providers whose terms prohibit them from using data submitted through their APIs to train their own models, and configures the Services accordingly. Whether Ploid may use your content to improve its own models depends on your plan, and is governed by Section 6 of the EULA — the single authoritative statement of those terms.

11. Changes, and how to reach us

We will update this document as the Services change. Each version carries a version identifier and an effective date, and material changes are notified in the application.

Ploid AI Biotech, S.L. Avenida Diagonal, 612 - P. 1 PTA. 3, 08021 Barcelona, Spain CIF B19786623 (EU OSS VAT ID ESB19786623) Email: legal@ploid.ai

12. Related documents

• End User License Agreement — the terms governing your use, including the intended purpose and prohibited uses (Sections 4, 4A, 9A) and the per-plan data-use terms (Section 6)
• Privacy Policy — how Ploid processes personal data as controller, and how to exercise your rights
• Data Processing Agreement — Ploid's Article 28 processor terms for personal data inside your content
• Subprocessors — the providers Ploid engages, their role and their location
• Open-Source Software Notices — the third-party components incorporated in or operated as part of the Services