# Bloom: Regulatory intelligence for Anvisa ## From scattered documentation to a submission-ready dossier. Bloom turns study reports, manufacturing records and specifications into eCTD dossiers for Anvisa. AI agents draft each section with its source cited, and a regulatory expert verifies and signs off before delivery. - Every data point reaches the dossier with its source file, page and passage - No dossier is delivered without a regulatory expert's approval ## Every inaccuracy becomes an exigência, and every exigência delays the registration. An exigência is Anvisa's formal request for clarification on a submission. Here are three failures that cost months, and what the platform does about each one before the dossier reaches your desk. ### The problem: A value differs between two sections **With Bloom:** The discrepancy becomes a recorded open item When the same data point appears in two documents with different values, the platform records the discrepancy and points to both sources. The choice stays with whoever has the context to decide, and it appears in the audit kit. ### The problem: A claim the source document does not support **With Bloom:** A citation is mandatory for every claim Every sentence produced carries a reference to the passage that supports it. A sentence without a reference is treated as a defect by the next layer, which takes away the agent's option to fill in what merely sounds plausible. ### The problem: A field filled in by assumption **With Bloom:** The agent declares what is missing When the source does not support a field, the agent records the gap and moves on. The field reaches the expert flagged as open, with a note of what is missing, and is resolved before filing. ## Four layers between the document and the dossier No single step guarantees accuracy. What holds the result up is the sequence: each layer checks the work of the one before, and the last one is human. - **In:** Study reports, Manufacturing records, Specifications and methods, Previous submissions - **Production cycle:** Failed evaluation? The section goes back for redrafting and is evaluated again. Only what passes moves on. - **Out:** eCTD-format dossier, Audit kit. The regulatory expert signs off before delivery. ### Layer 1 · Extraction: Every data point is born with an address Every value read keeps the file, page and passage it came from, and that reference travels with it into the final dossier and the audit kit. - Value-level references: ranges, batches, expiry dates - Discrepancies between documents recorded with both sources ### Layer 2 · Drafting: One specialised agent per section Each section is filled in by an agent given that section's rules: what Anvisa expects, the required format, approved examples and where the information usually sits in the CTD. - Mandatory citation for every claim - Gaps declared, never filled by assumption ### Layer 3 · Evaluation: The judge did not write it Evaluator agents that took no part in drafting check the output against the sources and against the section's acceptance criteria. - Compared with approved dossiers and expert answer keys - A failed section goes back for redrafting ### Layer 4 · Verification: An expert reviews and signs A regulatory expert reviews section by section, resolves what the evaluators flagged, closes the open items and signs. Responsibility for what is delivered stays with a person. - The source document one click from every claim - Every edit recorded with date and author ## Every section of the dossier has its own rules A generic agent gets it wrong exactly where the detail matters. At Bloom, the agent that fills a section is given what Anvisa expects from it, and the agent for the next section is given a different set. Some examples: - **API specification** (M3.2.S.4.1). What the agent is given: Mandatory parameters, limits, pharmacopoeial reference and the table format. - **Analytical method validation** (M3.2.S.4.3). What the agent is given: The parameters required by RDC 166/2017 and the summary for each validation report. - **Batch analyses** (M3.2.P.5.4). What the agent is given: Batches, scale, manufacturing site and result per parameter. - **Manufacturing data** (M3.2.P.3.2 · P.3.3). What the agent is given: Master formula, flow chart, critical steps and equipment. - **Stability** (M3.2.P.8). What the agent is given: Conditions, time points, packaging and proposed shelf life. **Specialised without training on your data.** The rules reach the agent as instructions, examples and criteria at the time of the work. No model is fine-tuned or trained on client documents. ## A change to the platform goes through the same proof your department requires An eval is the systematic, repeatable assessment of output against an expected result. The logic is that of a quality control test: there is a reference standard, an acceptance criterion defined before the test, and the result is recorded. ### Previously approved dossiers Real cases that went through Anvisa, with a known regulatory outcome. The most demanding reference available. ### Expert answer keys Correct answers built by hand for test cases, including the hard ones and the known traps of each section. ### Criteria per section Mandatory fields present, required format respected and the source cited for every claim. ### Change control When a model, an instruction or a rule changes, the whole reference set is reprocessed. If a section that used to be correct starts to diverge, the change is held until the cause is understood. ### What the evaluator checks in every section - Every claim is supported by the cited passage - The section's mandatory fields are filled in - The section does not contradict another part of the same dossier - Format and terminology follow what Anvisa requires - Units, dates and identifiers are consistent - Anything marked not applicable has a justification Evaluation is done by a second agent that took no part in drafting. The separation is deliberate: the same agent tends to accept its own reasoning. ### Who does what **AI agents** (The heavy lifting): - Read and organise your company's documentation - Draft each section with its source cited - Evaluate the output against sources and criteria - Declare gaps and record discrepancies **Regulatory expert** (The decision): - Reviews section by section against the cited sources - Resolves what the evaluators flagged - Decides what goes back to your company as a question - Approves and signs the delivery, with every edit recorded No dossier is delivered without a regulatory expert's approval. ## Alongside the dossier, the record of how it was built So that any claim can be defended months later, by someone who took no part in the work. ### Provenance map Every claim in the dossier linked to the file, page and passage it came from. It answers in seconds where a questioned number came from. ### Points of uncertainty Where the source was ambiguous, where more than one reading was possible and where two documents disagreed, with the sources involved and the reading adopted. ### Items requiring confirmation Data only your company can confirm, listed as explicit open items before filing, with a note of who usually holds the information. ### Coverage and verification report Sections completed, sections marked not applicable with their justification, and what each verification layer flagged along the way. ### What it is for after delivery - **Answering an exigência:** The evidence for every claim is already located. - **Supporting an inspection:** The path from data point to source is on record. - **Continuity for the product:** Whoever takes over later finds the decisions made and the reasons for them. ## Dossier validation and accuracy URL: https://bloompharma.ai/en/validation How Bloom checks every claim before the dossier reaches your desk. An imprecise dossier costs time. A value that differs between two sections, a data point with no support in the source document or a field filled in by assumption becomes an exigência, and every exigência delays the registration. This page describes how the platform handles accuracy: how each data point is extracted together with its source address, how the agents fill each section, how the result is evaluated before it counts as ready, and what the regulatory expert checks before signing. At the end, the audit kit that ships with the delivery. Written for Regulatory Affairs and Quality Assurance teams, and for those who audit them. The PDF version (Portuguese) is the companion to our Data protection and information security document, available on request by email. ### Layer 1 · Extraction: Every data point is born with an address Every value read carries a reference to the exact point it came from, and that reference travels with it into the final dossier and the audit kit. - **References at the value level:** A specification range, a batch number or an expiry date keeps the file, page and passage it came from. Checking any field in the dossier leads back to the exact line in the source document. - **A discrepancy becomes a recorded open item:** When the same data point appears in two documents with different values, the platform records the discrepancy and points to both sources. The choice between them stays with whoever has the context to decide, and it appears in the audit kit as an open item. ### Layer 2 · Drafting: Drafting by specialised agents Each section of the dossier has its own requirements for content, format and source. A generic agent gets it wrong exactly where the detail matters, so each section is filled in by an agent given that section's rules. - **Every section has its own set of rules:** The agent filling in the API specification is given what Anvisa expects from that table, the required format, approved examples and a map of where that information usually sits in the CTD. The agent for the next section is given a different set. - **Specialised without training on your data:** Specialisation comes from instructions, examples and rules given to the agent at the time of the work. No model is fine-tuned or trained on client documents. - **Mandatory citation:** Every claim produced carries a reference to the passage that supports it. A sentence without a reference is treated as a defect by the next layer, which removes the agent's option to fill in what merely sounds plausible. - **The agent declares what is missing:** When the source does not support a field, the agent records the gap and moves on. The field reaches the expert flagged as open, with a note of what is missing. A declared gap is resolved before filing; a gap filled in by assumption shows up later as an exigência. Examples of what changes from one section to the next: | Dossier section | What the agent is given | CTD location | | --- | --- | --- | | API specification | Mandatory parameters, limits, pharmacopoeial reference and the table format. | M3.2.S.4.1 | | Analytical method validation | The parameters required by RDC 166/2017 and the summary for each validation report. | M3.2.S.4.3 | | Batch analyses | Batches, scale, manufacturing site and result per parameter. | M3.2.P.5.4 | | Manufacturing data | Master formula, flow chart, critical steps and equipment. | M3.2.P.3.2 · P.3.3 | | Stability | Conditions, time points, packaging and proposed shelf life. | M3.2.P.8 | ### Layer 3 · Evaluation: Automated evaluation: the evals An eval is the systematic, repeatable assessment of the system's output against an expected result. The logic is that of a quality control test: there is a reference standard, there is an acceptance criterion, and the result is recorded. - **Compared with previously approved dossiers:** Real cases that went through Anvisa, with a known regulatory outcome. The most demanding reference available. - **Compared with expert answer keys:** Correct answers built by hand for test cases, including the hard ones and the known traps of each section. - **Criteria per section:** Mandatory fields present, required format respected, the source cited for every claim. - **The judge took no part in drafting:** Evaluation is done by a second agent, which receives the section produced, the source documents and that section's criteria, and answers objective questions. Separating who drafts from who judges is deliberate, because the same agent tends to accept its own reasoning. - **Reference set and change control:** The reference set is fixed and grows with every new case. When something changes in the platform, whether a model, an instruction or a rule, the whole set is reprocessed and the outputs are compared with the approved ones. If a section that used to be correct starts to diverge, the change is held until the cause is understood. No change goes into production because it looked better on a single case. What the evaluator checks in every section: - Every claim is supported by the cited passage - The section's mandatory fields are filled in - The section does not contradict another part of the same dossier - Format and terminology follow what Anvisa requires - Units, dates and identifiers are consistent - Anything marked not applicable has a justification ### Layer 4 · Verification: Human verification before sign-off The three previous layers exist so that the regulatory expert receives material that is already consistent and can spend their time on what takes judgement. Responsibility for what is delivered stays with a person. - **Reviews section by section:** Checks the content against the cited sources, with the source document one click from every claim. - **Resolves what the evaluators flagged:** Every recorded discrepancy reaches them with the flag and the source, without having to be hunted down. - **Closes the open items:** Decides what can be completed with the material available and what has to go back to your company as a question. - **Approves and signs:** The delivery goes out with a name, a date and a record of what was changed relative to what the agents produced. Each evaluator's opinion and the justification for every section marked not applicable are recorded too. - **Controlled access:** Granted on a need basis, to the smallest number of consultants assigned to the project, with scope limited to that dossier's documents. Individual credentials with a unique, random, strong password plus two-factor authentication, and a log of who accessed what, when, and what was viewed or changed. ### Delivery · Audit kit: The audit kit Alongside the dossier, your team receives the record of how it was built, so that any claim can be defended months later, by someone who took no part in the work. - **Provenance map:** Every claim in the dossier linked to the file, page and passage it came from. It is the document that answers, in seconds, where a questioned number came from. - **Points of uncertainty:** Where the source was ambiguous, where more than one reading was possible and where two documents disagreed. Each point comes with the sources involved and the reading adopted. - **Items requiring confirmation:** Data only your company can confirm, listed as explicit open items before filing, with a note of who usually holds the information. - **Coverage and verification report:** Sections completed, sections marked not applicable with their justification, and what each verification layer flagged along the way. - **What it is for after delivery:** The kit supports the answer to an exigência, because the evidence for every claim is located. It supports an inspection, because the path between data and source is on record. And it supports continuity, because a team that takes over the product later finds the decisions made and the reasons for them. ## Let's talk about your next dossier Write to our team. The conversation can be with Regulatory Affairs, with Quality Assurance or with whoever audits both. - Email: contato@bloompharma.ai - Website: https://bloompharma.ai/en/ Write to ask for: - **A demo:** The platform and the audit kit, on a case similar to yours. - **A conversation with your Quality Assurance team:** How the validation method works and what is recorded at each step. - **A supplier qualification questionnaire:** We answer the questionnaire from your supplier assessment. - **The method documents:** Dossier validation and accuracy, and Data protection and information security. ## Frequently asked questions What Regulatory Affairs and Quality Assurance teams usually ask. Can't find yours? ### Does Bloom replace my Regulatory Affairs team? No. The agents do the heavy lifting of reading, organising and drafting, so people spend their time on what takes judgement. The decision about what goes to the regulator belongs to a qualified person, and the open items only your company can resolve come back to it as explicit questions. ### Who is responsible for the dossier's content? A regulatory expert. They review section by section against the cited sources, resolve what the evaluators flagged, close the open items and sign. The delivery goes out with a name, a date and a record of everything changed relative to what the agents produced. ### What happens when the documentation does not support a field? The agent records the gap instead of filling it in by assumption. The field reaches the expert flagged as open, with a note of what is missing, and anything only your company can confirm appears in the audit kit as an explicit open item before filing. ### What if two documents give different values for the same data point? The platform records the discrepancy and points to both sources. The choice stays with whoever has the context to decide, and the discrepancy appears in the audit kit with the reading adopted. ### Are my company's documents used to train models? No. The agents' specialisation comes from instructions, examples and rules provided at the time of the work. No model is fine-tuned or trained on client documents. ### Who has access to my documents? The smallest possible number of consultants assigned to the project, with scope limited to the contracted service: each one reaches only that dossier's documents. Credentials are individual, with a unique, random, strong password plus two-factor authentication, and every access is logged: who, when, and what was viewed or changed. Our Data protection and information security document covers where the data is stored. ### How do you make sure a platform update does not make results worse? With a fixed reference set that grows with every new case. When a model, an instruction or a rule changes, the whole set is reprocessed and the outputs are compared with the approved ones. If a section that used to be correct starts to diverge, the change is held until the cause is understood. ### What do I receive at the end? The dossier in eCTD format and the audit kit: the provenance map for every claim, the points of uncertainty, the items that need your company's confirmation and the coverage and verification report. ### Which CTD sections does the platform prepare? Each section of the dossier is filled in by an agent with its own rules. Some examples: API specification (M3.2.S.4.1), analytical method validation under RDC 166/2017 (M3.2.S.4.3), batch analyses (M3.2.P.5.4), manufacturing data (M3.2.P.3.2 and P.3.3) and stability (M3.2.P.8). To find out whether your case is covered, write to our team. ### Can my Quality Assurance team qualify Bloom as a supplier? Yes. We answer your company's supplier qualification questionnaire, present the validation method and show the audit kit. Write to our team. ### Is Bloom affiliated with Anvisa? No. Bloom prepares the material your company submits to Anvisa and follows the format and terminology the agency expects. Anvisa is a trademark of the Agência Nacional de Vigilância Sanitária. ## Your next dossier can reach your desk already checked. We walk your Regulatory Affairs or Quality Assurance team through the platform and the audit kit. **Request a demo:** contato@bloompharma.ai ## Blog ### Evals, explained for Quality Assurance URL: https://bloompharma.ai/en/blog/evals-for-quality-teams Published: 2026-10-18 · Author: Bloom team Tags: quality assurance, evals, AI, eCTD dossier How do you prove an AI change did not make the dossier worse? Like a QC test: a reference standard, a criterion set in advance, a recorded result. Anyone in Quality Assurance has a ready question for any system that touches a regulated process: **how do you know it still works after a change?** For an analytical method, the answer is well known. For an artificial intelligence system, the answer tends to be vague: "we tested it", "it seems better", "users liked it". At Bloom, the answer has a name: an eval. ## What an eval is An eval is the systematic, repeatable assessment of a system's output against an expected result. The logic is that of a quality control test: - there is a **reference standard**; - there is an **acceptance criterion**, defined before the test; - the **result is recorded**. If that structure sounds familiar, it is because it is the same one your department requires of a validated process. ## What the output is compared against A section drafted by the platform is compared with three kinds of reference: 1. **Previously approved dossiers.** Real cases that went through Anvisa, with a known regulatory outcome. The most demanding reference available. 2. **Expert answer keys.** Correct answers built by hand for test cases, including the hard ones and the known traps of each section. 3. **Criteria per section.** Mandatory fields present, required format respected, the source cited for every claim. ## The judge did not write it Evaluation is done by a second agent, which receives the section produced, the source documents and that section's criteria, and answers objective questions: - Is every claim supported by the cited passage? - Are the mandatory fields filled in? - Does the section contradict another part of the same dossier? - Do format and terminology follow what Anvisa requires? - Are units, dates and identifiers consistent? - Does anything marked not applicable have a justification? Separating who drafts from who judges is deliberate. The same agent tends to accept its own reasoning, just as nobody reviews their own report well. ## Change control This is the part Quality cares about most. The reference set is fixed and grows with every new case. **When something changes in the platform, whether a model, an instruction or a rule, the whole set is reprocessed** and the outputs are compared with the approved ones. If a section that used to be correct starts to diverge, the change is held until the cause is understood. No change goes into production because it looked better on a single case. That is the difference between "the new model looks good" and "the new model was reprocessed against the entire reference set and no approved section regressed". ## And after the eval The eval does not replace the person. It exists so the regulatory expert receives material that is already consistent and can spend their time on what takes judgement. Each evaluator's opinion is recorded next to the section, and no dossier is delivered without an expert's approval and signature. If your department would like to go through the method in detail, or needs to send a supplier qualification questionnaire, [write to our team](mailto:contato@bloompharma.ai?subject=Quality%20Assurance%20conversation). The full description is on the [Dossier validation and accuracy](/en/validation) page. ### Declared gap or filled-in field? URL: https://bloompharma.ai/en/blog/declared-gap-vs-assumption Published: 2026-10-14 · Author: Bloom team Tags: eCTD dossier, Anvisa, accuracy, AI When the documentation does not support a dossier field, recording the gap resolves it before filing. Filling it by assumption becomes an exigência. Every dossier has fields the documentation does not fully support. A stability report that does not state the packaging used in the study. An API specification that refers to a certificate nobody attached. A batch cited in the analysis that does not appear in the manufacturing record. Faced with a field like that, there are two ways out. The first is to record that something is missing, say what, and move on. The second is to fill the field with whatever seems most likely. For anyone drafting in a hurry, and for any language model, the second is always the more tempting: the dossier looks complete. ## Why the assumption costs more A field filled in by assumption does not go away. It waits. When the reviewer at Anvisa finds a value that does not match the rest of the dossier, or that no attached document supports, the result is an *exigência*, Anvisa's formal request for clarification. And every exigência delays the registration. A declared gap takes the opposite path. It reaches whoever can resolve it **before** filing, with an exact note of what is missing. Sometimes the answer is in a document the company has and did not send. Sometimes the field needs to be revisited. Either way, the decision happens in-house, not in a reply to the regulator. ## How Bloom handles a gap On the platform the rule is simple: **when the source does not support a field, the agent records the gap and moves on.** It does not fill in what merely sounds plausible. Three mechanisms make sure of that: 1. **Mandatory citation.** Every claim produced carries a reference to the passage that supports it: file, page and passage. A sentence without a reference is treated as a defect by the next layer. An agent that filled in by assumption would have nothing to cite. 2. **Independent evaluation.** A second agent that took no part in drafting checks every section against the sources and asks, among other things, whether each claim is supported by the cited passage. Whatever fails goes back for redrafting. 3. **A visible open item.** The field reaches the regulatory expert flagged as open, with a note of what is missing. The expert decides what can be completed from the material available and what has to go back to the company as a question. ## What shows up in the delivery Alongside the dossier, the audit kit includes a list of **items requiring confirmation**: data only the company can confirm, listed as explicit open items before filing, with a note of who usually holds the information. It also includes the **points of uncertainty**, where the source was ambiguous or two documents disagreed, with the reading adopted. That record matters beyond filing. Months later, when someone asks why a field has that value, the answer is written down: where it came from, what was in doubt and who decided. ## In short A declared gap is resolved before filing. A gap filled in by assumption shows up later as an exigência. The difference between the two is not a matter of technology but of rules: the system has to be forbidden from inventing, and there has to be a layer that checks it obeyed. If you would like to see how this looks in a real dossier, [write to our team](mailto:contato@bloompharma.ai?subject=Bloom%20demo). --- Bloom is an artificial intelligence platform for regulatory affairs. It turns scattered documentation into forms and dossiers ready for submission to Anvisa, with AI agents that draft, organise and review the material and a regulatory expert who verifies the result before delivery. Anvisa is a trademark of the Agência Nacional de Vigilância Sanitária. The content of this site is for information only and does not replace the terms set out in contract.