Breaking the Determinism Wall: How Indago Drafts Consistent, Traceable Regulatory Submissions

Regulatory teams are putting AI drafting tools to work, because the writing load is heavy and the teams are small. A stack of study reports goes in and a draft of the whole submission comes out. The trouble shows up in review. Every fact in a filing has to trace back to its source, and it has to come out the same way every time someone checks.

The cause sits inside the models themselves. A system is deterministic when the same input always gives the same output. A calculator is deterministic. Type 12 times 4 and you get 48 every time, on anyone's desk. General-purpose AI models are built another way. Each answer is a fresh prediction of the words most likely to come next, so the same question asked twice can come back with two different answers, and sometimes two different numbers. A model that works this way can't promise that the NOAEL it writes matches the study report, or that it will write the same one tomorrow.

We call this the determinism wall. It is the point where AI output has to become evidence and can't, because nobody can count on it saying the same thing the next time. Most general AI tools stop at the wall. A summary that states one NOAEL on Monday and another on Tuesday isn't evidence of anything. Every regulatory and quality lead we meet raises the wall, usually within the first ten minutes, and they are right to.

"The question I hear most is some version of this: if I run it again tomorrow, do I get the same NOAEL? If the answer is no, the AI draft is one more thing to check. We built Indago so every value that lands in a filing comes out the same way twice."

Life sciences runs on a record that has to hold up for years

A program reaches an IND after 18 to 36 months of assessment work. Pharmacology, pharmacokinetics, toxicology and CMC data arrive study by study, each from a different CRO in a different format. Someone has to keep working out what it all adds up to. Then comes a pre-IND meeting, then the IND itself, then amendments, Investigator's Brochure updates and annual reports for as long as the program runs.

Every one of those steps restates the same facts. The NOAEL from a 28-day tox study shows up in the nonclinical summary, the integrated summary, the dose rationale and the IB. The IND has to give FDA enough pharmacology, toxicology and manufacturing information to judge whether it is reasonably safe to start human trials, and FDA has 30 days to decide.^[1] When a number disagrees with itself across those places, reviewers notice. About 1 in 11 initial INDs is placed on clinical hold in its first 30 days, and CMC is the most common reason.^[2]

That is why regulatory work is built around evidence. A reviewer at FDA, EMA or any other agency needs to see where each claim came from and get the same answer when they look again. Rules for electronic records say the same thing in their own terms: FDA's Part 11 expects secure, time-stamped audit trails that show who changed a record and when.^[3]

So the determinism wall hits harder here than almost anywhere else. The same NOAEL has to read the same in five documents over three years, and it has to match the study report every time someone checks. A tool that can't get past the wall has no place in a filing.

Each kind of content in a submission needs a different guarantee

The way past the wall starts with being specific about what has to be reproducible. A Module 2 summary holds four kinds of content, and they carry different obligations.

This split matters because it tells you where to demand determinism and where to demand something else. Values, references and structure can and should be deterministic. Interpretation is where judgment lives, so it needs traceability and a human signature. A tool that treats all four the same way either blocks useful drafting or lets unsupported claims through.

For the team that signs, review turns into checking

Picture the regulatory lead at a 20-person biotech three weeks from filing. A revised stability report lands from the CRO on a Tuesday.

Today, she reopens every section that might cite it, hunts for the old values by hand and hopes the consultant catches the rest. Each reviewer re-derives the numbers they are responsible for, because nobody can see where the last draft got them.

Now run the same Tuesday in Indago. Every value in her drafts was pulled from a specific page of a specific report, and Indago kept that link. So when the revised stability report goes into the program, Indago can see which statements rested on the old one. It flags each of them, in every section where they appear. Nothing else moves.

That link is what traceability means in practice. Her reviewers click the flagged value, see the new page it points to, and accept or fix it, without re-deriving the shelf-life claim from scratch. Her CMC consultant does the same in the same workspace, so no Word file comes back with tracked changes to reconcile. When she signs, every fact in the section points to a page in the current source documents, and every structural check it passed will give the same result if FDA or anyone else runs it again.

The same holds after filing. When a health authority question arrives, the answer starts from the same source record the IND was built on.

Determinism belongs wherever the output becomes evidence

We have not built a deterministic language model, and we aren't claiming to. Ask Indago to draft the same section twice and the wording will differ, the way it would from two writers given the same brief.

What stays fixed is what the draft rests on. Everything Indago drafts is grounded in your source documents, and every value and reference in it can be followed back to the page it came from. This is the split from the table above: values and references are held fixed, and wording can vary. That is determinism where it matters in a filing, in the facts a reviewer checks and signs for.

Indago is built in four layers. Models do the reading and the writing. Everything they produce is stored, linked to its source and checked, and that stored record is what stays fixed.

One check, run twice

The validation layer is deliberately plain. Each rule traces to published guidance, it checks only what can be verified from the text, and its output may not hedge. Words like "appears," "may" and "likely" are banned. A result reads like this:

Run 1 · Monday 09:14

Requirement

NOAEL stated for all repeat-dose studies.

Result

Not Met.

Detail

NOAEL not identified for Study 1042.

=

Run 2 · Tuesday 16:02

Requirement

NOAEL stated for all repeat-dose studies.

Result

Not Met.

Detail

NOAEL not identified for Study 1042.

Run it again on the same text and you get the same line. It runs on every save and can't be skipped. When content changes, the old result is discarded and the check runs fresh, so a compliance status is never stale.

Your documents become assertions

A study report is written for a person who reads it from start to finish. A sentence on page 40 leans on a table on page 12 and a definition on page 3. That suits a reader. It suits software badly when the job is to find one fact in thousands of pages.

So the first thing Indago does with your corpus is take it apart. Human writing packs several facts into one sentence. Take a line from a study report:

A person reads that as a single thought. Indago breaks it into the separate statements it contains:

• Study 1042 dosed male and female rats.
• The dose in Study 1042 was 10 mg/kg.
• The route of administration in Study 1042 was intravenous injection.
• Dosing in Study 1042 was once weekly.
• Dosing in Study 1042 continued for four weeks.

We call these assertions. Nobody would write a report this way. The statements are flat and repetitive because they are written for a machine to match and retrieve, and each one can be found, checked and cited without the others. Every assertion keeps a link to the page it came from.

Source documentStudy 1042 report · p. 40 · §7.2

Study 1042 was a 28-day repeat-dose toxicity study in Sprague-Dawley rats. Male and female animals received 10 mg/kg of the test article by intravenous injection once weekly for four weeks. No treatment-related mortality was observed, and the no-observed-adverse-effect level (NOAEL) was 10 mg/kg.

Split into assertions

Assertion store0 stored

1. a3f9c1eStudy 1042 was a 28-day repeat-dose toxicity study.p. 40
2. 7b2d04fThe species in Study 1042 was Sprague-Dawley rat.p. 40
3. e81c6aaStudy 1042 dosed male and female animals.p. 40
4. c45f2b9The dose in Study 1042 was 10 mg/kg.p. 40
5. 19d7e3cThe route of administration in Study 1042 was intravenous injection.p. 40
6. f0a8b62Dosing in Study 1042 was once weekly for four weeks.p. 40
7. 5c3e91dNo treatment-related mortality was observed in Study 1042.p. 40
8. b7d2a08The NOAEL in Study 1042 was 10 mg/kg.p. 40

Drafted from the assertions

Module 2.6.6 · Toxicology written summary2.6.6.3 Repeat-dose toxicity · draft

 

The assertions are then indexed two ways. A keyword index finds exact terms such as a study number, a lot number or a compound name. A semantic index finds statements by what they mean, so a search for liver findings also returns an assertion about hepatocellular hypertrophy. Both are fast and cheap to run, which matters because they run constantly.

Assertions work better than raw pages for everything that comes later. A statement that says one thing is easier to match, easier to rank and harder to misread than a sentence that says five. Before our drafting agent writes, it is handed the assertions that bear on the section, a step the industry calls retrieval-augmented generation, or RAG. The agent writes from those assertions and carries their links forward, so each sentence in a draft arrives with its source attached. When you search your program, you are using the same index the agent uses.

Live grounding: sources for the sentences you write yourself

Citing a source is easy when our agent wrote the sentence, because the agent started from assertions. The harder case is a sentence that comes out of your own head. A reviewer rewrites a paragraph or adds a conclusion from memory, and that sentence has no link to anything.

Live grounding covers that case. When you finish typing a sentence in Indago, it is checked against the assertions in your program and marked where it sits. Green means a source supports it. Red means a source contradicts it. No color means Indago found nothing either way. Click a marked sentence to see the assertion it was matched to, and click once more to open the source PDF at that page with the passage highlighted.

Module 2.6.6 · Toxicology written summaryEditing · live grounding on

Source 

Nothing checked yet.

 

This takes milliseconds, and no large language model is involved. Live grounding runs on a set of small, specialized models inside Indago's own environment, each with one narrow job. One finds the assertions most likely to be relevant, another judges whether your sentence and an assertion mean the same thing, and a third looks only for contradiction. Numbers are compared as numbers, so 13 weeks against a source that says six weeks is a conflict however alike the two sentences read. The models are ours and run locally, so the sentence you typed is never sent to an outside provider.

"We hold live grounding to one rule above the others: a wrong color is worse than no color. When Indago isn't sure, it leaves the sentence unmarked. Vouching for something your documents don't say would do more harm than staying quiet."

Judgment stays with a named person

Some decisions should never collapse into a single computed answer. Regulatory strategy, how to frame an unexpected finding, which questions to bring to a pre-IND meeting and how to answer FDA when it pushes back all belong to people with their names on the submission.

Indago leaves those calls where they are. Every finding is accepted by a person before a package is final, and the interpretation in each section carries an author. What changes is the ground that person stands on. The facts under their judgment are fixed, sourced and checkable, so review time goes into the judgment itself.

Regulators are drawing the lines now

FDA put a generative AI tool in front of staff across the agency in June 2025, for work such as protocol reviews and adverse event summaries.^[4] Its January 2025 draft guidance sets a credibility framework for AI that produces evidence behind regulatory decisions, and names drafting a regulatory submission as outside that framework's scope, as long as it doesn't affect patient safety, drug quality or study reliability.^[5] EMA's reflection paper on AI across the medicinal product lifecycle, adopted in September 2024, takes a similar risk-based view and puts responsibility on the applicant for any AI it uses.^[6]

Regulators are using AI themselves, and no agency certifies a vendor's tool. The sponsor answers for every word it files, so a drafting tool earns its place by making that responsibility easier to carry.

The place to start is one section

Pick the section you trust least and put any AI tool, ours included, through four questions:

• Run the same check on the same content twice. Do you get the same result?
• Can every value and statement resolve to a specific page of your source documents?
• Can you reconstruct how the section was produced, who changed it and when?
• Is your data retained by the model provider, or used to train anything?

Indago was built so the answers are yes, yes, yes and no.

About Indago

Indago, Inc. builds regulatory drafting software for biotech teams preparing Pre-IND, IND and NDA/BLA submissions. Every statement it drafts links back to its source document, so regulatory professionals can check a claim in one click. See it on your own data at indago.bio/resources, or write to sales@indago.bio.

Sources

1. 21 CFR 312.23 (IND content and format) and 312.40 (30-day review). ecfr.gov
2. Lapteva L, Pariser AR. Investigational New Drug applications: a 1-year pilot study on rates and reasons for clinical hold. Journal of Investigative Medicine, 2016 (125 of 1,410 initial INDs, 8.9%). doi:10.1136/jim-2015-000010
3. 21 CFR 11.10(e), audit trails for electronic records. ecfr.gov
4. FDA launches gen AI Elsa to support clinical, regulatory tasks. TechTarget, June 2025.
5. FDA. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products; Draft Guidance. Federal Register, January 7, 2025.
6. EMA. Reflection paper on the use of Artificial Intelligence in the medicinal product lifecycle, adopted September 2024.