Biology has no compiler

TL;DR: Coding AI succeeds because its output is verifiable against deterministic ground truth. Biology offers no such oracle, so agent workflows produce impressive results whose provenance evaporates. In an industry governed by 21 CFR Part 11, an unauditable result is a liability rather than a shortcut. We don't think the fix is more skills, or a second agent reviewing the first. We think it's an underlying deterministic layer that tracks, traces, and verifies data provenance. That's what Dalea is building: Part 11 auditing across human, AI, and API actions, plus a provenance engine that dissolves the agentic black box.

The landscape

The agentic gold rush reaches the lab

2026 is the year agentic AI stopped being a demo in life science. Agents that autonomously reason, act, and learn across complex research workflows are now a recognized category in the literature rather than a novelty [1], and the capital has followed the enthusiasm: AstraZeneca acquired Modella AI to accelerate oncology research in the same week that NVIDIA and Eli Lilly committed to a billion-dollar joint research lab [2].

The excitement is earned. AI genuinely compresses early-discovery timelines and makes sense of complex biological data. But the systematic evaluations are sobering. When researchers tested eight open-source AI research frameworks on two real tasks reproducing published work, not one completed a full research cycle from literature through computational execution to validated results. Every system was competent at the conceptual parts, planning and summarizing, and every system broke down at robust implementation [3]. The capability is real. So is a gap that the excitement tends to paper over.

Why biopharma runs on the audit trail

To see the gap clearly, start with what makes a record trustworthy in this industry, because it isn't the result itself.

Under 21 CFR Part 11, the FDA's rule for electronic records, a system must use secure, computer-generated, time-stamped audit trails to independently record every action that creates, modifies, or deletes a record, and changes must never obscure what came before [4]. The entire point of the rule is to make electronic records as trustworthy and reliable as paper ones [5]. Who did what, when, and why, for every touch of the data, in a form no one can quietly rewrite later.

This is not bureaucratic overhead bolted onto science. In biopharma, the trail is the evidence. A number with no defensible history behind it isn't a weaker result; regulators treat it as no result at all. Which is exactly why the arrival of a technology that produces results without a history should give the industry pause.

The problem

What worked for code doesn't transfer to biology

The agentic playbook was written on software, and software has a property biology doesn't: a ground truth you can check against. Code compiles or it doesn't. Tests pass or they fail. An agent can write for hours, and at the end a deterministic oracle tells you, unambiguously, whether the output is correct. That feedback loop is what makes coding agents so effective.

Biology has no compiler. The point has been made sharply by researchers working on formal protocol languages: biology still describes its procedures in ambiguous natural language, a medium other engineering disciplines gave up decades ago in favour of compiler-verified specification languages, and that deficit fractures reproducibility along protocol accuracy, pre-execution verification, and portability [6]. There is no unit test for whether a data transformation preserved biological meaning, no build step that fails when an agent quietly misreads a column.

This is why the industry's instinctive responses fall short. Reach for a bigger model and you get more fluent output over the same unverifiable ground. Add more agents and you multiply the untraceable steps. Add more skills and you widen what the system can do without improving your ability to check what it did. Even the popular reflex of having a second agent adversarially review the first only produces another confident, non-deterministic opinion rather than a verifiable record. None of these give you the thing software had for free: a way to know, deterministically, how a result was reached.

The black box at the end of the run

Here is what actually happens in a long agentic workflow. A set of agents runs for hours, burns through millions of tokens, and performs hundreds of transformations: retrieving, joining, reshaping, summarizing. Structured, semi-structured, and unstructured source data gets churned into a fresh pile of semi-structured data. And out the other end comes something that looks authoritative.

What went missing along the way is provenance. Where did each input come from? Which step produced this particular value? Did the model reason correctly at every hop, or invent a number in hour three and carry it forward as fact?

That last question is not hypothetical. In the framework evaluation above, one system produced a complete manuscript containing fabricated FRET signals and biosensor outputs, along with simulation details for analyses that were never run. Another hallucinated confidence scores and implied structural validation without performing any of the underlying computation [3]. The outputs were confident, well-formatted, and unsubstantiated. Nothing about their appearance signalled which parts were real.

This is a structural problem, not a bad-luck one. Commercial AI systems are easy to adopt precisely because they hide their machinery, and in doing so they obscure data provenance, lack transparency, and can generate false information [7]. Generic LLM agents run on implicit state and stochastic execution, which introduces silent state changes and incomplete provenance records by default [8].

In principle you could reconstruct what happened by combing through hours of raw agent logs and chain-of-thought. In practice nobody does, and a transcript of an agent's reasoning is not an audit trail anyway. So you're left with the one thing regulated science cannot accept: an impressive output and no defensible account of where it came from.

Regulators have already noticed. The sixth of the ten guiding principles that the FDA and EMA jointly issued in January 2026 is explicit that data source provenance, processing steps, and analytical decisions must be documented in a detailed, traceable, and verifiable manner, in line with GxP requirements [9]. If you can't trace it, in this industry, it didn't happen.

The solution

The fix isn't a bigger agent

So the real problem isn't that agents aren't capable enough. It's that biology never had the deterministic verification layer software takes for granted, and no amount of model scale conjures one into existence.

We don't think the answer is more skills. We don't think it's a second agent reviewing the first. We think the answer is an underlying deterministic set of tools that tracks, traces, and verifies data provenance, and that expresses scientific work in a language built for science rather than as free-form text an agent has to be trusted on. Scientific experiments are supposed to be reproducible, auditable, and deterministic [8]. The substrate the agents work on should enforce exactly that, no matter who or what is doing the work.

Dalea: a language for science, audited by default

This is the mission Dalea is on. In our first paper we argued that the natural unit of scientific work is the package, not the page: protocols, data models, queries, and code as connected, first-class objects rather than flattened documents. That same idea is what makes verification possible. When work is expressed in structured, deterministic primitives, every operation on it is inherently traceable.

On that foundation, two things follow.

First, everything is audited by default. Every action lands in the same tamper-evident, time-stamped record, regardless of who or what performed it. A scientist editing a result, an AI agent transforming a table, and a script writing through the API all produce the same kind of entry: identity, action, timestamp, and the before-and-after of what changed. Human, agent, and programmatic access are not three separate compliance stories. They are one audit substrate, so working in Dalea means working in an environment that already satisfies what Part 11 was written to guarantee. An API that bypassed the log, or agents whose actions went unrecorded, wouldn't be a convenience. It would be a gap with good intentions, and Dalea closes it by construction.

Second, a provenance engine dissolves the black box. Beyond logging every action, provenance is about how those actions connect. The mature grammar for this already exists. The W3C PROV standard models the origin of any datum as a graph of entities (data), activities (transformations), and agents (the humans or software responsible) [10], the same vocabulary now being extended to make AI-agent decisions first-class, traceable nodes rather than opaque prompt context [11]. A provenance engine built on that lets you take any value in a workspace and walk the graph both ways: back through every transformation to its source, and forward to everything derived from it, across human hands and agent runs, without the lineage breaking at the boundary between them. When an agent runs for six hours, you don't get a polished artifact and a shrug. You get the chain, without combing through a single line of reasoning log.

Provenance is the verification biology never had

Software could trust its agents because a compiler and a test suite stood behind them. Biology has no such oracle. What it has always demanded, and rarely gets to keep once AI enters the loop, is a complete and honest account of how a result was reached.

That account is the verification layer life science has been missing. It's also the difference between an AI-native platform you can build a submission on and one you can only be impressed by. Biology has no compiler, so we're building the next best thing: a deterministic substrate where every step, human or machine, leaves a trail you can trust.

References

1. Seal S, Huynh DL, et al. AI agents in drug discovery: applications and case studies. Drug Discovery Today, 2026;31(3):104650. doi.org/10.1016/j.drudis.2026.104650

2. Reuters. US, European regulators set principles for good AI practice in drug development. January 14, 2026. On the concurrent wave of AI investment in pharma, including the AstraZeneca–Modella acquisition and the NVIDIA–Eli Lilly research lab.

3. Can AI Conduct Autonomous Scientific Research? Case Studies on Two Real-World Tasks. bioRxiv, January 2026. Evaluation of eight open-source AI research frameworks; none completed a full research cycle, and several fabricated results. biorxiv.org

4. Electronic Code of Federal Regulations. 21 CFR Part 11, Electronic Records; Electronic Signatures, section 11.10(e). ecfr.gov

5. U.S. Food and Drug Administration. Guidance for Industry: Part 11, Electronic Records; Electronic Signatures, Scope and Application. fda.gov

6. Towards Autonomous Biology: Compiler-Verified Protocols as a Foundation for Real-World AI Execution. bioRxiv, May 2026. On biology's reliance on ambiguous natural-language protocols and the resulting deficits in verification and reproducibility. biorxiv.org

7. Pickard J, et al. Automatic biomarker discovery and enrichment with BRAD. PMC, 2025. On commercial AI systems obscuring data provenance and generating false information. ncbi.nlm.nih.gov/pmc/articles/PMC12064167

8. R-LAM: Reproducibility-Constrained Large Action Models for Scientific Workflow Automation. arXiv:2601.09749, 2026. On implicit state, stochastic execution, and incomplete provenance in generic LLM agents. arxiv.org/abs/2601.09749

9. FDA and EMA. Guiding Principles of Good AI Practice in Drug Development, January 2026. Principle 6, Data governance and documentation. fda.gov | ema.europa.eu

10. Moreau L, Missier P, et al. PROV-DM: The PROV Data Model. W3C Recommendation. Entities, activities, and agents as the core model for data provenance. w3.org/TR/prov-dm

11. PROV-AGENT: Unified Provenance for Tracking AI Agent Interactions in Agentic Workflows. arXiv:2508.02866, 2025. Extending W3C PROV to make AI-agent decisions first-class, traceable nodes. arxiv.org/abs/2508.02866