PEN-STACK: A provenance-enforcing tool layer for language-model agents in genome writing

arXiv:2608.20412 · q-bio.GN · Submitted 2026-08-15 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "PEN-STACK: A provenance-enforcing tool layer for language-model agents in genome writing".

Marcus: Here's the gist of the paper. PEN-STACK is an open tool layer designed to supply guaranteed provenance to language-model agents engaged in genome writing, functioning as a substrate rather than just an application.

Ines: It enforces a no-fabrication invariant: every quantitative value reported must originate from a validated tool call, making ungrounded numbers structurally unrepresentable.

Marcus: This is achieved through mechanisms such as preventing candidate claims via errors on.as claim, freezing the nested provenance record to ensure immutability, and explicitly returning out-of-scope quantities as known unknowns.

Ines: The system exposes ten genome-writing design stages through twenty-two scope-aware tools, accessible via a software development kit, a Model Context Protocol server, and a Representational State Transfer interface.

Title and authors: Ines: So, we’re looking at this paper today called "PEN-STACK: A provenance-enforcing tool layer for language-model agents in genome writing." It sounds like they're trying to fix a big problem with how these AI models are used in biology.

Marcus: Yeah, it’s about making sure when an AI agent is designing a new genome, every single number it spits out actually comes from a tool you've already validated. It’s basically building this substrate so the model can't just make things up.

Yuki: From my side, I’m thinking about what that means for us. If these agents are making design plans for new DNA sequences, we need to know those quantities—like where to write something or what enzyme to use—are actually feasible and grounded in reality, not just plausible sounding text.

Ines: Exactly. The core idea is this layer enforces a "no-fabrication invariant." This means if the agent reports a quantity, it has to show the exact tool call that produced it. If you can't show that tool call, the claim is structurally unrepresentable, meaning it can't be true >

Marcus: It’s like making sure every claim in a report has its source line right next to it. They achieve this by stopping candidate claims with errors on something called ".as claim", and they freeze the nested record so you can’t change the history later, which keeps everything immutable >

Yuki: I wonder how that changes things for designing new life forms. Right now, we rely a lot on the model to bridge those gaps between design steps. If it can't invent a required component, it has to admit it doesn't know >

The paper's summary: Ines: The paper summarizes that they’ve built ten genome-writing design stages and twenty-two scope aware tools. These tools are all accessible through a software development kit, a Model Context Protocol server, and a representational state transfer interface >

Marcus: And the main result they show is how much fabrication happens without this layer. They found that in tests involving four planning goals needing six tool only quantities each—that’s twenty-four required quantities total—models fabricated between ninety point eight percent and ninety eight point eight percent of those numbers when just using a naive prompt >

Yuki: That high fabrication rate is concerning because it shows how much the models are guessing when they don't have a reliable source for the necessary design parameters >

Ines: Right, but then they show that if you drive the tools—meaning you force the agent to use those validated tools—the same models fabricated nothing on a four goal audit >

Marcus: That’s where the real difference is. Driving the validated tools results in every model family passing an audit when quantities have to go through those tools, and they even used a grounded zero as a comparison point >

Yuki: So, it suggests that if we give the agent access only to verified design steps, it stops inventing data points that are supposed to be fixed by the tools >

Ines: It really points toward grounding being better than just prompting or scaling up the model itself when you need verifiable biological quantities >

The paper's improvements: Marcus: Now let’s talk about what they suggest as improvements. One big one is grounding instead of prompting, which removes fabrication. The idea is that by requiring every quantity to "originate from a validated tool," agents can produce verifiable quantities, and the grounded arm, a four goal audit, warrants replication at the two hundred forty field scale >

Ines: That’s solid. They also implemented something called a type-enforced invariant which prevents claims. So if you try to generate a result that isn't from a tool call, it throws an error on ".as claim", and the generated sequence or structure remains just a proposal until writer verification scores it against measured data >

Yuki: That sounds like they’re trying to stop the models from asserting ungrounded facts as truth about biological systems >

Marcus: And they handle the things that are out of scope, too. They explicitly return quantities falling outside a tool’s scope card as known unknowns, and these are neither guessed nor omitted >

Ines: That’s important because it means they aren't making up data for parameters that the current tools can't handle; they just state what they don't know about >

Yuki: I like that explicit handling. It prevents them from inventing necessary data points, which is a big step toward more honest agent behavior >

Conclusion: Ines: So to wrap up, the PEN-STACK paper argues that grounding, not just prompting or making models bigger, removes fabrication by requiring a substrate for these agents >

Marcus: It provides that substrate as open and importable. They showed how driving validated tools stops fabrication on a four goal audit across different model families >

Yuki: For the wider field, this means if we build agent systems in biology, we can use this type-level solution as code to ensure the agents aren't just hallucinating necessary components >

Ines: And they showed that even with some noise, like a correlation of zero point five seven one for the human K562 head at exact-site resolution, the system still functions as intended >

Marcus: But they did have some limitations. They noted that while exact-site resolution is good, when you use coarser resolution features, like those served by default, the validation flag isn't returned reliably and you get a machine readable downgrade flag instead >

Yuki: That’s a fair limitation. It means if we rely on the defaults for speed or simpler analysis, we might not have that guaranteed provenance >

Ines: Yeah, so while it solves the fabrication problem by enforcing the tool layer structure, there's still a trade-off between resolution and certainty >

Marcus: It’s a lot of work to build this substrate, but if you can import it into other agentic biology systems, it could be very powerful for ensuring reliability >

Anees Ahmed Mahaboob Ali, Radhakrishnan Delhibabu, Everette Jacob Remington Nelson

School of Bio Sciences and Technology, Vellore Institute of Technology · School of Computer Science and Engineering, Vellore Institute of Technology

q-bio.GN

Submitted: 2026-08-15

Updated: 2026-11-08

Code: https://github.com/ahmedanees-m/pen-stack

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: PEN-STACK is an open tool layer designed to supply guaranteed provenance to language-model agents engaged in genome writing, functioning as a substrate rather than just an application.

Key concepts

No-Fabrication Invariant
This core rule mandates that every quantitative value reported by an agent must originate directly from a validated tool call. If a model tries to state a number without this direct source, the system flags it as ungrounded and structurally unrepresentable, making fabrication impossible.
Scope-Aware Tools
PEN-STACK exposes twenty-two tools covering ten genome-writing design stages. These tools are context-aware, meaning they only operate on relevant parameters for a specific design step. This structure allows agents to interact with the system in a controlled, verifiable manner.
Design-Stage Biosecurity Screen
Before any protocol is sent out, a screening mechanism checks the design against safety standards. It maps decisions onto a predefined vocabulary (IBBIS Common Mechanism ScreenStatus) to immediately flag or refuse designs that might pose risks, such as enhancing pandemic transmissibility.
Grounding vs. Prompting
Instead of relying on models to guess necessary data through prompting, PEN-STACK requires agents to ground their answers in tool outputs. This means agents produce verifiable quantities by forcing them to use validated tools rather than just generating text based on a prompt.

Terminology

Summary

PEN-STACK is an open tool layer designed to supply guaranteed provenance to language-model agents engaged in genome writing, functioning as a substrate rather than just an application. p. 0

Where the paper says it

Genome writing sharpens both: a write plan must specify a location, writer enzyme, cargo, and delivery vehicle, all quantitative and interdependent, so without an integrated tool layer, the agent must supply them. We introduce PEN-STACK, an open tool layer that supplies them with guaranteed provenance.

It enforces a no-fabrication invariant: every quantitative value reported must originate from a validated tool call, making ungrounded numbers structurally unrepresentable. p. 1
Where the paper says it

PEN-STACK provides ten genome-writing design stages as twenty-two scope-aware tools, accessible via a software development kit, a Model Context Protocol server, and a Representational State Transfer interface, under a type-enforced invariant: every quantity must originate from a validated tool. Without tools, three model families fabricated 90.

This is achieved through mechanisms such as preventing candidate claims via errors on .as claim, freezing the nested provenance record to ensure immutability, and explicitly returning out-of-scope quantities as known unknowns.

The system exposes ten genome-writing design stages through twenty-two scope-aware tools, accessible via a software development kit, a Model Context Protocol server, and a Representational State Transfer interface. p. 1

Where the paper says it

PEN-STACK provides ten genome-writing design stages as twenty-two scope-aware tools, accessible via a software development kit, a Model Context Protocol server, and a Representational State Transfer interface, under a type-enforced invariant: every quantity must originate from a validated tool. Without tools, three model families fabricated 90.

The core functionality is tested by evaluating whether models can fabricate the specific quantities that only validated tools can produce. p. 9
Where the paper says it

Grounding removes fabrication where prompting does not The principal result concerning the substrate lies in what the tools prevent (Figure 2; Table 2), and we asked whether the no-fabrication property might be an artefact of prompting, model scale, or architecture. Under a naive prompt and given no tools, models asked to produce a genome-writing plan fabricated the six required quantities at 98.

This is demonstrated by measuring fabrication rates across different model families under naive prompting, coaching conditions, and tool-driving conditions. p. 5
Where the paper says it

The rates reported here are computed over 330 of these transcripts, 110 per model: 40 goals under each of the two prompt conditions, together with the 30 ungroundable questions, which are scored under the naive condition only. The deposited corpus holds 448 in total: 420 across the three families reported here, of which the remaining 30 per model are the coached ungroundable transcripts, deposited for completeness but not scored; and 28 from a fourth model.

For instance, in tests involving 40 planning goals requiring six tool-only quantities each, models exhibited high fabrication rates (e. p. 5
Where the paper says it

The corpus consists of forty goals for genome writing. Each goal requires six tool-only quantities that a model cannot determine on its own and that only a validated computation can provide.

g., Claude Haiku-4.5 at 98.8% for naive prompting) unless explicitly coached to refuse ungrounded requests; however, driving the validated tools as agents resulted in the final audit passing in every model family when quantities must be routed through the tools, with a grounded zero serving as a load-bearing comparison. p. 1
Where the paper says it

8% of the 240 required quantities under a naive prompt; coaching left a residual of 0 to 4, with no model certified at zero. Driving the tools, the same models fabricated nothing on a four-goal audit.

The platform incorporates robust safety checks, including a design-stage biosecurity screen that maps decisions to the IBBIS Common Mechanism ScreenStatus vocabulary before any protocol is emitted. p. 4

Where the paper says it

5 This study presents PEN-STACK, an open tool layer for genome writing in which ten design stages are exposed as twenty-two scope-aware tools, reachable from a software development kit (SDK), Model Context Protocol (MCP) server, or a Representational State Transfer (REST) interface, and governed throughout by a single type-enforced no-fabrication invariant together with a designstage biosecurity screen. We describe the substrate and the composition contract it imposes on any client, measure the fabrication it prevents with an adversarial probe across several model providers, report how far the biosecurity screen’s verdicts agree with established synthesis-screening practice, and set out the external evidence for the science these tools incorporate.

The expression-robustness axis was validated at exact-site resolution with a Spearman correlation of rho = 0. p. 1
Where the paper says it

A preemission biosecurity screen matched expert labels on all eight designs. The expression-robustness axis validated at exact-site resolution (ρ = 0.

571 for the human K562 head, although coarser resolution yielded a machine-readable downgrade flag. p. 1
Where the paper says it

571, n = 1,506) but not at the coarser resolution served by default (ρ ≈ 0. 16), which returns a machine-readable downgrade flag.

The system's structure is comprehensive, featuring forty-four REST endpoints and various modules for different tasks (e.g., /generate designs, /predict Digital-twin outcome). Verification mechanisms are provided through endpoints like POST /verify and POST /verify/proof. The system includes oracle capabilities for prediction, such as protein-ligand binding affinity, and a grounded conversational system that derives fields from tools rather than language models. p. 7

Where the paper says it

The conversational agent consumes the substrate and does not constitute it, since any external agent can call the same tools. Code, and not a system prompt, carries the invariant with three mechanisms supporting it.

Supporting data includes five resources: a multi-provider fabrication corpus (448 transcripts), a curated writer-efficiency knowledge base, the writable-genome atlas, the writer atlas, and a derived human K562 position-effect feature table. p. 6

Where the paper says it

The writable-genome atlas holds per-locus safety, expression-robustness, and integrated writability scores across the human genome per cell type; the writer atlas holds 33,370 systems across eight families, together with a 1,058-entity writer-and-vector comparison set; and both are vendored as static artefacts that the runtime never re-derives. The third new resource is a derived human K562 position-effect feature table of 9,298 integrations, which joins per-integration reporter expression from published K562 thousands-of-reporters-integrated-inparallel (TRIP) data [18] to ENCODE chromatin features [19] at exact-site resolution.

The potential implications suggest that the no-fabrication invariant is not specific to genome writing but applies directly to other agentic biology systems where this type-level solution can be imported as code. p. 1
Where the paper says it

PEN-STACK provides that substrate as open, importable code for agentic genomeengineering systems. Keywords genome engineering; agentic artificial intelligence; tool grounding; hallucination; Model Context Protocol; biosecurity screening; calibrated uncertainty; reproducibility Background Genome engineering has evolved from making cuts in DNA to writing specific sequences.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Grounding instead of prompting removes fabrication: the grounded arm, a four-goal audit, warrants replication at the 240-field scale. This allows agents to produce verifiable quantities by requiring every quantity to originate from a validated tool.

  2. Implementing a type-enforced invariant prevents claims: a generative result raises an error on.as claim, and a generated sequence, structure or backbone remains a proposal until writer-verification has scored it against measured data. This makes it impossible for models to assert ungrounded facts as truth.

  3. Providing explicit known unknowns for out-of-scope quantities: Quantities falling outside a tool’s scope card are returned as explicit known-unknowns and are neither guessed nor omitted. This ensures that predictions are not made for unvalidated parameters, preventing the agent from inventing necessary data points.

  4. Utilizing a design-stage biosecurity screen: The screen runs first in verification, before any protocol is emitted, ensuring that designs such as those intended to enhance pandemic transmissibility are refused or flagged immediately by mapping decisions to the IBBIS Common Mechanism ScreenStatus vocabulary.

  5. Enforcing resolution-aware validation flags: The system withholds the validation flag when serving coarse features: The served provenance is resolution-aware, and the system enforces this bound rather than merely disclosing it, meaning a user is never told that a coarse-resolution prediction is validated when it is not.

  6. Creating an external governance layer for safety: A separate governance layer, BioFirewall, consumes plans issued by PEN-STACK to screen them externally so that any tool it supervises, PEN-STACK among them, acquires signed design passports together with tiered access.

  7. Enabling agentic construction via a grounded tool layer: The system provides twenty-two composable, scope-aware tools whose every result is grounded, cited, and non-fabricating, allowing an agent builder to construct a system where the models such a builder will use fabricate almost every quantity a genome-writing plan needs.

  8. Providing replayable fabrication benchmarks: The platform releases the fabrication corpus of 448 transcripts as an immediate, replayable baseline for genomewriting agents, which allows researchers to measure what an ungrounded model asserts when it cannot compute them.

Abstract

Writing a gene into a human genome is a coupled design problem across locus, writer enzyme, cargo and delivery vehicle. Language-model agents are widely applied to such problems, but genome writing has no grounded tool layer. We present PEN-STACK, an open tool layer with enforced provenance that exposes ten design stages as twenty-two scope-aware tools over SDK, Model Context Protocol and REST interfaces. It integrates chromatin, enzyme, off-target, delivery and immunogenicity evidence under one typed contract: every reported quantity originates from a validated tool, and anything out of scope returns an explicit known-unknown. Given no tools and a naive prompt, three model families fabricated 97.1% to 100.0% of 240 required quantities. Coaching against fabrication reduced this to 0 of 240 on each of two families and 5 of 240 on a 1.5B-parameter local model, but each measured zero carries a 95% upper bound of 1.58%, so prompting yields no model-independent guarantee. With tool access on the same forty goals, all three families fabricated nothing, but 81% and 84% of required quantities resolved through a tool call for the two larger models against 27% for the local model: grounding confers fabrication safety, not capability. A pre-emission biosecurity screen matched its own probe specification on all eight designs, a self-consistency check with unmeasured error rates. A human position-effect model validated on held-out chromosomes (Spearman rho = 0.571, n = 1,506) but not on the features served by default (rho 0.16), where the service withholds its validation flag. Ten claims were pre-specified in hash-locked specifications, though priority over scoring is not independently established; every outcome, including eight non-passes, is a machine-readable flag. Insertional-oncogenesis coverage is limited to documented loci. Code and three data resources are released openly.

Sources

Related papers