Let the Model Design the Poster—Not the Evidence

TL;DR for operators

A scientific poster can look polished and still fabricate the plots or diagrams readers interpret as evidence. PosterHarness separates those responsibilities: the image model designs the layout and decides where evidence should appear, but it must leave those regions blank for source-paper figures to be inserted later by deterministic code.

In a three-paper prompt probe, direct poster generation produced 34 synthesized scientific figures. An explicit placeholder contract reduced that count to zero while preserving all 10 expected placeholder IDs, blank slots, and aspect labels. This shows that poster reliability can be expressed through observable rules rather than entrusted entirely to better prompting.

The contract did not solve full-poster reliability. Data-like decorations remained, domain-specific styling reintroduced scientific-looking elements, and a correctly sourced figure did not guarantee that the surrounding text, caption, or claim was accurate. Source-figure provenance therefore establishes where an inserted asset came from, not that every statement around it is correct.

For operators, the decision is whether to approve a polished generative output directly or redesign the workflow around a placeholder-first architecture: generative composition followed by verified figure insertion, contract checks, rejection logic, and auditable manifests. The broader lesson is to let the model design the poster—not the evidence.

A polished poster can still fabricate its evidence

A team building an AI poster tool faces two demands that do not naturally align. The poster must look designed rather than assembled from rigid boxes. It must also avoid inventing plots, diagrams, axes, or experimental results that resemble scientific evidence.

A stronger image model does not remove that conflict. In the paper’s small prompt-component probe, direct generation produced 34 synthesized scientific figures and 38 data-like decorations across three papers. These were not merely stylistic flourishes. They occupied the visual role that readers normally treat as evidence.

The first operational decision, then, is not which model produces the most attractive poster. It is whether the model is allowed to create evidence-bearing visuals at all.

PosterHarness answers by restricting the model’s authority. The model designs where figures should go, but it does not draw them. It must produce labeled, blank placeholders with declared aspect ratios. A deterministic compositor later inserts figures extracted from the paper.

This is the paper’s central contribution: scientific poster generation becomes a set of observable contracts. A slot can be missing. Its identifier can be wrong. It can contain prohibited graphics. Its shape can violate the requested ratio. An inserted figure can overflow its boundary. Each condition can be checked, logged, retried, or rejected.

The placeholder contract changes the failure surface

The three-paper prompt probe is best read as limited mechanism evidence. It compares four conditions: direct generation, basic placeholders, an explicit placeholder contract, and a domain-specific contract.

Basic placeholders were not enough. They produced all 10 expected IDs and blank slots, but none of the required aspect-ratio labels. The VLM auditor still counted seven synthesized scientific figures.

The explicit contract produced all 10 IDs, all 10 blank regions, and all 10 aspect labels, while reducing synthesized figures to zero. Four data-like decorations remained outside the figure slots. Adding domain grammar preserved placeholder compliance but reintroduced two synthesized figures and increased data-like decoration to 10.

That non-monotonic result is useful. Domain adaptation can improve field-specific appearance while weakening abstention behavior. A physics-flavored visual grammar, for example, may encourage scientific-looking motifs that resemble plots or data displays even when the prompt still preserves blank placeholders.

The practical response is not to abandon domain profiles. It is to evaluate domain adaptation separately from evidence controls. A profile can specify narrative order, density, terminology, and visual conventions, but a second rule must prohibit data-bearing decoration wherever it appears. One instruction cannot safely stand in for the other.

The 12-paper pilot shows a trade-off, not a winner

The main benchmark pairs PosterHarness and Paper2Poster on the same 12 curated papers: six high-energy physics papers, five CS/ML papers, and one biology/ML paper. Both systems produced 12 accepted final PNGs, and both received a 12/12 source-figure provenance proxy after correction of the Paper2Poster cache.

The systems differ in architecture and output medium, so the comparison does not isolate a single component. PosterHarness generates a raster editorial design and inserts verified figures. Paper2Poster follows a more deterministic PPTX-style pipeline. The results characterize two operating regimes.

The visual metrics show a different design regime, not scientific usefulness by themselves. Higher color entropy, lower white space, and greater resolution indicate richer artifacts. They do not establish that researchers can find claims faster, interpret figures more accurately, or remember more of the paper.

The single-VLM preference result carries a similar boundary. The judge came from the same model family used in the generation pipeline, and the paper reports no human evaluation. The finding supports visual preference within this pilot, not human utility across scientific communication settings.

Paper2Poster’s slightly stronger information score also deserves attention. A poster can look more editorial while carrying less answerable content. For teams deploying visual summaries, aesthetic quality and information transfer should be separate acceptance criteria.

Auditability has a visible operating cost

PosterHarness accepted 12 final posters, but the run manifests recorded 22 selected-paper job attempts: 12 successes, eight failures or exceptions, and two cached or skipped attempts. Among failed jobs, seven contained placeholder-geometry rejection signals, six contained strict placeholder-QA rejection, and five contained template-critic rejection. Counts overlap because one failed job can trigger several categories.

Those rejected attempts are not incidental noise. They are part of the control system.

A workflow that hides retries behind a final success rate can make a brittle generator appear dependable. PosterHarness instead exposes geometry failures, dirty placeholders, critic failures, public-text issues, infrastructure errors, and launcher exceptions as distinct states. That improves diagnosis and governance, but it also raises latency and compute cost.

For an operations team, the relevant unit is therefore not “one poster generated.” It is the distribution of attempts, rejection causes, final acceptance rate, and time to an accepted artifact. The paper’s mean of 27.96 minutes ranged from 10.19 to 64.77 minutes, which is compatible with offline publishing but less suitable for interactive authoring.

Provenance is narrower than factual reliability

The most likely overclaim is that deterministic source-figure insertion makes the poster trustworthy. It does not.

PosterHarness can provide a provenance proxy for accepted figure regions, conditional on correct extraction, figure assignment, placeholder detection, insertion, and containment checks. It can say, in effect, that an accepted region contains a source asset rather than a model-drawn substitute.

It cannot certify that the selected figure is the right one for the surrounding claim. It cannot guarantee that a caption is accurate, a number is copied correctly, a conclusion is supported, or the generated text faithfully represents the paper. Upstream planning and extraction can also fail.

This narrower definition is still valuable because it attaches a specific guarantee to a specific region. It replaces a vague claim of “grounded output” with an inspectable chain: source asset, assignment record, placeholder ID, geometry check, inserted crop, and manifest status.

What product and governance teams can take from the architecture

What the paper directly shows: in a small probe, explicit placeholder instructions sharply reduced synthesized scientific figures; in a curated 12-paper pilot, the full system produced visually richer outputs, exposed rejected attempts, and preserved accepted-region source provenance proxies. It also ran slower and retained slightly less answerable information than Paper2Poster.

Cognaptus inference for business use: the architecture generalizes beyond posters. A generative model can control composition, wording suggestions, or presentation while verified components supply charts, numbers, citations, product identifiers, compliance text, and other evidence-bearing elements. Contracts should be machine-checkable, and rejection states should remain visible in logs and dashboards.

This pattern is relevant to research summaries, financial reports, medical communication, sales collateral, and executive dashboards. The affected decision is whether a polished output can be approved directly. The safer alternative, when unsupported evidence would be costly, is to require verified insertion and explicit acceptance checks before release.

What remains uncertain: the benchmark does not establish population-level performance across scientific domains. It does not quantify the benefit of domain profiles, isolate the critic loop, test multiple judges, or validate human preference. The current raster workflow also remains slower, costly, and less editable than a mature authoring tool.

The product is the acceptance process

PosterHarness is most useful when read as a governance architecture rather than a claim that one poster generator has solved scientific communication.

Its key move is to deny the image model authority over scientific evidence while preserving its ability to design the page. The benchmark then evaluates compliance through blank slots, identifiers, aspect labels, prohibited decoration, geometry, containment, retries, and manifests.

That design creates a more defensible output, but not a fully certified one. Teams adopting the pattern still need separate checks for text, numbers, captions, conclusions, and figure selection. They must also decide whether the added latency and rejection handling are justified by the risk of fabricated visual evidence.

The broader lesson is practical: when generative quality and factual control pull in different directions, reliability is more likely to come from assigning them to different components than from asking one model to balance both inside a prompt.

Cognaptus: Automate the Present, Incubate the Future.