Lattice Graph × Radical AI

Lattice Graph operates a materials knowledge graph spanning millions of compositions, connecting structures, properties, synthesis routes, and provenance across the published and proprietary scientific record. What makes that graph useful to an organization like Radical AI is not just the breadth of positive results — it is the honesty of what we have curated alongside them: more than 23,000 labeled failure edges, kill results, and documented dead ends that the public literature systematically buries. Every node in the graph carries cross-source evidence, disagreement flags, and calibrated prediction bounds derived from comparing multiple independent property sources, so the graph tells you not just what is known but how well it is known and where the sources diverge. On the physics side, we validate candidate materials through multiple independent computational engines — machine-learning interatomic potentials including MACE and CHGNet, plus density functional theory — and require consensus across phonon spectra and thermodynamic stability before elevating a candidate. That multi-engine requirement is what generates our disagreement signals in a principled way: when MACE and DFT disagree on a formation energy or a phonon mode, that disagreement is recorded, flagged, and retrievable. For a materials foundation model, those cross-engine disagreement signals are as informative as the consensus results, because they mark the regions of composition space where the evidence is thin, contradictory, or genuinely hard. We also maintain a freedom-to-operate screening layer across more than 300,000 materials patents, which is less central for a data-and-evaluation use case but becomes relevant when Radical's models produce a candidate that looks novel and their customers ask whether the IP path is clear. The combination — a provenance-grounded knowledge graph, an honest negatives atlas, calibrated uncertainty signals, and patent coverage — gives Radical a data supplier whose curation decisions are independent of the public sources their models already trained on.

Radical AI is building the infrastructure layer for autonomous materials discovery: foundation models that propose candidates, self-driving labs that run them, and a feedback loop that tightens with every experiment. The economic return on that loop depends almost entirely on two things — the quality of the data the model learns from, and the honesty of the signal it is evaluated against. Both of those are exactly where publicly available materials data falls short. The open literature is structurally biased toward successes and computed properties; failed experiments rarely get published, negative results do not appear in databases, and benchmarks assembled from the same public sources competitors trained on cannot distinguish genuine generalization from distribution-memorization. Lattice Graph's role here is narrow and complementary: we are a data and evaluation supplier, not a foundation model builder. The gap we fill is the one Radical's own pipeline cannot easily close from public sources — labeled negatives, cross-source disagreement quantification, and a knowledge graph with real provenance trails rather than scraped values of uncertain origin. A self-driving lab runs physical experiments at real cost; a pre-screen against documented failure regions and a calibrated uncertainty layer that flags where the model is extrapolating both reduce the rate at which expensive robotic runs return uninformative results. The strategic positioning is also clean. Radical's customers and investors increasingly ask for generalization proofs that go beyond benchmarks constructed from the same public corpora the model trained on. An evaluation set built from Lattice Graph's negatives atlas is genuinely independent because most of those failure results were never published. That independence is hard to replicate, which makes it durable as a differentiator in Radical's own sales and funding conversations.

Radical AI is building the infrastructure layer for autonomous materials discovery: foundation models that propose candidates, self-driving labs that run them, and a feedback loop that tightens with every experiment. The economic return on that loop depends almost entirely on two things — the quality of the data the model learns from, and the honesty of the signal it is evaluated against. Both of those are exactly where publicly available materials data falls short. The open literature is structurally biased toward successes and computed properties; failed experiments rarely get published, negative results do not appear in databases, and benchmarks assembled from the same public sources competitors trained on cannot distinguish genuine generalization from distribution-memorization. Lattice Graph's role here is narrow and complementary: we are a data and evaluation supplier, not a foundation model builder. The gap we fill is the one Radical's own pipeline cannot easily close from public sources — labeled negatives, cross-source disagreement quantification, and a knowledge graph with real provenance trails rather than scraped values of uncertain origin. A self-driving lab runs physical experiments at real cost; a pre-screen against documented failure regions and a calibrated uncertainty layer that flags where the model is extrapolating both reduce the rate at which expensive robotic runs return uninformative results. The strategic positioning is also clean. Radical's customers and investors increasingly ask for generalization proofs that go beyond benchmarks constructed from the same public corpora the model trained on. An evaluation set built from Lattice Graph's negatives atlas is genuinely independent because most of those failure results were never published. That independence is hard to replicate, which makes it durable as a differentiator in Radical's own sales and funding conversations.

For Radical's scientists and model engineers working interactively, the knowledge-graph explorer dashboard surfaces the same evidence their models consume in a form researchers can inspect and interrogate. Composition-360 views show all structure, property, synthesis, and provenance data associated with a formula in one place, and evidence neighborhood maps let a scientist understand why a model call looks surprising — what the supporting evidence actually is, where it conflicts, and which sources the graph trusts most. Dossiers on individual candidate materials compile that evidence automatically, so the bottleneck of manually tracing provenance disappears. The batch-screening workflow is the most operationally important surface for a self-driving lab context. Teams can submit a candidate list — the output of a foundation model's proposal step — and run it in bulk against the negatives atlas and trust signals before any experiment is queued. The workflow returns a ranked view that separates candidates already in documented failure zones, candidates in high-disagreement regions worth investigating, and candidates with clean provenance and consensus support. That ranked output maps directly onto the robotic queue prioritization problem, which is precisely the bottleneck that makes or breaks throughput in an autonomous lab.

The right structure for a Radical AI engagement is a data and evaluation license, not asset licensing or co-development. The recommended starting point is a scoped negatives audit: Lattice Graph runs Radical's recent candidate list or proposal set against the negatives atlas and trust signals, quantifies the overlap with documented failure regions and high-disagreement zones, and delivers an independent written assessment of how much physical experimentation that pre-screen would have flagged in advance. Context on this engagement suggests a starting range of $40,000 to $75,000 for the audit phase. That scoped engagement converts naturally into an ongoing data and evaluation license covering API access to the negatives atlas, the trust and disagreement signals layer, and the knowledge-graph API as a provenance and grounding resource. Deliverables from the audit phase include a failure-region overlap report, a calibrated uncertainty map over the submitted candidate set, and a quantified estimate of experiment-run savings attributable to the pre-screen. From there, the ongoing license can be structured as a held-out evaluation set, an always-on pre-screen integrated into Radical's candidate pipeline, or both. Volume, exclusivity, and refresh cadence are all negotiable in a direct conversation, and all figures cited here are estimates for planning purposes rather than quoted prices.