Lattice Graph × SandboxAQ

Lattice Graph operates a materials knowledge graph spanning millions of compositions, knitting together experimental records, computed properties, synthesis conditions, and patent claims into a single governed structure. Every composition in the graph carries provenance — the source, the method, the conditions — so that any property value can be traced to the evidence that supports it rather than treated as an undifferentiated number. That architecture makes the graph useful not just for lookup but for systematic disagreement analysis: where multiple independent physics engines or literature sources report conflicting values for the same material, we flag the tension and quantify it rather than resolving it arbitrarily. Multi-engine validation sits at the core of how we establish confidence in new candidate materials. We run predictions through multiple machine-learning interatomic potentials — including MACE and CHGNet — alongside density functional theory calculations, requiring phonon stability and thermodynamic consistency to hold across independent methods before a candidate earns a positive signal. The consensus criterion is the key detail: agreement between methods is treated as evidence of reliability, and disagreement is surfaced as a calibrated uncertainty signal rather than hidden. For an organization building enterprise simulation products, that distinction is the difference between a trustworthy engineering tool and a research demo. The asset that most organizations cannot replicate is our labeled negative corpus — more than 23,000 failed-experiment and kill edges catalogued from internal programs and sources that never appear in the published DFT literature. Positive results get published; negative results do not. Our negative atlas directly addresses the selection bias baked into every model trained on the open computational record, giving practitioners a ground truth for what the field has already tried and retired. Paired with cross-source trust signals and the full knowledge-graph API, this constitutes a calibration and benchmark infrastructure that is difficult to assemble from public data alone.

SandboxAQ's decision to build Large Quantitative Models for materials and chemistry simulation represents a bet that quantum-informed, physics-grounded AI can meet enterprise credibility standards — a higher bar than academic benchmarks, because enterprise customers will eventually test model confidence against real experimental programs. The company has the physics depth and computational infrastructure to generate predictions at scale; what is harder to accumulate is the labeled failure record that tells a model what chemistries have already been ruled out, and an independent uncertainty layer that lets a customer trust the model's own reported confidence. Those two signals are structural gaps in the public data ecosystem, not gaps specific to SandboxAQ. Lattice Graph addresses both gaps directly, without competing with SandboxAQ's modeling stack. We are a data and calibration supplier, not a simulation vendor. Our negatives corpus provides the labeled failure record that publication bias systematically excludes, and our cross-source trust and disagreement signals provide a source-aware calibration target that SandboxAQ can score their LQM's uncertainty estimates against. The knowledge graph ties everything to provenance, so when SandboxAQ's enterprise customers ask how a prediction was validated, the answer is auditable rather than opaque. The strategic fit is precise: SandboxAQ is at the moment in its commercial trajectory where a model's benchmark performance needs to hold up under scrutiny from sophisticated buyers in pharma, energy, and advanced materials. A positives-only training and evaluation record looks strong in the lab and becomes fragile the first time a customer asks about a composition that the experimental record has already killed. Licensing the negatives moat and trust layer before paid pilots rather than after a credibility event is the cleanest path to an enterprise-grade accuracy story.

SandboxAQ's decision to build Large Quantitative Models for materials and chemistry simulation represents a bet that quantum-informed, physics-grounded AI can meet enterprise credibility standards — a higher bar than academic benchmarks, because enterprise customers will eventually test model confidence against real experimental programs. The company has the physics depth and computational infrastructure to generate predictions at scale; what is harder to accumulate is the labeled failure record that tells a model what chemistries have already been ruled out, and an independent uncertainty layer that lets a customer trust the model's own reported confidence. Those two signals are structural gaps in the public data ecosystem, not gaps specific to SandboxAQ. Lattice Graph addresses both gaps directly, without competing with SandboxAQ's modeling stack. We are a data and calibration supplier, not a simulation vendor. Our negatives corpus provides the labeled failure record that publication bias systematically excludes, and our cross-source trust and disagreement signals provide a source-aware calibration target that SandboxAQ can score their LQM's uncertainty estimates against. The knowledge graph ties everything to provenance, so when SandboxAQ's enterprise customers ask how a prediction was validated, the answer is auditable rather than opaque. The strategic fit is precise: SandboxAQ is at the moment in its commercial trajectory where a model's benchmark performance needs to hold up under scrutiny from sophisticated buyers in pharma, energy, and advanced materials. A positives-only training and evaluation record looks strong in the lab and becomes fragile the first time a customer asks about a composition that the experimental record has already killed. Licensing the negatives moat and trust layer before paid pilots rather than after a credibility event is the cleanest path to an enterprise-grade accuracy story.

For SandboxAQ's research and machine-learning engineering teams, the most practical daily surfaces are the knowledge-graph explorer and the composition-intelligence reports. The graph explorer lets researchers visually navigate the evidence neighborhood around a predicted material — seeing which sources agree, which disagree, and what the disagreement flags look like for the chemistries an LQM is being benchmarked on. Composition-intelligence reports surface a full 360-degree view of a composition: structure variants, property distributions across sources, synthesis conditions, and patent encumbrance, all in one place. These are the tools for triage work: when a benchmark number looks suspicious or an LQM output diverges from graph evidence, the explorer and the composition report are where a researcher goes to understand why. At the workflow level, SandboxAQ's teams can run batches of LQM candidate proposals through the platform's batch screening interface, scoring each against the negatives atlas and the cross-source trust signals in a single pass. The platform's formation-energy predictions serve as an independent cross-check against LQM outputs, and the negatives coverage statistics give eval leads an at-a-glance view of how much of a given chemistry lane is represented in the labeled failure record before they stake a benchmark claim on it. The result is an integrated eval-ops workflow rather than a collection of one-off API calls, which matters as SandboxAQ's eval surface grows across more chemistry families and customer verticals.

This engagement is structured as a data and evaluation license, not a patent or asset transaction. The natural entry point is a time-boxed scoped evaluation: SandboxAQ licenses the Negatives and Eval-Data Atlas and Trust and Disagreement Signals for one or two chemistry lanes — for example electrolytes and cathode materials — runs them against a current LQM benchmark, and quantifies how labeled negatives and disagreement flags shift measured accuracy and calibration scores. The goal of the scoped evaluation is to produce a documented, lane-specific accuracy delta that SandboxAQ can use internally to justify the expansion and, over time, share with enterprise customers as part of its model validation story. Given SandboxAQ's scale and the breadth of chemistry lanes across its enterprise products, the pricing shape is a bespoke data license rather than a fixed-fee audit; a reasonable starting estimate for an initial scoped engagement is in the low-to-mid six figures, with an expansion path to a recurring full-atlas plus knowledge-graph API license priced against usage volume and lane coverage. The expansion structure would bundle negatives atlas data, trust and disagreement signals, and knowledge-graph API access into a recurring license with refreshed negatives drops as internal programs add to the failure record. Lattice Graph does not co-develop models or take IP positions in SandboxAQ's stack; the relationship is a supplier arrangement — we provide the labeled failure corpus, the calibration signals, and the provenance infrastructure, and SandboxAQ incorporates that into its own training, evaluation, and enterprise product workflows.