Lattice Graph × CuspAI

Lattice Graph operates a materials knowledge graph that spans millions of compositions and links composition to crystal structure, thermodynamic and mechanical properties, synthesis routes, and patent claims. What makes this graph unusual is not just breadth but the validation layer sitting beneath every property record: candidate materials are evaluated by multiple independent physics engines — machine-learning interatomic potentials including MACE and CHGNet alongside density functional theory — and a finding is only propagated forward when these engines reach consensus on phonon and thermodynamic stability. For a platform like CuspAI that is generating novel MOF and capture-material candidates at scale, this cross-engine agreement protocol is the difference between a property prediction that has been stress-tested and one that reflects a single model's blind spots. The second structural asset is what Lattice Graph holds in negative space. The public materials literature is almost entirely composed of reported successes: frameworks that synthesized cleanly, sorbents that hit a CO2-uptake target, compositions that survived activation. The far larger population of frameworks that collapsed on activation, lost capacity under humidity cycling, failed synthesis entirely, or were abandoned without publication is effectively invisible to any model trained on the open corpus. Lattice Graph has curated over 23,000 labeled failed-experiment edges — what we call the negatives atlas — capturing exactly this dark matter. For a generative discovery platform, this is the signal that determines whether the model learns to avoid known dead ends or confidently re-proposes them. On top of the knowledge graph and the negatives moat, Lattice Graph provides calibrated uncertainty signals derived from cross-source disagreement in property data. When multiple authoritative sources give materially different values for the same property on the same composition, that disagreement is itself a signal: it means the candidate sits in a region of chemical space where predictions are unreliable and human review is warranted before synthesis investment is made. These trust and disagreement signals, surfaced as calibrated bounds on individual records, give CuspAI the honest error bars that separate a ranked list from a defensible shortlist.

CuspAI has built its platform around a generative-plus-retrieval loop designed to surface viable CO2-capture candidates and novel MOFs faster than conventional screening. The quality of that loop is constrained above all by the data the models learn from — specifically by how well the training and eval corpus represents failure as well as success. A platform that recommends materials should penalize the failure modes that actually kill sorbents in practice: poor synthesis yield, framework collapse on desolvation, rapid capacity degradation under real-flue-gas humidity, and thermal instability under regeneration conditions. None of those failure modes are well-represented in the published literature, because negative results rarely get published. That is the structural gap where Lattice Graph's value sits for CuspAI. We hold labeled negative results at a scale and specificity that cannot be reconstructed from the open corpus, along with cross-source trust signals and a full provenance graph that grounds generated candidates in traceable evidence. For CuspAI, this translates directly to hit-rate credibility: when the platform surfaces a capture-material shortlist to an industrial partner, each hit should arrive with a citation trail, a calibrated confidence level, and confirmation that it does not sit on a known kill edge. That combination — grounded, honest, negative-result-aware recommendations — is what makes a discovery platform defensible to a sophisticated counterparty rather than simply fast. The strategic fit extends to the competitive landscape. As AI-driven materials discovery becomes more crowded, the differentiation that is hardest to replicate is not model architecture; it is curated data that took years and significant experimental infrastructure to accumulate. Licensing Lattice Graph's negatives atlas, trust layer, and knowledge-graph access gives CuspAI's platform a data moat it cannot close through web scraping or literature mining — because the core asset is precisely the results that were never published.

CuspAI has built its platform around a generative-plus-retrieval loop designed to surface viable CO2-capture candidates and novel MOFs faster than conventional screening. The quality of that loop is constrained above all by the data the models learn from — specifically by how well the training and eval corpus represents failure as well as success. A platform that recommends materials should penalize the failure modes that actually kill sorbents in practice: poor synthesis yield, framework collapse on desolvation, rapid capacity degradation under real-flue-gas humidity, and thermal instability under regeneration conditions. None of those failure modes are well-represented in the published literature, because negative results rarely get published. That is the structural gap where Lattice Graph's value sits for CuspAI. We hold labeled negative results at a scale and specificity that cannot be reconstructed from the open corpus, along with cross-source trust signals and a full provenance graph that grounds generated candidates in traceable evidence. For CuspAI, this translates directly to hit-rate credibility: when the platform surfaces a capture-material shortlist to an industrial partner, each hit should arrive with a citation trail, a calibrated confidence level, and confirmation that it does not sit on a known kill edge. That combination — grounded, honest, negative-result-aware recommendations — is what makes a discovery platform defensible to a sophisticated counterparty rather than simply fast. The strategic fit extends to the competitive landscape. As AI-driven materials discovery becomes more crowded, the differentiation that is hardest to replicate is not model architecture; it is curated data that took years and significant experimental infrastructure to accumulate. Licensing Lattice Graph's negatives atlas, trust layer, and knowledge-graph access gives CuspAI's platform a data moat it cannot close through web scraping or literature mining — because the core asset is precisely the results that were never published.

Beyond the licensed data products, CuspAI's scientists and ML engineers have direct access to the Lattice Graph web application for interactive exploration and batch screening. The knowledge-graph explorer lets a researcher walk the composition-structure-property-patent-synthesis graph around any generated candidate: inspecting the evidence neighborhood, viewing trust and disagreement flags on individual property values, and tracing provenance from a predicted property back to the original experimental or computational source. This surface is particularly useful for auditing model outputs that are technically plausible but sit in sparsely covered regions of chemical space — the explorer makes coverage gaps visible before they become credibility gaps in a partner presentation. For throughput work, the batch screening dashboard allows CuspAI to push an entire generated shortlist through negatives filtering, disagreement scoring, and provenance grounding in a single operation, rather than resolving one candidate at a time. The composition-intelligence reports produced by this workflow give CuspAI a structured output that shows, for each candidate, its kill-edge status, cross-source agreement level, evidence density, and patent exposure — a package that is both internally useful for triage and externally presentable to industrial partners evaluating the platform's recommendations. The synthesis-aware views in the application, which surface known recipe paths and flag compositions with no credible synthesis route in the graph, are especially relevant for MOF-focused work where the gap between a thermodynamically interesting framework and a practically synthesizable one is often large.

The natural entry point is a scoped negatives audit and data-and-eval license, which the context estimates at $40,000 to $75,000 to start (this is a framing range, not a committed price). In that engagement, CuspAI receives a capture- and MOF-scoped export of the Negatives and Eval-Data Atlas along with access to the Trust and Disagreement Signals product and metered Knowledge-Graph API access for provenance grounding during the evaluation period. The deliverable from CuspAI's side is a measurement of lift: how many candidates in a recent generated shortlist would have been caught by the kill-edge filter, and how the model's calibration changes when hard negatives are folded into training. That measurement converts the data license from a theoretical benefit into a demonstrated improvement in hit rate, and it gives CuspAI a concrete result to share with its own partners and investors. If the audit demonstrates measurable value — which the negatives coverage and trust signals are structured to produce — the relationship scales into a recurring data-and-eval-license subscription: ongoing access to refreshed negatives lanes as the atlas is updated, a live trust-and-disagreement feed keyed to CuspAI's composition universe, and metered KG API queries sized to the platform's training cadence and query volume. Larger annual data-license structures, which across comparable engagements fall in the range of six figures, are scoped by lane coverage depth and update frequency rather than fixed list pricing. All figures here are estimates for planning purposes; final scope and pricing are established during the paid evaluation.