Lattice Graph × Periodic Labs

Lattice Graph is a computational materials-discovery platform built around a knowledge graph spanning millions of compositions, connecting structures, synthesis routes, experimental measurements, theoretical predictions, and patent records into a single queryable fabric. Most platforms show you what succeeded; ours also encodes what failed. We hold 23,196 labeled kill edges — failed-experiment records that document where materials don't work, under what conditions, and why — a dataset class that has largely been kept internal and is absent from every major public corpus. For a team building foundation models for materials, that distinction is not cosmetic. The real decision boundary for a model is defined at least as much by the negatives as by the positives, and the negatives are precisely what survivorship-biased public data sources like the Materials Project or the open literature cannot provide. Our validation stack pairs machine-learning interatomic potentials — MACE, CHGNet, and related architectures — with density functional theory to reach multi-engine consensus on phonon and thermodynamic stability. Where those engines disagree, we generate explicit cross-source disagreement signals and calibrated prediction bounds. For a model-evaluation team, these high-disagreement compositions are gold: they mark the regime where model confidence is hardest to earn honestly and easiest to fake with clean training data. We expose those uncertainty signals as a structured evaluation layer so teams can build benchmarks that actually stress-test calibration rather than just measuring accuracy on the easy cases. The knowledge graph layer ties all of this together with provenance. Every negative, every disagreement flag, every property record traces back to its source so researchers can audit a data point before it enters a training or evaluation split. Composition-360 views, evidence neighborhoods, and natural-language graph queries let teams interrogate the graph interactively — not just retrieve records but explore the context around a candidate and understand why a particular composition sits in the uncertain or the failed region of the space.

Periodic Labs has made a public bet that the field's foundation models are trained on a fundamentally incomplete signal — that positive experimental outcomes dominate public corpora while failed runs, dead ends, and synthesis routes that don't close go largely unrecorded and unlabeled. That argument is correct, and Lattice Graph is one of the few places that holds the other side of the ledger at meaningful scale. The strategic fit is direct: Periodic needs labeled negatives as a training and benchmark-hardening input, and we hold 23,196 of them, largely internal and unavailable from any public scraping exercise. The deeper alignment is epistemological. Periodic's autonomous-lab thesis assumes that experimental throughput is the input worth maximizing, but throughput generates value only when the model directing experiments is well-calibrated about uncertainty — when it knows what it doesn't know and stops sending the robot after already-known dead ends. Our trust and disagreement signals operationalize exactly that calibration layer: compositions where DFT sources cross-disagree or where measured and predicted values diverge are the cases where model uncertainty claims are hardest to validate, and where an honest eval set needs to concentrate. Licensing those signals lets Periodic build benchmarks that distinguish a model that is genuinely uncertain from one that is confidently wrong. Lattice Graph's role for Periodic is as a data and evaluation partner, not an IP or asset supplier. The patent-whitespace tooling and invention portfolios that serve other customer archetypes are not the relevant surface here. What matters is the labeled-negatives corpus, the uncertainty and disagreement layer, and the knowledge-graph API for provenance — a combination that no public corpus delivers and that Periodic cannot reconstruct from in-house runs without re-spending compute on known dead ends at significant cost.

Periodic Labs has made a public bet that the field's foundation models are trained on a fundamentally incomplete signal — that positive experimental outcomes dominate public corpora while failed runs, dead ends, and synthesis routes that don't close go largely unrecorded and unlabeled. That argument is correct, and Lattice Graph is one of the few places that holds the other side of the ledger at meaningful scale. The strategic fit is direct: Periodic needs labeled negatives as a training and benchmark-hardening input, and we hold 23,196 of them, largely internal and unavailable from any public scraping exercise. The deeper alignment is epistemological. Periodic's autonomous-lab thesis assumes that experimental throughput is the input worth maximizing, but throughput generates value only when the model directing experiments is well-calibrated about uncertainty — when it knows what it doesn't know and stops sending the robot after already-known dead ends. Our trust and disagreement signals operationalize exactly that calibration layer: compositions where DFT sources cross-disagree or where measured and predicted values diverge are the cases where model uncertainty claims are hardest to validate, and where an honest eval set needs to concentrate. Licensing those signals lets Periodic build benchmarks that distinguish a model that is genuinely uncertain from one that is confidently wrong. Lattice Graph's role for Periodic is as a data and evaluation partner, not an IP or asset supplier. The patent-whitespace tooling and invention portfolios that serve other customer archetypes are not the relevant surface here. What matters is the labeled-negatives corpus, the uncertainty and disagreement layer, and the knowledge-graph API for provenance — a combination that no public corpus delivers and that Periodic cannot reconstruct from in-house runs without re-spending compute on known dead ends at significant cost.

Periodic's team spends most of their platform time in the data and evaluation surfaces rather than the IP or workflow dashboards that other customer archetypes use. The knowledge-graph explorer is the primary analytical environment: researchers navigate composition-360 views and evidence neighborhoods interactively, inspecting the provenance and experimental context around a record before it enters a training or evaluation split. This is especially useful during dataset curation, where the difference between a high-quality negative and a mislabeled or ambiguous result determines downstream benchmark validity. The trust and disagreement view operationalizes cross-source uncertainty scoring as a visual triage surface, letting the team sort candidate compositions by disagreement severity and identify the clusters where model calibration claims are most vulnerable to challenge. On the operational side, the negatives atlas export with lane-level slicing supports repeatable, versioned dataset construction — teams can pull defined coverage slices, version them for reproducibility, and refresh them as the atlas grows with new kill edges. Batch screening against the negatives set lets Periodic pre-filter large candidate lists from their generative models before allocating autonomous-lab runs, so experimental capacity concentrates on genuinely uncertain territory. Provenance reporting gives a defensible audit trail for any negative or disagreement signal cited in model documentation or shared with external evaluators, which becomes increasingly important as foundation-model claims face more rigorous external scrutiny.

The natural engagement structure for Periodic is a scoped data and evaluation license covering the Negatives and Eval-Data Atlas together with the Trust and Disagreement Signals, delivered with Knowledge-Graph API access for provenance grounding under defined training and evaluation usage terms. A sensible entry point is a time-boxed evaluation phase: Periodic receives a held-out sample from the negatives corpus and the high-disagreement signal set, runs it against their existing benchmarks and model pipelines, and measures the calibration lift before committing to the full license. This removes the need to take the asset's value on faith and gives both teams a concrete, reproducible measurement of what the negatives corpus actually moves. The context frames the headline engagement at an estimated range of $100,000 to $500,000 for the data license — that is a framing estimate and not a committed price. After the evaluation phase, the ongoing engagement becomes a recurring license with versioned export access and trust-signal refreshes as the atlas expands. The Negatives and Eval-Data Atlas and the Trust and Disagreement Signals are the core licensed entitlements; Knowledge-Graph API access is included for grounding and audit purposes. The engagement does not require involvement of Lattice Graph's patent-screening or invention-portfolio products, keeping the scope clean and the procurement path straightforward. As the autonomous-lab pipeline generates new experimental runs, Periodic and Lattice Graph can discuss structured data-exchange arrangements that allow kill edges from Periodic's own hardware to be incorporated into the atlas, expanding coverage and potentially reducing license costs on a contributed-data basis.