Lattice Graph × Flagship Pioneering

Lattice Graph is a computational materials-discovery platform built around a governed knowledge graph that spans millions of compositions and connects them to structures, properties, synthesis routes, and patent records in a single queryable fabric. The graph's depth comes from its integration of sources that most informatics platforms treat as separate silos: computed DFT results, machine-learning interatomic potentials including MACE and CHGNet, experimental property records, and supply-chain data — all reconciled to a common provenance layer so that every property claim can be traced to the source and the method that produced it. Candidate validation is the other core capability. Rather than relying on a single model or engine, Lattice Graph runs multiple independent physics engines in parallel and requires consensus across them before a candidate is promoted. Phonon stability, thermodynamic stability, and mechanical property targets are all cross-checked against at least two independent predictors, which means that cherry-picked results from whichever engine returned the best number become visible as outliers rather than conclusions. For an evaluator asking whether a target's computational results are calibrated or opportunistic, that cross-engine disagreement signal is a direct, quantitative answer. Lattice Graph also screens intellectual property at scale, covering more than 300,000 materials patents to map freedom to operate and identify genuine whitespace. And critically, the platform holds a curated atlas of labeled negative results — failed experiments and kill edges that the public scientific literature systematically underreports. Where most foundation models and most company pitches are trained or built entirely on positive outcomes, Lattice Graph's negative-result coverage is a structural differentiator: it makes it possible to test whether a model or a thesis is actually learning something the public corpus does not already give away, rather than reproducing survivorship bias at scale.

Flagship Pioneering's company-creation model imposes a distinctive diligence problem. At the ProtoCo stage, the relevant question is not whether a proposed materials thesis sounds defensible in a deck — it is whether the underlying science is genuinely inventable, whether the computational edge is real or re-derived from widely shared public data, and whether there are specific, well-capitalized buyers who demonstrably pay for the class of material in question. Those questions must be answered early, when the decision to harden or abandon a thesis is still cheap to change, and they must be answered from a source that is independent of the ProtoCo team's own models and incentives. The same pressure applies to external diligence. As AI-for-materials has made it inexpensive to generate a plausible-looking platform story from public data, the cohort of inbound AI-materials and critical-minerals companies whose decks Flagship evaluates now includes a substantial fraction whose computational edge, on close inspection, is a re-narration of the public corpus rather than a genuinely proprietary signal. Distinguishing those two cases requires an independent technical baseline that does not itself depend on the public positive-result literature — which is precisely what Flagship cannot get from a reference call with the company's own scientific advisors or from a standard quantitative diligence checklist. Lattice Graph occupies that position. Its knowledge graph, negative-result atlas, and opportunity-ranking infrastructure are maintained independently of any ProtoCo or portfolio company, and the 23,196 kill-edge negatives that underpin the technical-defensibility screen are largely absent from the public literature, making them a hard-to-game benchmark. A foundry that applies this read consistently — across both the theses it originates internally and the companies it evaluates externally — builds a calibrated, reproducible diligence standard at the moment in the materials-AI cycle when that standard is most valuable and most scarce.

Flagship Pioneering's company-creation model imposes a distinctive diligence problem. At the ProtoCo stage, the relevant question is not whether a proposed materials thesis sounds defensible in a deck — it is whether the underlying science is genuinely inventable, whether the computational edge is real or re-derived from widely shared public data, and whether there are specific, well-capitalized buyers who demonstrably pay for the class of material in question. Those questions must be answered early, when the decision to harden or abandon a thesis is still cheap to change, and they must be answered from a source that is independent of the ProtoCo team's own models and incentives. The same pressure applies to external diligence. As AI-for-materials has made it inexpensive to generate a plausible-looking platform story from public data, the cohort of inbound AI-materials and critical-minerals companies whose decks Flagship evaluates now includes a substantial fraction whose computational edge, on close inspection, is a re-narration of the public corpus rather than a genuinely proprietary signal. Distinguishing those two cases requires an independent technical baseline that does not itself depend on the public positive-result literature — which is precisely what Flagship cannot get from a reference call with the company's own scientific advisors or from a standard quantitative diligence checklist. Lattice Graph occupies that position. Its knowledge graph, negative-result atlas, and opportunity-ranking infrastructure are maintained independently of any ProtoCo or portfolio company, and the 23,196 kill-edge negatives that underpin the technical-defensibility screen are largely absent from the public literature, making them a hard-to-game benchmark. A foundry that applies this read consistently — across both the theses it originates internally and the companies it evaluates externally — builds a calibrated, reproducible diligence standard at the moment in the materials-AI cycle when that standard is most valuable and most scarce.

For Flagship scientists hardening a ProtoCo thesis, the most-used surfaces are the opportunity and funded-buyer leaderboard views and the knowledge-graph explorer. The leaderboard gives an analyst an immediate ranked read on where a proposed materials lane sits in the whitespace landscape and which known buyers have funded or purchased in adjacent areas — a framing exercise that takes hours rather than weeks and drops cleanly into a ProtoCo memo or an investment-committee exhibit. The knowledge-graph explorer lets a scientist trace any specific claim — a composition, a stability result, a property target — back to its provenance, its source method, and any negative results attached to it in the atlas, so that "our model shows this is stable" becomes "our model shows this is stable, and here is whether the kill-edge record contradicts that." For partners and analysts evaluating outside deals, the composition-intelligence reports and the cross-engine trust and disagreement dashboard are the most practical daily tools. A one-page composition report, generated from the knowledge graph, gives a provenance-backed summary of what is known, computed, and patented for a given material and drops directly into a data-room review. The disagreement dashboard surfaces candidates where different physics engines return conflicting predictions — the signal most likely to indicate that a target's results are calibrated against the most favorable engine rather than validated across independent methods. Together these surfaces let one analyst, without a standing materials-informatics team or wet-lab infrastructure, produce a technical exhibit that is reproducible and auditable.

The natural entry for Flagship is a scoped per-thesis engagement rather than a broad platform subscription. A single paid pass covers one ProtoCo thesis Flagship is actively hardening, or one outside AI-materials or critical-minerals target under evaluation. Lattice Graph runs the opportunity and buyer-affinity read, applies the negatives-atlas reproducibility test against the relevant candidate list, and delivers an independent technical memo structured around the three questions a foundry needs answered: is the opportunity genuinely inventable in whitespace, does the computational edge survive a negative-result check, and are the named buyers real and well-funded. The trust and disagreement scoring, supply-chain and concentration-risk layers, and freedom-to-operate screen across the materials patent corpus are added where the thesis warrants them — feedstock concentration and clear IP paths are the natural extensions when a ProtoCo sits in critical minerals, refining, or separations. Scoped engagements of this type are estimated in the range of roughly $30,000 to $60,000 per pass; all figures are estimates for framing and would be confirmed in a scoping conversation. For a foundry evaluating multiple materials theses and inbound decks across a year, the more efficient structure is a standing diligence retainer that bundles a defined number of named-thesis or named-target reads with metered API access to the opportunity-intelligence and negatives-atlas products for Flagship's own scientists and partners. A sensible sequence is a single pilot on one live thesis to calibrate Lattice Graph's read against Flagship's own judgment, then conversion to an annual arrangement with modular add-on access to the supply-intelligence, cross-engine trust, freedom-to-operate, and knowledge-graph layers as specific theses demand them. There is no exclusivity or IP transfer implied — this is an independent technical-diligence service and metered data access, with the commercial structure set in a scoping conversation rather than a fixed price list.