Lattice Graph × Lila Sciences

Lattice Graph operates a computational materials-discovery platform built around a knowledge graph that spans millions of compositions, linking structures, properties, synthesis routes, patent claims, and experimental outcomes into a single governed evidence graph. The core differentiator is not just breadth but depth of evidence: every node and edge carries provenance, so any property value can be traced back to the source calculation, experiment, or literature entry that produced it. For a company like Lila Sciences, whose foundation models need to learn not just what materials do but why certain predictions should be trusted and which experiments have already been run, that provenance chain is the layer that turns a static dataset into a queryable source of truth. Validation on the platform runs through multiple independent physics engines in parallel. Machine-learning interatomic potentials including MACE and CHGNet, combined with density-functional-theory calculations, produce phonon spectra, thermodynamic stability envelopes, and formation-energy predictions that are cross-checked against each other before any signal reaches downstream consumers. The disagreement between these engines is itself a signal: where MACE and DFT converge, confidence is high; where they diverge, a flag surfaces. That cross-engine adjudication layer gives science-foundation models something they rarely have access to in standard computed datasets — a calibrated uncertainty signal grounded in physics, not just statistical variance. Perhaps the most unusual asset on the platform is the labeled negative corpus: more than 23,000 failed-experiment and kill edges representing compositions, dopants, annealing conditions, and interfaces that were tried and did not work. Public materials databases and the scientific literature are overwhelmingly records of success; failures are suppressed by publication bias and largely absent from any training corpus built by re-scraping the open web. Lattice Graph captures these negatives through internal experimental pipelines, labels them with the same provenance discipline applied to positive results, and makes them available through a governed API — giving training and evaluation pipelines access to the side of materials reality that most models have never seen.

Lila Sciences is building toward scientific superintelligence: science-foundation models coupled to autonomous, self-driving experimentation loops intended to close the hypothesis-test-refine cycle far faster than human-paced R&D. That ambition is compelling, but it rests on a structural vulnerability in how publicly available materials data is distributed. The positive-result bias in the scientific literature means that any foundation model trained predominantly on public sources will develop an optimistic prior — it learns what works and is systematically underexposed to what does not. For a platform whose value proposition is the throughput and reliability of an autonomous experimental loop, that optimistic prior is not a minor calibration issue; it is a direct driver of wasted bench cycles and eroded credibility every time the loop proposes a known dead end. Lattice Graph's strategic fit with Lila is not as a competitor or a parallel discovery engine but as the missing data infrastructure layer. The three products that map most directly to Lila's needs address three distinct gaps: the absence of labeled failure data in training and evaluation sets, the absence of a principled uncertainty signal for ranking predictions before committing autonomous-lab time, and the absence of grounded provenance for the computed values Lila's models ingest and generate. These are precisely the gaps that Lila cannot close by scraping more literature or running more DFT, because the failures were never published and the uncertainty was never quantified in the sources that already exist. The commercial relationship Lattice Graph envisions for Lila is a data and evaluation license, not an asset sale. Lila keeps its models, its lab infrastructure, and its IP; it licenses the missing training and evaluation signal — the negatives, the disagreement flags, the evidence graph — and uses that signal to harden its own benchmarks and improve the fidelity of its own experiment-prioritization logic. That structure is straightforward to scope, price, and expand incrementally as the engagement matures and Lila identifies additional data products that move its metrics.

Lila Sciences is building toward scientific superintelligence: science-foundation models coupled to autonomous, self-driving experimentation loops intended to close the hypothesis-test-refine cycle far faster than human-paced R&D. That ambition is compelling, but it rests on a structural vulnerability in how publicly available materials data is distributed. The positive-result bias in the scientific literature means that any foundation model trained predominantly on public sources will develop an optimistic prior — it learns what works and is systematically underexposed to what does not. For a platform whose value proposition is the throughput and reliability of an autonomous experimental loop, that optimistic prior is not a minor calibration issue; it is a direct driver of wasted bench cycles and eroded credibility every time the loop proposes a known dead end. Lattice Graph's strategic fit with Lila is not as a competitor or a parallel discovery engine but as the missing data infrastructure layer. The three products that map most directly to Lila's needs address three distinct gaps: the absence of labeled failure data in training and evaluation sets, the absence of a principled uncertainty signal for ranking predictions before committing autonomous-lab time, and the absence of grounded provenance for the computed values Lila's models ingest and generate. These are precisely the gaps that Lila cannot close by scraping more literature or running more DFT, because the failures were never published and the uncertainty was never quantified in the sources that already exist. The commercial relationship Lattice Graph envisions for Lila is a data and evaluation license, not an asset sale. Lila keeps its models, its lab infrastructure, and its IP; it licenses the missing training and evaluation signal — the negatives, the disagreement flags, the evidence graph — and uses that signal to harden its own benchmarks and improve the fidelity of its own experiment-prioritization logic. That structure is straightforward to scope, price, and expand incrementally as the engagement matures and Lila identifies additional data products that move its metrics.

For a data-and-evaluation engagement, the most useful interfaces are the knowledge-graph explorer and the composition-intelligence reporting layer, which serve as the human-readable window onto the same signals Lila's team consumes through the API. Lila's ML engineers and research scientists would use composition-360 views and evidence-neighborhood traversal to audit any sampled material candidate — inspecting its stability evidence, the specific negatives attached to it, and the cross-engine disagreement profile on its computed properties. This makes eval failures traceable: instead of a black-box benchmark score, the team sees the specific provenance node or kill edge that drove a discrepancy, which is exactly the kind of interpretable feedback needed to iterate on training data curation and benchmark design. The batch-screening workflow and the formation-energy predictor complement the explorer for day-to-day use. Lila's team can run candidate families through batch screening and apply the negatives check and trust signals as a programmatic filter before routing candidates to autonomous-lab queues, while the explorer makes the kill and negatives data human-inspectable so the team can validate coverage, assess label quality, and sample the distribution of failure modes during an evaluation period. These are grounding and auditing tools positioned to harden Lila's own model and loop, not to replace the discovery logic Lila is building.

The natural entry point for Lila is a scoped negatives audit: a bounded, time-limited engagement in which Lila evaluates the Negatives and Eval-Data Atlas and the Trust and Disagreement Signals against its own model outputs and benchmarks, with the Knowledge Graph API providing provenance context throughout. The goal is concrete and measurable — quantify how much the labeled-failure and calibrated-uncertainty data move Lila's eval metrics before committing to a broader license. Deliverables from the audit include a coverage report mapping the negatives corpus against Lila's current training and evaluation sets, a disagreement-signal integration study showing where cross-engine flags would have changed Lila's experiment-queue rankings, and a provenance-sampling exercise anchoring a representative slice of Lila's computed inputs to their underlying evidence. That audit is scoped in the range of forty to seventy-five thousand dollars as a starting engagement. Following a successful audit, the relationship converts to a recurring data and evaluation license covering the full suite of products, priced against seat count, query volume, and refresh cadence. The license is non-exclusive and refreshes as the negatives corpus and trust signals grow. If Lila later wants to pull in additional capabilities — freedom-to-operate and patent whitespace screening, synthesis-route intelligence, or the opportunity-identification engine — those are scoped as incremental add-ons to the base license rather than requiring a new negotiation from scratch. No IP transfer is implied at any stage; Lila licenses the missing signal, keeps its models and lab infrastructure, and retains full ownership of what it builds on top.