The Hypothesis
Our #1 invention lead — a Na-ion patent-clear composition and route finder — survived four rounds of competitive analysis, market sizing, and expert critique. It was rated as plausible by web-search-based evaluation. Then we ran it against real data.
The Test
We ran the core anti-join against a 4.5 GB DuckDB warehouse with 168 tables, 15.1M material search rows, 306K patents, and 9.5K Ceder synthesis recipes. The query: find Na cathode compositions that have no patent coverage and at least one documented synthesis route.
Results: the anti-join produced false "patent-clear" results for compositions that are heavily patented.
Three Root Causes
| Root Cause | Impact | Example |
|---|---|---|
dim_canonical_composition is EMPTY (0 rows) | The canonical crosswalk was never built — no composition-to-composition identity resolution exists | NaFePO4 in one source cannot be matched to NaFePO4 in another |
| Patent formulas stored as un-reduced supercells | Exact string matching is structurally impossible — Na2Fe2(PO4)2 ≠ NaFePO4 | Known-patented cathodes like NaFePO4 falsely returned "patent-clear" |
12.4% of formula_reduced values are OQMD filesystem paths | Data quality defect pollutes all formula-based queries | /home/oqmd/data/... strings in the formula column |
The Binding Constraint Moved
Before the data-plane test, the binding constraint on C3 appeared to be market competition — are there funded incumbents doing this? After the test, the binding constraint moved to data-plane readiness — can the warehouse even answer the core query?
This is the critical insight: web-search-based competitive analysis rated 4 candidates as plausible that warehouse queries instantly killed. The data plane can complement market research and is worth querying early, once its inputs are checked.
What This Means for Invention Loops
- Ground-truth data contact moves to the front of the funnel. Before Round 5, data-plane validation happened after competitive analysis. After this finding, it happens before.
- An idea that passes market analysis but fails data-plane validation is dead. The market opportunity doesn't matter if the data can't answer the core question.
- Composition identity resolution is a prerequisite, not an optimization. Without a canonical crosswalk, every composition-based query produces structural false positives or false negatives.
- Data quality is not a background concern — it is the binding constraint. 12.4% filesystem-path contamination in formula columns means that 1 in 8 formula-based lookups is matching against garbage.