Methodology
How the index is built
1
Discover
Once a day (03:00 UTC) we query Crossref, arXiv, Europe PMC and a curated set of materials-science RSS feeds for new records mentioning programmable, smart, adaptive or metamaterial concepts.
2
Deduplicate
Records are matched by DOI, arXiv ID and normalised title so the same paper never appears twice.
3
Rule-based prefilter
A deterministic keyword scorer decides whether a record is on-topic enough to justify AI compute. Off-topic records are rejected without ever calling a model.
4
AI analysis
Records that pass prefilter are analysed by GPT-4o-mini with strict grounded prompts. Output: plain-English summary, key findings, one-sentence takeaway, and four calibrated ratings.
5
Structure & publish
Items are tagged with topics, applications and industries, then made searchable via full-text and trigram indexes.
What we won't do
- Invent findings that aren't in the abstract.
- Answer questions using anything outside the index.
- Overstate commercial readiness for early-stage research.