About / Methodology

Scoring methodology

Lattice helps you triage university listings. Search finds candidates. Scores help you judge how unusual, corroborated, and actionable a listing looks. Missing evidence is labeled not scored — we do not invent a low number.

What you see

01 · Default

One novelty label

frontier / emerging / established — plus optional badges

02 · Details

How we scored this

Novelty (N), Impact quality (V), and Opportunity (O) sit in the dossier card

03 · Dossier

Expand Evidence

Papers, trials, patents, or the source page — with outbound links when we have a DOI, trial id, patent number, or listing URL

04 · Always

Open the source

Read the university page (and linked patents) before acting on a rank

Three dimensions

Each axis is independent. A listing can score on one and stay not scored on another.

Novelty (N)

How unusual is this vs the catalog?

We embed the listing text and look at how crowded its neighborhood is
Sparser neighborhood → higher N; crowded neighborhood → lower N
Labels are percentiles of scored listings, not scientific truth

Too little text, or no usable neighbors → N not scored (not the same as “established”)

Impact quality (V)

Patent lens only

Scored only if the catalog row already lists linked patents
Then we use citation patterns (forward/backward) and CPC rarity
Examiner-rejection history is optional — included only when we have it
Literature or trials elsewhere in the product do not fill V

No patent linkage → V not scored (not a fail, not a low score)

Opportunity (O)

How actionable does the listing look?

Uses license status, how recently we see the listing, and readiness cues when those fields exist

Missing inputs → O not scored, not a placeholder number

Percentile labels

Standing in our scored set — not absolute scientific truth.

Frontier

≈ top ~5%

Sparsest neighborhood among scored listings

Emerging

≈ next ~28%

Less crowded than most, not the extreme tail

Established

≈ remaining ~67%

Crowded neighborhood relative to the scored set

Not scored vs low score

Empty state

Not scored

Not enough inputs for that dimension. We leave it blank rather than invent a number.

Measured

Low score

We had inputs, and they look crowded, weakly corroborated, or thin on opportunity.

Rule of thumb: missing evidence ≠ 0

Search vs scoring

Retrieval (hybrid search ± query expansion / rerank) is separate from N/V/O
A hit can be highly relevant and still have thin novelty text
A hit can look novel and still have no V if patents are not linked on the record
Use both together

Advisory reading (optional)

A second, categorical opinion on the listing text (e.g. frontier / emerging / established)
Not mixed into the public numeric novelty rank
If the model declines (thin text), we show unavailable — not “established”

What we do not claim

Not investment advice; not a substitute for diligence
We do not fabricate patents, trials, or papers when linkage is missing
Percentiles = standing in our scored set, not absolute truth

Try Lattice

Questions on a record: dossier card first, then the university page. Product overview: About Lattice.