adjacent-atlas

Scoring model

Every node receives an adjacency score in 0–100, computed as a weighted mean of five components, each normalised to 0–1.

adjacency = 100 × ( w_m·momentum + w_r·recency + w_c·connectivity
                    + w_n·novelty + w_f·feasibility )

Weights are normalised to sum to 1. Defaults (DEFAULT_WEIGHTS):

Component	Weight
momentum	30%
recency	20%
connectivity	20%
novelty	15%
feasibility	15%

Components

Momentum — the windowed sum of recent activity divided by the previous window, smoothed by +1 and passed through a logistic of the log-ratio. Flat activity reads 0.5; sustained growth approaches 1; decline approaches 0.

Recency — 0.5 ^ (ageDays / halfLifeDays), with a 180-day half-life by default. Activity today reads ~1; one half-life old reads 0.5.

Connectivity — a node’s total incident edge weight divided by the largest weighted degree in the graph. Hubs approach 1; isolated nodes read 0.

Novelty — 0.5 ^ (maturityYears / halfLifeYears), with a 4-year half-life. Newer entities score higher.

Feasibility — 1 − e^(−implementations / saturation), saturating at 5 implementations by default. Zero implementations read 0.

Properties

The model is transparent (every component is inspectable on the atlas and in each brief), deterministic (a snapshot pinned to its generatedAt reproduces exactly — scripts/validate-snapshot.ts checks this), and tunable (weights are passed through ScoreOptions).

It is a heuristic, not a predictor. It encodes a defensible point of view about what is worth looking at next; changing the weights changes that view.

Limitations

The score is a useful lens and a blunt instrument. In short:

• It measures attention, not merit — a high score is a prompt to look closer, never a verdict.
• The inputs are coarse public-signal proxies (repository and keyword-search counts), not curated bibliometrics, and must not be read as citation counts.
• The weights are a point of view; change them and the ranking changes.
• The seed graph is small and authored, so connectivity is sensitive to a few hand-set edges.
• There is no causal or predictive claim — opportunity means young, moving, and buildable by the proxies above, nothing more.

A fuller version is in field-notes/scoring-limitations.md.

This site is open source. Improve this page.