Methodology

How LLM Adoption Rating is computed from Git activity

LLM Adoption Rating (0–100)

The score is a single headline number that summarizes whether a profile shows a sustained, step-change in activity patterns near the user-reported “90% LLM point” (dominant_at).

This is observational. We do not infer LLM usage from code, and we make no quality or causality claims.

Inputs

We use the declared dominant_at plus weekly aggregates derived from Git history:

•commits and changed (insertions + deletions) per week
•Optional: per-week technology breakdown used for “breadth” (entropy). If missing, the score is still computed without the breadth term.

Active-Week Windows (Holiday Safe)

Calendar windows break when someone takes time off. Instead, we build windows in terms of active weeks (weeks with commits above a per-profile threshold).

K = 8 active weeks before dominant_at (“pre”)

K = 8 active weeks on/after it (“post”)

K2 = 8 more active weeks after that (“post2”)

Subscores

We combine several subscores (0..1). Core signals use log transforms (log1p) and robust statistics (median + MAD).

Strength: effect size of the pre → post change (commits + changed, and entropy when available)
Alignment: how close the strongest local “jump” is to dominant_at, down-weighted when the best jump is near-zero
Persistence: how often weeks in post2 remain uplifted vs the pre baseline
Breadth (optional): whether technology entropy increases
Confidence: reduces scores when evidence is thin (low volume or a large calendar span to collect the active weeks)

Final Score

We compute a weighted sum and then apply confidence:

raw = 0.45*Strength + 0.20*Alignment + 0.25*Persistence + 0.10*Breadth
rating = round(100 * Confidence * raw)

If breadth inputs are unavailable, the breadth term is dropped and the remaining weights are renormalized.

Accuracy Disclaimer

All metrics are observational and based on user-declared inflection points. The service makes no claims of causality or performance guarantees. Correlation does not imply causation. Many factors affect developer productivity beyond LLM usage.