Jev Cold Start Prior — Testing skill for Claude Code

jev-cold-start-prior

**Can a TypeSafe Jev prior, read from a README on day one, tell which new agent-skill repos will gain stars — and how many labels does a trained model need to catch up?**

Two parts:

part	status	what it answers
Retrospective proxy (this repo, results below)	done, n = 1,132	Directional: does the day-one prior rank repos by later popularity, vs day-one metadata and vs a model trained on the same README text?
Prospective test (running in Agent Skills Hub's sync)	first outcomes 2026-10-02	The clean version: prior vs star count at entry, 14-day outcome, no look-ahead

Why a proxy

The obvious experiment — rebuild each repo's star count on the day we first saw it and 14 days later — can no longer be done after the fact:

• GitHub's GET /repos/{owner}/{repo}/stargazers now returns 404 for every repo, including octocat/Hello-World (authenticated, 2026-09-18). GraphQL returns stargazerCount but an empty stargazers list. `src/jcp/stars.py` is the route we built for it.
• Star-history services rebuild from GitHub's public event firehose; OSSInsight marks its own data severely degraded since 2026-05-01 (a repo with 157 stars shows 1).

So the retrospective target is **stars today, adjusted for how long each repo has existed**, and the strongest real baseline — the star count at entry — is missing. That baseline is exactly what the prospective part records.

Results (retrospective proxy)

**Sample**: every repo the Agent Skills Hub catalog first saw 16–75 days before 2026-09-18 that was created at most 14 days before being seen (genuinely new, not a backfill), has a README over 300 characters and fewer than 3,000 stars — **1,132 repos**, median 17.5 stars.

**Target**: `log1p(stars today)` residualised on `log(days since creation)`.

**Jev prior**: one `/alpha/decisions` call per repo (`typesafe/jev-1.13-20260917`, $0.085 total)