Calibration data loop

Making “viral potential” good means fitting weights to outcomes, not stuffing thousands of tweets into Jev’s small context window.

What to collect (per post)

Field	Why
text	Same input the analyzer sees
author_followers	Normalize engagement
posted_at	Freshness / cohort
likes, replies, reposts, quotes, bookmarks, views	Labels
media (bool / type)	Optional stratum

Derived label (example):

engagement_rate = (replies*3 + reposts*2 + quotes*2 + bookmarks*2 + likes) / max(followers, 1)

Prefer reply/share-weighted labels — closer to X ranking signals than likes alone.

Legal sources (allowed)

1. X API — your own posts, and whatever your access tier permits.
2. First-party outcomes — drafts you posted after using this tool (best feedback loop).
3. Public research archives / licensed datasets — only with clear redistribution rights.

Not allowed

• Bulk scraping of X HTML/mobile APIs in violation of ToS
• Purchased “tweet dumps” of unclear provenance
• Impersonation / credential stuffing

How Jev fits

1. Run the same question set in server/analyze.ts offline over the corpus.
2. Store per-dimension scores + composite under current WEIGHTS.
3. Fit new weights (regression / simple search) so composite correlates with engagement_rate within follower buckets.
4. Update server/weights.ts only — no prompt stuffing required.

Optional later: treat Jev probabilities as features for a small classical model (TypeSafe AutoResearch cookbook).

Similarity panel (phase 2+)

Once you have a corpus with embeddings or category + score fingerprints, surface nearest neighbors in the right-hand “Similar posts” stub. Retrieval stays in code; Jev stays for judgments.