# #1461 v3 births to AI scorers (ppfeeds, 11 Oct)
- Feed already exists: fw/a1feed/early.jsonl (+60 s and +300 s checks, from v3live.db) and cross.jsonl (300k cross). Fields: creator, distinct_buyers, buy_usd, sell_usd, mcap_usd, pool_weth, fee_tier, launch-block buyers. Lag p50 ~2-14 s. Fee tiers seen there (since 8 Oct): 100 x1439, 10000 x72, 3000 x22, 500 x20 (a1feed applies a filter, so far fewer than the 9100 pools in v3live.db).
- Raw: /opt/sw/v3live.db (research-box) tables pools/swaps/liq, current to <1 min, all v3 and launcher pools incl. VORTA/hoodcat/PAWS.
- NOT done: wiring them into aipick/aipaper. aipick feeds the paper books and scores with a v4 model (hook, pad, fee features) that does not transfer to v3, so a score would be meaningless without a v3 model. Needs: testmgr shadow book (own DB) + model from analysts (train on v3hist.db + v3live.db). Existing precedent: v3mc5 paper book (cap $250k-1M at +5 min).
- Open question for analyst1: should a1feed's filter be loosened so every v3 pool gets a +60/+300 row (needed for training and a shadow scorer)?