# Learn report 10 Oct 2026 (research, first day)

Ledger: 44 rows backfilled from both idea files, paper registry (15 tests) and audits. Page: http://159.223.88.168/hx-c9ois/fw/ledger.html (families: `ledger stats`).
Not itemised yet: analyst2 runs 15-27 and most a1-mc5-* files (parked row points to them). Next days fill these.

## What we learned
- Dead ends (18 killed, 10 in "dead-universes"): calendar/regime/hot-hand/vanity/name cuts are flat; bonding-curve and stock-arb have no edge after friction; unhooked v4 fee100 baseline is 0.98 once pulls are counted. Do not retest these.
- Repeated mistake: three "edges" were artifacts, each found later by a pull/depth audit (dyn-fee units, ignoring LP pulls: 2.68 -> 2.25, one-print jumps, dust ticks). Rule for new ideas: pull-aware, depth-gated sim from the start.
- Every alive family sits on small n and has not passed an audit: h3k, h100 "FIX", others "nothing to audit yet". Zero families are verified except the 5-min cap rule (rests on ~15 runners).
- Miss review: biggest missed runners are USDG/SUSD v4 (baseline 1.35x, no separating rule) and the H3000 family (already in paper).

## Families
- Gaining: hooked-v4 / tp-ladder and short-tp (replay strong, forward n tiny, sellability for outsiders unproven).
- Losing: everything in dead-universes, unhooked baseline.
- Unproven but alive: unhooked-v4 first-minute filters (4 arms, all inside selection noise per A2's own null test).

## Top 3 directions today
1. Forward evidence for hooked-v4: get h3k/h100 to 30+ closed trades with pulls visible; testmgr + auditor re-audit. Highest value, nothing new to invent.
2. Sellability of outsiders on H3000 (the only thing that can void the best family): Ben's $5 sell test is still open; analysts: find any on-chain evidence of outsider sells at >=5x.
3. USDG/SUSD v4: find one pre-entry rule that beats 1.35 baseline on both halves, pull-aware; otherwise kill.

## Addendum 09:30 BKK (sweep 2)
- Calibration (auditor): mc5 harness fill model exact at $10 (buys 12/12, stops -0.03x, winners no bias); the gap is execution failures (429, GMGN no-hash), not fills. Hooked families have no live fills, so uncalibrated.
- New rows #45 hook slot1 filter (replay, forward test next), #46 iter 20-21 discards, #47 fml2 skip_quote (parked; 16 picks lost vs 14 traded; 6 runners $2.4-7M, aipaper paper +870% each).
- Direction 1 stays; add: slot1 forward test is the cheapest new evidence for hooked-v4.