# Test manager review: A1-SIZE ramp for live mc5 (8 Oct 2026, 18:30 Bangkok)

Reply to fw/a1-to-tm-sizing.md. **Verdict: ramp is sound in principle, but not ready to start yet.** Three things block it. All three can be settled today. The decision on the stop rule (blocker 1) is Ben's.

## What I checked
- **Live record (mc5.db, 12 closed trades):**
  - 2 winners: 29.4x and 19.5x.
  - 10 losers: median 0.48x. One was a 0.00x stop wipe, one 0.22x.
  - Net: +40.9 tickets. Better than the model's +1.9 tickets per trade, but only two winners carry it.
  - The last 4 trades are losses (all 3 at $50 lost).
- **Live fills:**
  - Buy fill vs sim stays within -12% to +7.5% on all 12 trades.
  - Stop slippage is usually -5% to -17%. The two outliers are the 0.00x rug and the 0.22x trade.
  - This matches A2's finding that fills hold at today's size.
- **Model trades (it7/o/out50.csv):**
  - 154 trades, 23% win rate, mean 2.87x, median 0.48x, longest real losing run 13.
  - Half B (30 Sep-6 Oct) is weak: +31 tickets in 72 trades.
  - A2-MC5GATE shows the same rule lost money from May to August. The edge is the September regime.

## Blocker 1: the 10-loss stop will very likely fire during the ramp, even if the model is right
- A rule with a 23% win rate has a 68% chance of 10 losses in a row within 60 trades. This is from a bootstrap of the model's own trades, 20,000 runs.
- Ramp $100 x20 -> $250 x20 -> $500 x20 under today's stops:
  - Only 30% of runs finish.
  - Median +$7.7k, 5th percentile -$636.
- Same ramp with a 20-ticket drawdown stop instead:
  - 99% of runs finish.
  - Median +$29.9k, 5th percentile +$6.7k.
- But those runs assume the September regime holds. The streak stop is a crude brake for when it ends.
- **Ben decides one of these:**
  - (a) Keep the 10-loss stop as it is. A stop means review and re-arm, not "the edge is dead".
  - (b) For this rule only, replace it with A2-MC5GATE plus a ticket drawdown stop. A2-MC5GATE (buy only while the last 20 closed trades average at least 1.0x) is at stage 3; it needs to be ready first.

## Blocker 2: the fill-gap gate has no live data
- The gate reads /opt/exec/exitgap.json. That file was last written 7 Oct 08:48 and nothing updates it, so it misses 9 of the 12 live trades.
- **Proposed fix:** the test manager builds the gate from the bot's own per-trade fields in mc5.db: fill_vs_sim, buy_gap, sell_gap and stop_slip. It goes on the live tab, with step up / step down shown automatically. The move-up/down rule stays as briefed (gap within 10%, back down above 20%).

## Blocker 3: settings and money for larger tickets
- **Wallet:** with max_open 3, a $500 ticket needs about $1,500 in flight plus gas. Ben funds the wallet. Only Ben changes `size_usd`.
- **Slippage cap:** buy_slip is 5%. Pools hold 10-15 WETH. A $500 buy (about 0.2 WETH) plus copy bots could go past 5% and revert, which means missed winners.
  - At the $250 step, watch for reverted or skipped buys. Do not widen the cap blindly.
- **Drawdown stop:** kill_dd_tickets is 10, so it scales by itself: $1,000 at $100, $5,000 at $500. Ben should confirm he is comfortable with those dollar amounts.

## Proposed order
1. Ben picks (a) or (b) for blocker 1.
2. The test manager ships the live fill-gap gate (blocker 2).
3. Ben funds the wallet and sets $100. 20 trades at that size is about 3-4 days at the live pace.
4. Steps after that go as briefed, with the gate shown on the dashboard.