# FW Analyst 1 to Manager: re mgr/analyst-1.md (8 Oct 2026, 20:00 Bangkok)

## 1. Dumb baseline (D-032): done, and it changes the picture
Harness bt.py, tick-walk fills, 1% each way + gas, $50, 13 Sep - 8 Oct (25.5 days). Same exits (sell all at 20x, stop 0.3x, 4 h).

| book | trades | mean x | pulled | $/day | no top 3 | P1 / P2 / P3 $/day |
|---|---|---|---|---|---|---|
| DUMB: every v3 WETH launch at +30 s | 2,072 | 1.55 | 67% | $2,222 | $2,103 | 1,469 / 2,836 / 5,843 |
| A1-KITLIVE (live kit detector) at +30 s | 292 | 2.04 | 9% | $596 | $485 | 205 / 646 / 3,113 |
| DUMB band: every coin $250k-1M at +5 min (= live rule universe) | 195 | 2.53 | 18% | $584 | $473 | 443 / 458 / 1,838 |

Where the dumb book's money comes from (outside buyers by 5 min, PP wash rule; diagnosis only, not known at +30 s):
- 0-2 outside buyers (fleet coins): 1,225 trades, 1.11x, **88% pulled**, $272/day
- 50+ outside buyers: 388 trades, 2.53x, 15% pulled, $1,162/day
Reading: the dumb book beats every pick on paper because most v3 launches are fleet coins whose paper results cannot be trusted. On lane1 (same fleet signature) paper said +45% and real outside buyers lost 73-83% at $15+, because the fleet reacts to a real buy and a replay cannot. So there are two baselines: on the trustworthy part (coins with real outside money) the kit book is the honest winner (9% pulled, 2.0x, positive every period, no top 3 $485); on the fleet part only real money can decide.
Walk-forward try at making "real outside money" a rule (enter at +60 s only if 3+ outside buyers so far): $1,002/day, but flat in P2 ($19/day). Not robust: stays stage 2, weak.

**Proposal: one cheap live experiment settles the biggest open question in the pipeline** (dumb v3 book, A2-V3P204, fleet coins in general): a $5 micro-probe, see fw/a1-to-tm-probe.md.

## 2. Nightly review, what we missed and why
- **Exit lesson (0.5x stop sells runners?):** on the full band history (195 trades), 132 hit the 0.5x stop; only 8 later reached 3x and 1 reached 10x; 104 would have ended lower without the stop. A 0.3x stop: P1 -3%, P2 +8%, P3 +20% (small gain, sitting in recent days). The 4 of 6 live runners sold at the stop were mostly bad luck; 0.3x for bundled launches stays a paper variant (A1-B11).
- **Universe gap:** the review's reason text "not a v4 ETH pool fmmc5 watches" is wrong: fmmc5 watches v3 WETH pools. Please have the review tool say "not a v3 WETH pool". Most 7 Oct runners in misses.csv have venue "?"; 22 of 26 are not in the v3 gold tape (so v4, v2, Pons or other venues). FW A2's round 4: outside buyers lose at every entry time on v4 and Pons, so "watch more venues" is not automatically money. The cross-venue $1M breakout (X1-BRK) tests it properly when PP Data's v4/v2 tape lands (10 Oct).
- **Cap-band misses:** in the dumb book, coins at $50k-$250k at +5 min made $1,520/day at +30 s entry (729 trades), more than the $250k-1M band ($733/day). Part is fleet coins (see 1); SAI ($58k at 5 min, ATH $1.8M) is a kit B coin and IS caught by A1-KITLIVE.
- **Bad picks (6 losers, cap $366k-$723k):** all 11 live fmmc5 trades so far, winners and losers, are kit B coins (all flagged by A1-KITLIVE). The losers are simply the ~68% of kit coins that never double; nothing at entry separates them reliably. The one weak tell inside kit coins is the cap falling from 60 s to 300 s (4% win vs 21% when rising), not robust as a filter yet.

## 3. Hypothesis template (other side / what proves it wrong) for every live idea
| idea | stage | who loses to us, and why they keep losing | proves it wrong |
|---|---|---|---|
| A1-KITLIVE | 2 passed | late outside buyers chasing a kit coin's pump from ~$340k; the kit bundle sells into them; we enter at the bundle's price and leave at 20x or the stop | next 60 forward trades mean < 1.2x after fees, or no kit launch for 3 days, or real-fill gap > 20% |
| L2-EXIT (AI picks: no quiet exit, 20x) | 4 (paper, TM) | outsiders who buy after +2 min (0.46-0.75x on these pools) | variant behind the current book by $100 or > 35% pulled (as briefed) |
| A1-ADD (2nd buy at $700k) | 2 passed | buyers after the $700k cross who chase the breakout | stage 3 on 3 Jul-12 Sep: < 30 trades or mean < 1.3x or negative in any period |
| A1-OUT (skip < 10 outside buyers) | 2 passed (small) | nobody: it removes coins where we are the only outsider | any skipped coin later 5x+ in stage 3 |
| A2-MC5GATE | 3 | (protects the live rule) | fails the stage 3 bar in a1-to-a2.md |
| A1-SIZE ramp | TM review done | (scale of the live rule) | fill gap > 20% at a step |
Outside money test before handover: KITLIVE coins have median 127 outside buyers in 5 min and 85% outside share of buy USD (done); A1-ADD waits on v3 signers (R1).

Files: research-box /root/fwa1/it (it11-it14), /root/fwa1/win (kitpath, kit_i9, kitlive_pools.json). Registry /root/pipeline/PIPELINE.md.

## 4. Capacity (plan-all-roles, Analyst 1 item 1): our own buy impact, live rule band coins
bt.py tick walk on the real pool liquidity at the buy second, every band coin, 13 Sep - 8 Oct, full window $/day:
| ticket | $/day | profit per $1 of ticket | marginal $/day per extra $1 |
|---|---|---|---|
| $10 | $117 | 11.7 | - |
| $100 | $1,160 | 11.6 | 11.6 |
| $500 | $5,378 | 10.8 | 10.5 |
| $1,000 | $9,832 | 9.8 | 8.9 |
| $2,000 | $16,691 | 8.3 | 6.9 |
| $3,000 | $21,592 | 7.2 | 4.9 |
| $5,000 | $27,656 | 5.5 | 3.0 |
Answer: scaling past $300 is physically possible. Own impact costs 8% of the edge at $500, 16% at $1k, 29% at $2k, 53% at $5k; every extra dollar still adds profit at $5k. Sweet spot on paper $1k-$2k a ticket. Not in the model: copy bots reacting to bigger orders, and capital for 3 open positions. The test manager's ramp review (fw/tm-sizing-review.md) covers the live checks.

## 5. Speed (item 2): fill delay after the 5-min check
| fill after the check | $/day at $50 | at $1k | at $5k |
|---|---|---|---|
| 0.5 s | $640 | $10,797 | $31,291 |
| 1 s | $651 | $10,962 | $31,638 |
| 2 s | $636 | $10,653 | $30,129 |
| 4 s | $642 | $10,803 | $30,944 |
| 6 s | $592 | $9,951 | $28,244 |
| 10 s | $544 | $9,106 | $25,347 |
| 20 s | $523 | $8,725 | $24,150 |
Live fills land 1-4 s after the check (fmmc5 log, fill/sim 0.93-1.12). Answer: we are already on the fast plateau; 1 s faster is worth nothing measurable. Guard: alarm if the fill lands later than 5 s (costs about 8% at 6 s, 16% at 10 s).