winning-wallet-finder/HANDOFF.md

Session handoff — 2026-07-08

Self-contained pickup for a fresh session. Read README.md gotchas 1–11, FINDINGS.md, and LIVE_ROLLOUT.md first. Project memory (polymarket-resolution-truth, polymarket-us-expansion) has deeper history; the repo is authoritative.

Where things stand (all green)

• Live paper bot: Fly.io app wwf-copybot, region arn (Stockholm), push mode. Following Set D (2026-07-08): LSB1, imwalkinghere, Kruto2027, 42021, 0xbadaf319, BikesAreTheBikes — 6 moderate-bet volume wallets chosen by simulation ($27.8k / +2680% 30d backtest). flyctl logs --app wwf-copybot.
• P&L is honest end-to-end. The sharps table's All-Time P&L is now the realized track record (sum of Polymarket realizedPnl per closed position) PLUS abandoned-loser fold-in (_open_split in validate_timing) — matches PM where PM is honest, reads lower where PM hides abandoned losers (oliman2 $181k →$20k). 27/31 sharps match PM within a few %, 4 honestly-below, 0 over-count. New Open P&L column = genuine in-flight exposure.
• Ledger reconciles (ledger_drift 0.00, self-checked every heartbeat).
• Daily pipeline (launchd 08:00) runs the corrected code; exits cache is backfilled (86/96 wallets complete, incremental from here). Verified wired.
• Data caches: live/cache.duckdb — bets (still windowed/capped, used for z-score selection only), exits (FULL history, incremental, realized_pnl), resolutions (chain payouts). closed_exits is the shared exit/realized source for both the backtest and the sharps table.

THE PROPER FIX (do before real money) — the orphan / cash-debit gap

Symptom seen 2026-07-08: after the Set D deploy the ledger self-check flagged ⚠ LEDGER DRIFT $+36.35. Root: a position ("Hanyi Liu vs Aoran Wang: Total Sets", ArbTrader — a now-dropped wallet) sat in state["my_pos"] but its cost was never debited from state["cash"] and it had no bet record and no conds entry. A second market on the same event, sold, was left status:open (a "ghost"). Both were cleaned by hand (copybot_state.json surgery, committed); drift is 0 now. But the mechanism that creates orphans is guarded, not fixed.

Why it happens (as far as traced): on_wallet_activity books cash + creates the bet record via:

for f in self._drain_fills():          # _drain_fills debits cash for EVERY fill
    if f["token"] == tok:              # but _record_lag (the bet record) fires
        self._record_lag(wallet, t, f) # ONLY for the fill matching THIS trade's token

When two markets of one event are processed in a batch (or a fill is left in ex.fills across iterations, or an exception fires between handle_trade and _drain_fills), a buy fill can end up drained-but-not-recorded (no bet record), or added-to-my_pos-but-never-drained (cost not debited). The _record_lag if f["token"] == tok filter is the fragile seam.

Guards already in place (make recurrence VISIBLE, not silent):

• Copybot.ledger_drift() — checked every heartbeat, logged ⚠ LEDGER DRIFT, published in the feed (ledger_drift field). Catches the cash gap in ≤60s.
• write_feed synthesizes a bet record for any open my_pos position missing one (so open_count always matches the visible table).
• my_pos entries now carry wallet (copytrade.py) for attribution.

The proper fix — two options, recommend A for Phase 1:

Option A (surgical, lower risk): in on_wallet_activity, record EVERY drained buy fill, not just the tok-matching one. The tok-match gets _record_lag (real lag/slippage); others get a bet record synthesized from the fill (lag unknown). AND assert the invariant after each trade: every my_pos token has (cost in cash via a drained fill) + (bet record) + (conds entry); if not, log and self-correct. This closes the "drained-but-not-recorded" and "added-but-not-drained" seams without touching the cash model.

Option B (architectural, eliminates the class): stop accumulating state["cash"] per fill; derive it each cycle from the book: cash = bankroll + Σ(settled bet pnl) − Σ(open my_pos cost+fee) + adjustments. Then ledger_drift is 0 by construction and an orphan simply (correctly) reduces available cash. Higher effort — the derivation must reproduce every flow the accumulation handles (entry fee, sold-leg proceeds, refund/settlement payoffs, redeems). Worth it long-term; risky to rush before real money.

To reproduce/investigate: add a log line in _handle_their_buy on every my_pos add with len(executor.fills) before/after, and in _drain_fills log each fill booked. Watch a batch where one wallet trades ≥2 markets of the same event in <10 min (ITF tennis sub-markets, in-play soccer). The orphan appears when the second market's fill misses _record_lag.

Phase 1 remaining (LIVE_ROLLOUT.md) — 6 of 8 done

Done + verified: 1.1 feed separation, 1.2 env secrets, 1.4 on-chain cash anchor (chain_cash_gap), 1.5 fatal geocheck (refuses from a blocked box), 1.6 exit retries (⚠ EXIT STUCK), 1.7 config sync (config.live.example.json).

1.3 — live deps in the worker image (NOT done). fly.Dockerfile is stdlib-only. Add pip install py-clob-client web3 (pin versions) — needed by LedgerLiveExecutor + redeem.py. Keep the paper path import-lazy (already is). Verify: flyctl deploy --remote-only, then flyctl ssh console -C "python3 -c 'import py_clob_client, web3'" exits 0, and the paper bot still boots clean (banner + heartbeat, drift 0.00).

1.8 — dashboard live-money section (NOT done). trading/index.html (in the jaxperro repo) should read live/copybot_live_real.json (the live feed path from 1.1, set in config.live.example.json) into a fourth board chip + tables, clearly labeled REAL MONEY, reusing the standardized columns (Their/Our Entry, Stake, Result W/L/R/S, Placed/Settled) and the ledger_drift/adjustments fields. Verify against a copy of the paper feed with "mode":"live"; then confirm graceful empty-state when the file 404s (test not started).

After Phase 1: Phase 2 is funding (USER ONLY), then preflight → supervised first fill → the edge-case matrix (LIVE_ROLLOUT.md Phase 5).

Standing to-dos (user-only or external)

• ROTATE THE LEAKED ALCHEMY KEY. A real key briefly hit the public repo (config.live.example.json, removed from HEAD but in git history). Generate a new key at dashboard.alchemy.com, then update alchemy_key in the local gitignored config.json AND the ALCHEMY_RPC_URL Fly secret (flyctl secrets set ALCHEMY_RPC_URL=...). Blocks nothing but is a live key.
• Delete the stopped Railway project magnificent-kindness when comfortable (saves the idle spend; the Fly worker fully replaced it).

Operational quick-reference

• Change the follow set: edit live/copybot.paper.json wallets, then ./live/deploy_bot.sh (validates → pins floors → syncs Alchemy webhook → restarts Fly → confirms banner). backtest.json should mirror it.
• Backtest any set: python3 live/portfolio.py --wallets 0xa,0xb --days 30 --out /tmp/t.json.
• Never keep 2 Fly machines (2 book-writers): flyctl scale count 1.
• Cache is single-writer; don't run cache-touching scripts while the daily pipeline or a regen is running.
• Bot commits its own state/feed every few minutes — always git pull --rebase --autostash before pushing.