# Real-money rollout plan — executable by a fresh agent session > **STATUS 2026-07-10: executed through Phase 4** (bot ARMED, push mode, > first fills verified end-to-end — see [HANDOFF.md](HANDOFF.md) for current > state). **Rule 0.6's hard caps were RETIRED 2026-07-10 by user decision**: > sizing is paper-parity 4%-of-equity, floored at the venue $1 minimum. The > new-stack realities discovered during execution live in README gotchas > 16–17 (pUSD collateral / bridge conversion; in-play `delayed` acceptances > and the pending-order registry). This document stays as the historical > plan + the Phase 5 edge-case matrix, which is still worth running. Written 2026-07-07 after the pre-money audit. This document is **self-contained**: a session with no prior context (Sonnet/Opus) should be able to execute every phase from here + the repo. Read [README.md](README.md) gotchas 1–15 and [LIVE_TEST.md](LIVE_TEST.md) first. Project memory (`polymarket-resolution-truth`, `polymarket-us-expansion`) has deeper background; the repo is authoritative. **The goal:** a supervised minimum-size live test — deposit → copy a followed bet → mirror a sell before resolution → hold to resolution (win, loss, AND 50/50 refund) → withdraw — with worst-case loss bounded by the ~$30 exposure cap. Only after every edge case in Phase 5 passes does sizing-up get discussed. --- ## Phase 0 — invariants (read, don't skip) These were each earned the hard way. Violating any of them voids the test. | # | invariant | why | |---|-----------|-----| | 0.1 | **Live orders only from a geo-TRADABLE box.** `python3 host/geocheck.py` must print `VERDICT: TRADABLE`. The Fly worker (`wwf-copybot`, Stockholm) passes; a Mac in the US gets 403-geo-blocked on the order POST. | Polymarket IP-blocks order placement (README gotcha 10). | | 0.2 | **The user decides WHEN to arm** (jurisdictional posture: they trade live from Colombia months, paper from US months). The agent prepares everything; the arming step and its timing are user-only. | The machine's location doesn't relocate the person. | | 0.3 | **One writer per book.** The live book uses its own state (`copybot_state.live.json`) and must use its own feed file (Phase 1.1 — today both modes write `live/copybot_live.json`, which would clobber the July paper test's dashboard). The paper worker keeps running untouched throughout. | The paper July test is the control group; don't contaminate it. | | 0.4 | **Settlement truth is the chain.** CLOB `winner` flags never populate for operator-resolved in-play markets, and 50/50 refunds (28% of follow-set markets!) pay $0.50/share. `resolution_price()` in copybot.py already does CLOB→chain-payout tiering; live mode requires `ALCHEMY_RPC_URL` set or refunded/operator-resolved positions never settle. | 2026-07-06 audit: 4 positions sat locked for hours. | | 0.5 | **The ledger must reconcile to the cent.** `ledger_drift` in the feed must stay 0.00; in live mode it additionally must match the on-chain USDC balance (Phase 1.4). Any drift >1¢ = stop, diagnose, fix before the next order. | Real money makes bookkeeping bugs unrecoverable. | | 0.6 | **Hard caps stay at test size until Phase 6.** `config.live.json` risk block: $5/trade, $5/position, $25/day, $30 exposure, 6 open, prices 0.05–0.95, FAK orders. Never raise mid-test. | Worst case ≈ the $30 exposure cap. | | 0.7 | **Paper-mode safety interlocks stay intact**: live needs `mode:"live"` + `--live` + typed phrase. Never bypass, never automate the phrase. | The phrase is the human checkpoint. | --- ## Phase 1 — code prep (agent; all verifiable in paper mode, no money) Each item is a small PR-sized change with its own verification. Do them in order. ### 1.1 Feed separation `copybot.py` `FEED` is a module constant (`live/copybot_live.json`) — a live run overwrites the paper feed. Make it config-driven: `cfg.get("feed_path", "live/copybot_live.json")`, and set `"feed_path": "live/copybot_live_real.json"` in `config.live.json`. Same for `FILL_LOG` → `copybot_fills.live.jsonl`. **Verify:** run paper poll-once with a scratch config setting `feed_path`; confirm the alternate file is written and `live/copybot_live.json` is untouched. ### 1.2 Secrets via env for the worker path `config.live.json` is gitignored — the Fly worker clones the repo and won't have it. Extend `load_cfg`/`normalize_follow_config` so `LIVE_PRIVATE_KEY`, `LIVE_FUNDER_ADDRESS`, `LIVE_SIGNATURE_TYPE` env vars override `cfg["live"]`. **Verify:** unit-style: `LIVE_PRIVATE_KEY=0xdead python3 -c "…load_cfg…"` shows the override; no secret ever printed or committed. ### 1.3 Live deps in the worker image `fly.Dockerfile` is stdlib-only. Add `pip install py-clob-client web3` (pin versions) — needed for `LedgerLiveExecutor` + `redeem.py`. Keep the paper path import-lazy (it already is: LiveExecutor imports inside `__init__`). **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). ### 1.4 Live-mode ledger anchor: on-chain balance check In live mode, `summary()` should each heartbeat compare `state["cash"]` against the CLOB collateral balance (`client.get_balance_allowance`, cached ≥60s) and log `⚠ CASH≠CHAIN` when they diverge by more than pending-order float. This is the real-money version of `ledger_drift`. **Verify:** code-review level only until Phase 4 (needs a funded account); assert the check is skipped cleanly in paper mode. ### 1.5 Geocheck fatal in live mode `host/start.sh` warns on geo-block but continues (fine for paper). In `copybot.py` main(): if `want_live` and a fresh `geocheck` probe (import the 3-probe logic or shell out) is not TRADABLE → `sys.exit`. Never place a live order from a blocked box. **Verify:** from the (US) Mac, `--live` with a filled dummy key must refuse with the geo message before the confirmation phrase. ### 1.6 Exit-order resilience Mirror-sells use marketable FAK. On a thin in-play book a sell may fill 0 and the engine just logs ORDER FAILED — the position then silently rides to resolution. Add bounded retry: on a failed/partial SELL, retry next heartbeat (up to N=10 attempts, logged each time), then alert `⚠ EXIT STUCK` to Discord. **Verify:** paper-mode unit: monkeypatch executor.sell to fail twice then succeed; position exits on third heartbeat, sold-leg accounting intact. ### 1.7 Sync `config.live.json` follow set It still lists `shisan888` (long gone). Copy the current `wallets` list + pinned floors from `live/copybot.paper.json` (post-demotion: all volume class, Stavenson/0x4bFb/ArbTrader pinned at $4,760/$1,090/$168), keep the $5 caps and `bankroll_pct` small (0.02 ≈ $5 stakes on a $250 book — recompute so `bankroll_pct × bankroll_usd ≈ $5`). **Verify:** `./live/deploy_bot.sh`-style preview: `python3 -c` validation prints 8 wallets and the caps unchanged. ### 1.8 Dashboard: live-money section The dashboard reads `live/copybot_live_real.json` (from 1.1) 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). Statuses and the `ledger_drift`/`adjustments` fields render like the paper section. **Verify:** point it at a copy of the paper feed with `"mode":"live"`; preview renders; then empty-state renders when the file 404s (live test not started). **Phase-1 exit criteria:** all verifications green; paper bot untouched and heartbeating (`flyctl logs`); `git push` after each item; the July paper feed never skipped a beat. --- ## Phase 2 — account + funding (USER ONLY — agent stops here until told) Follow [LIVE_TEST.md](LIVE_TEST.md) §1–2, with these updates: 1. Dedicated Polymarket account (email login recommended → `signature_type: 1`). **Deposit through the Polymarket UI** (~$50 USDC) so the collateral token is whatever the venue expects — do not send tokens directly on-chain (USDC vs USDC.e confusion is a real loss mode). 2. Export the private key (Settings → Export private key). Fill `config.live.json` `live.private_key` + `live.funder_address` (profile deposit address). The file is gitignored; keep it that way. 3. Send ~1 POL to the **signing EOA address** (preflight prints it) for redeem gas. Skipping this = set `live.auto_redeem: false` and redeem manually. 4. Decide the venue box (0.1/0.2): **recommended = the Fly worker** (geo-clean, push-speed). Running on the Mac requires being in an unrestricted country. --- ## Phase 3 — preflight + dry runs (agent, with user watching) 1. `python3 preflight_live.py` — all green (auth, USDC balance, book access, POL gas). Fix reds before proceeding. 2. `python3 copybot.py --config config.live.json --state /tmp/dry.json --test-wallet ` — paper-mode routing of a real trade through filter+engine. Confirms config/filters without orders. 3. If running on Fly: `flyctl secrets set LIVE_PRIVATE_KEY=… LIVE_FUNDER_ADDRESS=…` (values piped, never echoed), plus a **separate Fly app or process group** decision: do NOT convert the paper worker — run a second app `wwf-copybot-live` from the same image with `mode:"live"` env config, so paper and live books never share a process. Update the Alchemy webhook? No — live app runs **poll mode (60s)** for the test; push webhooks stay pointed at the paper app. Poll lag ≈ 39s avg was measured fine. 4. Baseline sanity: live app boots, logs `on-chain settle fallback: ON`, `geocheck TRADABLE`, prints caps, **waits at the confirmation phrase** — which only the user types (via `flyctl ssh console` attach or Mac run). **Abort at any point:** Ctrl-C / `flyctl apps stop wwf-copybot-live`. Nothing is armed until the phrase is typed; nothing is at risk beyond the deposit. --- ## Phase 4 — first fill (supervised) Arm during a liquid window (evening EU/US — ITF + esports flow). Then wait for the first conviction signal. For the FIRST fill, verify every link end to end: - [ ] console `FOLLOW …` line → `OPEN … [LIVE] buy` with order id - [ ] fill appears in `copybot_fills.live.jsonl` with lag + slippage - [ ] position visible in the Polymarket UI (the account actually holds it) - [ ] feed `live/copybot_live_real.json` row status `open`; dashboard renders it - [ ] `state.cash` dropped by cost+fee; 1.4's CASH≠CHAIN check silent - [ ] Polygonscan: the fill's on-chain transfer matches shares/price ±rounding If any link fails: stop (Ctrl-C), snapshot state + logs, diagnose before re-arming. --- ## Phase 5 — edge-case matrix (the actual test) Run days-to-weeks until every row has a ✓ with evidence (log line + feed row + chain tx where applicable). Most trigger naturally from the follow set's flow; patience beats forcing. | case | trigger | acceptance evidence | |------|---------|---------------------| | BUY copy | any conviction signal | Phase-4 checklist | | ADD (signal scales in) | whale re-buys same market | position grows ≤ stake rule; feed cost/shares aggregate | | **SELL before resolution** | signal fully exits (they exit most positions — hours, not days) | `EXIT … [LIVE] sell` + proceeds land in balance + status `closed`/SOLD + 1.6 retry logs if book thin | | hold → WIN | resolution | `SETTLE … WON` + **auto-redeem tx hash on Polygonscan** + USDC balance up by shares×$1 | | hold → LOSS | resolution | `SETTLE … LOST`, no redeem attempted, book realizes −cost | | hold → **REFUND 50/50** | ITF walkover/abandonment (≈28% of markets — will happen) | `SETTLE … REFUND ↩` + redeem tx at $0.50/share + cash up by shares×0.5 | | missed (price guard) | price runs >5% in the copy lag | `skip … price moved` + missed row on dashboard, later settled hypothetically | | missed (cash gate) | exposure cap hit | `capital fully deployed` skip + missed row | | downtime reconcile | stop the live app 30+ min during flow, restart | boot logs `reconcile:` lines; downtime entries → missed("bot offline"); exits mirrored or settled | | FAK no-fill | thin book (observe, don't force) | order killed cleanly, no resting order in the UI, retry per 1.6 | | neg-risk market | if one gets copied | `⚠ NEG-RISK … redeem manually` warning; manual UI redeem works | | **WITHDRAW** | user, at the end | USDC withdrawn via UI back to source; final book cash ≈ withdrawal ± dust; screenshot + final feed snapshot committed | Throughout: `ledger_drift` 0.00 every heartbeat, CASH≠CHAIN silent, and the paper book running in parallel as control. --- ## Phase 6 — wind-down + verdict (user + agent) 1. Stop the live app; final feed + state snapshot committed (tagged `live-test-1`). 2. Write `FINDINGS.md` addendum: measured live slippage vs paper's −3.6%/65s, fee accuracy, fill rate (FAK kill rate), any stuck exits, refund handling. 3. **Graduation gate (all required):** every Phase-5 row ✓ · ledger drift 0 throughout · live slippage within 2× of paper model · no manual state surgery needed. Then — and only then — a sizing discussion (separate decision, separate doc; the their-bet ceiling and depth-aware fills from README "Scale carefully" apply). ## Standing notes for the executing agent - Never print/commit the private key; `git status` before every commit. - The daily 08:00 Mac pipeline and the bot's self-commits keep pushing to this repo — always `git pull --rebase --autostash` before pushing. - The paper worker's webhook/feed/state are load-bearing for the July test: if a change might touch them, it's a Phase-1-style isolated change with its own verification first. - When something looks wrong, the debugging order that worked in the audit: feed → state → fills ledger → data-api → CLOB → **chain payouts** (the chain is always right).