# Stanford demo: live signal → Polymarket order preview

This is an 8-minute, repeatable demo for Claude Code or Codex CLI. It uses the
BlockRun `trading` profile only: nine tools, no media schemas, one local wallet,
and no vendor API keys.

## What the audience sees

1. The agent discovers a current, liquid Polymarket question.
2. It combines live BTC spot, probability history, smart-wallet positioning,
   and liquidity into one evidence table.
3. It states both the thesis and the strongest counterevidence.
4. It builds a live, executable-size CLOB order preview through the read-only
   Polymarket tool.
5. It proves no order was signed or submitted.

At Stanford, stop at dry-run. Polymarket blocks new orders from the United
States; do not alter networking to bypass that restriction. A real-money write
test must be performed separately while physically operating from an eligible
jurisdiction and only after explicit approval of the exact trade.

The shipped default routes CLOB traffic through BlockRun's Finland egress, so
`setup` will report a permitted region even from the venue. For the demo to be
region-honest, set `POLYMARKET_CLOB_HOST=https://clob.polymarket.com`
explicitly on the presentation machine (the install commands below do this) and
let `setup` report the real status. Never type `confirm:true` regardless.

## Install the isolated Trading profile

Claude Code:

```bash
claude mcp add blockrun-trading -s user \
  -e BLOCKRUN_BUDGET_LIMIT=0.15 \
  -e POLYMARKET_MAX_BET_USD=5 \
  -e POLYMARKET_MAX_SESSION_USD=5 \
  -e POLYMARKET_CLOB_HOST=https://clob.polymarket.com \
  -- npx -y @blockrun/mcp@latest --profile trading
claude mcp list
```

Install the demo Skill from the BlockRun marketplace:

```bash
claude plugin marketplace add BlockRunAI/blockrun-mcp
claude plugin install signal-to-trade-demo@blockrun-mcp
```

Codex CLI:

```bash
codex mcp add blockrun-trading \
  --env BLOCKRUN_BUDGET_LIMIT=0.15 \
  --env POLYMARKET_MAX_BET_USD=5 \
  --env POLYMARKET_MAX_SESSION_USD=5 \
  --env POLYMARKET_CLOB_HOST=https://clob.polymarket.com \
  -- npx -y @blockrun/mcp@latest --profile trading
codex mcp list
```

Before the fixes are published, replace the `npx` command with:

```bash
node /ABSOLUTE/PATH/blockrun-mcp/dist/index.js --profile trading
```

Expected tool surface:

```text
wallet  price  dex  markets  surf  defi  rpc  polymarket_read  polymarket
```

If image, video, chat, or phone appears, the wrong profile is installed.

## Main prompt

Use the `signal-to-trade-demo` Skill and say:

```text
Use the signal-to-trade-demo Skill. Spend at most five US cents on public data.
Find the best current liquid Bitcoin threshold market, build the four-source
signal, and preview one five-US-dollar FOK order on the best-supported side
through the read-only preview action. Do not access account data, submit, or
retry. Redact identifiers and
finish with the exact line:
DRY RUN — no order signed or submitted.
```

For the most reliable stage pacing, use two prompts instead:

```text
Use the signal-to-trade-demo Skill. Spend at most five US cents on public data.
Find the best current liquid Bitcoin threshold market and return the four-source
signal. Do not access account data or place/preview an order.
```

```text
Using that exact market and the best-supported side, create one five-US-dollar
FOK dry-run through blockrun_polymarket_read action preview. This is an
execution preview, not permission to trade. Do not access wallet, setup,
positions, orders, or resources; do not call blockrun_polymarket; do not submit
or retry. Redact identifiers and end exactly:
DRY RUN — no order signed or submitted.
```

## Stage flow

### 0:00–1:00 — one-line product story

“BlockRun lets an agent pay for several live data sources with one self-custody
wallet, turn those sources into a traceable signal, and act on the same market
without creating API accounts.”

Show the nine-tool Trading profile. Point out that order preview lives inside
`polymarket_read`, separated from the funds-affecting trading tool, so MCP
clients do not request write approval for a dry-run.

### 1:00–3:30 — dynamic market and signal

Let the agent select a case live. The primary case is a near-dated Bitcoin
price threshold. On 2026-07-26, the tested candidate was “Will Bitcoin reach
$67,500 in July?”; reselect on presentation day instead of relying on it.

Require the agent to show:

- market probability and exact resolution source;
- current BTC price and percentage move required;
- probability change over a fixed window;
- smart-money buyer share plus aggregate PnL;
- recent volume/trades and an executable book.

### 3:30–5:30 — explain the evidence, not a magic score

The tested snapshot on 2026-07-26 showed:

| Evidence | Observation |
|---|---|
| Market | Yes about 31%; $2.03M total volume; about $37.4k and 541 trades in 24h |
| BTC spot | About $64,698; roughly 4.33% below the $67,500 trigger |
| Probability trend | About 24% → 31% over the tested 24-hour window |
| Smart-money cohort | 647 wallets; 70.48% net buyers; about $1.77M volume |
| Counterevidence | Cohort realized PnL about -$28.6k and total PnL about -$48.6k |

This is mixed evidence: positioning and probability momentum lean Yes, while
the remaining price move and negative cohort PnL argue for caution. The agent
should be allowed to return `NO TRADE`; forcing a bullish answer would make the
demo less credible.

### 5:30–7:00 — executable order preview

The agent calls `blockrun_polymarket_read` with `action:"preview"` for a fixed
$5 FOK dry-run. Five dollars reliably clears the common five-share minimum at
any valid probability while remaining inside the demo caps; the MCP still
rejects it if a market has a higher minimum or insufficient depth. It shows the
estimated fill and cannot sign or submit an order.

### 7:00–8:00 — close on safety

- local signing and self-custody;
- $5 per-order and $5 per-session demo caps;
- $0.15 API budget;
- pre-payment parameter validation;
- per-call payment authorization signed locally;
- explicit confirmation gate plus regional eligibility enforcement.

## Tested fallbacks

1. A different BTC threshold with clear Binance 1-minute-candle resolution.
2. A high-volume Fed/rates question, with cited live news kept separate from
   market-implied probability.
3. A live crypto up/down window only when activity and the order book are
   non-zero. Never select the first future placeholder from the raw feed.
4. If data is stale, liquidity is thin, or the region is blocked, finish with
   `NO TRADE` or dry-run. Do not hide the failed gate.

## Presenter checklist

- Update to the release containing these fixes; run `claude mcp list` or
  `codex mcp list`.
- Confirm wallet funding privately; never put an address on the projector.
- Run setup once before screen sharing; verify approvals and blocked-region
  status without showing identifiers.
- Keep a screenshot of the tested evidence table as a network fallback.
- Close unrelated terminals and notifications.
- Never type `confirm:true` during the Stanford session.