pm-manipulation

Read `memory/MEMORY.md` for context.
Read the last 3 days of `memory/logs/` to avoid re-flagging markets you already covered, and to compare current readings against prior ones.

## Voice

If `soul/SOUL.md` and `soul/STYLE.md` are populated, match the operator's voice in the notification and write-up. If empty or absent, use a clear, direct, neutral tone. The methodology itself is the same either way.

## Why this skill exists

English-language coverage of any prediction market is dense for US politics, sparse for everything else. Election-, geopolitics-, and conflict-related markets are most often pushed by actors closer to the home country than to the English-speaking financial press. By the time a coordinated narrative lands in mainstream financial outlets, the trade is over. Local-language press, regional outlets, and country-specific Telegram / X channels move first — sometimes by 24–72h.

This skill looks at the **past 3 days** of activity on a configurable prediction-market platform and asks: where does the price action diverge from organic English-press coverage, and does a multilingual sweep reveal a narrative push that explains it?

## Configuration

Read `memory/topics/prediction-markets.md` if it exists for an optional `## Platform` line naming the target API root. Defaults to Polymarket's public Gamma + CLOB APIs (`gamma-api.polymarket.com`, `clob.polymarket.com`, `data-api.polymarket.com`). The candidate selection, scoring rubric, and multilingual sweep are platform-agnostic; only the endpoints differ.

The keyword filter and locale table below are starting points — the operator can edit `memory/topics/prediction-markets.md` to add `## Keywords` and `## Locales` sections that override the defaults.

## Network note

`curl` works — there is no network sandbox. For every curl call, if it fails or returns empty, use **WebFetch** as a fallback for a flaky public GET. The Polymarket APIs above are public (no auth). For news searches, prefer **WebSearch** with locale-specific queries — a built-in Claude tool that needs no curl.

## Steps

### 1. Build the candidate set (markets active in past 3 days)

Pull the most-active markets across the categories most prone to manipulation: politics, geopolitics, conflict, elections (national and regional), regulatory rulings, and resolution-disputed markets.

```bash
# Top markets by 24h volume — repeat for last 3 days using volume24hr / volume1wk fields
curl -s "https://gamma-api.polymarket.com/markets?closed=false&order=volume24hr&ascending=false&limit=30"

# Also fetch by 7d volume to catch slower-burn manipulation
curl -s "https://gamma-api.polymarket.com/markets?closed=false&order=volume1wk&ascending=false&limit=30"
```

From the union of those two lists, pick **6–10 candidates** that meet at least one of:
- Question references a non-US country, region, or conflict
- Question references a regulatory body, election, or coup/conflict event
- 7d volume > $500k AND 24h volume > $50k (real money, not just liquidity-mining)
- Slug contains keywords (default set; operator can override in `memory/topics/prediction-markets.md`): `russia`, `iran`, `israel`, `china`, `taiwan`, `ukraine`, `venezuela`, `argentina`, `mexico`, `brazil`, `india`, `pakistan`, `nigeria`, `france`, `germany`, `italy`, `spain`, `eu-`, `nato`, `cartel`, `coup`, `nuclear`, `assassination`, `ceasefire`, `election`, `vote`, `referendum`, `oracle`, `dispute`, `umip`

Skip anything sports, weather, or pure crypto-price (BTC > $X by Y) — those have different manipulation signatures and aren't this skill's job.

### 2. Pull 3-day price + trade data per candidate

For each candidate market, get the YES-token (`clobTokenIds[0]`) and pull 3-day price history at hourly fidelity:

```bash
curl -s "https://clob.polymarket.com/prices-history?market=$TOKEN_ID&interval=1w&fidelity=60"
```

Compute, from the past 72h window only:
- **3d open / close / high / low** (price 0.0–1.0)
- **Largest single-hour move** (absolute pp change)
- **Move concentration** — what % of the 3d net move happened in the single largest 6h block? If > 70%, flag `concentrated-move`.
- **Volume spike ratio** — (peak hour volume) / (median hour volume). If > 8x, flag `volume-spike`.
- **Direction inversion** — did price reverse > 5pp within the 72h window? Flag `reversal` — common when a manipulator unwinds.

Also pull the trades feed if available (some markets only) for whale concentration:
```bash
curl -s "https://data-api.polymarket.com/trades?market=$TOKEN_ID&limit=500"
```
Compute **whale share** — % of 3d volume from the top-3 wallet addresses. If > 40%, flag `whale-concentrated`.

If the trades endpoint doesn't return per-market data, skip whale-share for that candidate — note it in the report rather than failing.

### 3. Pull comments and detect coordination patterns

For each candidate, fetch comments from the past 3 days:

```bash
# Get the event id first (comments live on Event, not Market)
curl -s "https://gamma-api.polymarket.com/events?slug=$EVENT_SLUG&limit=1"

# Top comments by reactions
curl -s "https://gamma-api.polymarket.com/comments?parent_entity_type=Event&parent_entity_id=$EVENT_ID&limit=50&order=reactionCount&ascending=false"

# Most recent
curl -s "https://gamma-api.polymarket.com/comments?parent_entity_type=Event&parent_entity_id=$EVENT_ID&limit=50&order=createdAt&ascending=false"
```

Filter to comments from the last 72h (`createdAt` within window). Then look for:
- **Burst posting** — > 10 comments from accounts with usernames matching the same regex pattern (e.g. `user\d{4,}`, common spam-bot signature) within a 6h window. Flag `bot-burst`.
- **Narrative concentration** — > 40% of top-reacted comments push the same one-sided talking point with near-identical phrasing. Flag `narrative-push`.
- **Cross-market spam** — same usernames repeating identical text across multiple unrelated markets (compare against any logged comments from prior `pm-manipulation` runs). Flag `cross-market-spam`.