Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs: correct the 3ws vanity-grind odds and state the strategy backtester's limits

nirholasfa8a4f0a9da9
diff --git a/docs/mint-mark.md b/docs/mint-mark.md
index 0cf9c1aab..62d0fc947 100644
--- a/docs/mint-mark.md
+++ b/docs/mint-mark.md
@@ -20,7 +20,7 @@ When you launch a coin on three.ws (via Studio or the agent wallet path), the pl
### How does it work?
-The server runs a fast WASM-based keypair grinder until it finds a match, then uses that keypair as the coin mint. The expected work is ~49 000 keypairs at ~25 000/s single-threaded, typically about a second on the Cloud Run server's CPU.
+The server runs a fast WASM-based keypair grinder until it finds a match, then uses that keypair as the coin mint. The expected work is about 14 500 keypairs (the exact Base58 model in `src/solana/vanity/base58-distribution.js` puts a case-insensitive `3ws` prefix at roughly 1 in 14 490, far easier than the naive 58³/4 because `3` is a common leading character of a 32-byte key) at ~25 000/s single-threaded, typically well under a second on the Cloud Run server's CPU.
### Is it on every coin?
@@ -61,7 +61,7 @@ const enforceMark = env.THREE_WS_MARK_ENFORCE !== '0' && env.THREE_WS_MARK_ENFOR
When enforcement is **on** (default):
- A client-supplied `mint_address` is validated with `hasThreeWsMark`; unbranded mints return `400 unbranded_mint`.
- When no mint is supplied, the server grinds one with `grindVanityNode({ ...THREE_WS_VANITY })`.
-- Grind cost is logged at info level: `{ publicKey, attempts, durationMs }` — expect ~49 000 attempts, < 1 000 ms on the Cloud Run server's CPU.
+- Grind cost is logged at info level: `{ publicKey, attempts, durationMs }`. Expect about 14 500 attempts on average (a long tail past 40 000 is normal), under 1 000 ms on the Cloud Run server's CPU.
When enforcement is **off**: a random `Keypair.generate()` is used (pure-legacy fallback). No brand constraint.
@@ -101,7 +101,7 @@ Every server-side grind emits a structured log line via `logger('pump.launch')`:
"publicKey": "3wsXrT4Gy...", "attempts": 49213, "durationMs": 892 }
```
-Watch this in the Cloud Run logs (Cloud Logging) to confirm the ~49 000-attempt / sub-second expectation holds. A sharp increase in `attempts` or `durationMs` signals WASM performance degradation.
+Watch this in the Cloud Run logs (Cloud Logging) to confirm the roughly 14 500-attempt average and sub-second expectation hold. A sharp increase in `attempts` or `durationMs` signals WASM performance degradation.
### The generic x402 launcher exemption
diff --git a/docs/strategy-objects.md b/docs/strategy-objects.md
index bec5eba11..b9a4925df 100644
--- a/docs/strategy-objects.md
+++ b/docs/strategy-objects.md
@@ -37,7 +37,7 @@ The schema enforces hard bounds on every field, so a stored config can never car
`/strategy-lab` is where you build and prove a declarative pump.fun strategy. The Lab speaks its own tiny strategy DSL (a spec with `scan`, `filters`, `entry`, and `exit` predicate rules), and one compiled evaluator powers validation, backtest, and live runs, so they cannot drift:
- **Validate** (`POST /api/pump/strategy-validate`) checks the spec with the same validator the runner uses, so a broken spec is caught with field-level issues before anything runs. No sign-in needed.
-- **Backtest** (`POST /api/pump/strategy-backtest`, sign-in required) replays the compiled strategy over real recorded on-chain history: for each candidate mint (from the request's `mints`, the spec's `scan.mints`, or a live scan of new or trending launches served by the pump.fun MCP's `get_new_tokens` and `get_trending_tokens`, which read pump.fun's public feed and the platform's own launch recorder, so no external indexer is needed) it fetches the token's real details, holders, bonding curve, and recent trades, evaluates the entry gate at the first recorded trade, then walks the subsequent real trades to trigger the exit rules. It never synthesizes launches or invents price paths; the response's `data` carries `spent`, `realizedPnlSol`, `roiPct`, `tradeCount`, `winRate`, `maxDrawdownSol`, the per-mint `trades`, and `mintsUsed`.
+- **Backtest** (`POST /api/pump/strategy-backtest`, sign-in required) replays the compiled strategy over real recorded on-chain history: for each candidate mint (from the request's `mints`, the spec's `scan.mints`, or a live scan of new or trending launches served by the pump.fun MCP's `get_new_tokens` and `get_trending_tokens`, which read pump.fun's public feed and the platform's own launch recorder, so no external indexer is needed) it fetches the token's real details, holders, bonding curve, and up to 200 recent trades, enters at the price of the first recorded trade, then walks the subsequent real trades to trigger the exit rules. It never synthesizes launches or invents price paths. Know its limits before reading a number: the entry filters are checked against holder, curve, and creator data as fetched when you run the backtest (not as they stood at launch, so later information leaks into the entry decision), fills carry no slippage or fees, and a trending scan only sees coins that survived long enough to trend. Treat it as a check that a strategy behaves as written, not as a statistical estimate of returns; for that, use the sniper backtester below, which reports its own caveats. The response's `data` carries `spent`, `realizedPnlSol`, `roiPct`, `tradeCount`, `winRate`, `maxDrawdownSol`, the per-mint `trades`, and `mintsUsed`.
- **Run** (`POST /api/pump/strategy-run`) executes the strategy for a bounded duration (5 to 600 seconds) and streams events over Server-Sent Events. In **simulate** mode it needs no wallet and spends nothing; in **live** mode it requires auth, an agent you own with a provisioned Solana wallet, and every buy passes a `policyGuard` that calls the same `checkBuyAllowed` spend-policy check the rest of the platform uses.
- **Close all** (`POST /api/pump/strategy-close-all`, sign-in required) exits open strategy positions.