Safety layer · scored in public
Pre-sign record
Every time an agent asked Toll's pre-sign check about a coin it was about to buy or approve, the answer was written down: ALLOW, WARN or BLOCK. 24 hours later we look at how the coin stood. This page is that score, the calls that went against us included.
The calibration, re-run every week
A coin finding makes the answer BLOCK only when coins carrying it rugged at 40% or more within 24 h. The rates in code come from the calibration of 2026-10-29 (2026-10-13 to 2026-10-29); each week the same measure runs again on the last 28 days of launches.
This week's run: 8 Nov 2026, 16:25 UTC, launches from 11 Oct 2026 to 8 Nov 2026 (28 days), 34,414 coins scored, 49.1% of them rugged within 24 h (the baseline). The thresholds in code are not changed by a run; a change would be made by hand and said here.
| Finding that makes it BLOCK | In code: rugged (coins) | This week's run: rugged (coins) | First half / second half |
|---|---|---|---|
| wallets bought a large share in the first 3 s of the pool | 93.8% (988) | 72.7% (2,267) | 69.5% / 75.6% |
| the dev sold 5% of supply or more | 68.1% (1,387) | 70.1% (2,375) | 66.8% / 72.7% |
| first-3 s pool buyers took 20% or more | 58.0% (50) | 79.5% (195) | 67.9% / 84.2% |
| a declared bundle under 1% of supply | 57.8% (554) | 54.4% (754) | 55.2% / 53.2% |
| the ten biggest wallets hold 70% or more | 51.8% (904) | 72.9% (5,012) | 58.6% / 89.6% |
| the dev among wallets selling in one tenth of a second, 20% or more | 50.9% (527) | 50.1% (698) | 51.4% / 48.6% |
| linked wallets hold 50% or more now | 45.9% (233) | 58.5% (552) | 47.8% / 70.5% |
| one wallet holds the main pool's liquidity and can pull it | 45.2% (1,996) | 83.3% (9,021) | 72.2% / 94.3% |
| a wallet sold 5% of supply or more into the pool or curve | 44.4% (2,777) | 56.5% (6,004) | 34.3% / 79.2% |
Under 30 coins a weekly rate says “too few”. One count per coin. Coins we could not read back at 24 h stay in the totals.
How it is scored
What is logged. A pre-sign check (MCP check_tx, POST /api/tx/check, the paid x402 route or the Protect RPC) whose whole answer names a coin bought in a swap or a coin approved. Kept: the coin, the answer, its top reason codes and when. The signing wallet is kept only as a keyed hash so repeats can be counted once, never the address; no calldata, no amounts. One call per asker, coin and answer an hour. An answer still waiting on a coin read, or past the free checks, is not a call and is not logged.
Scored 24 hours after the call, about every 10 minutes, the way the launch record scores a pool coin: rugged when DexScreener shows under $100 of liquidity; dead when under $50 traded in the 24 hours, or its market cap fell to a tenth of what it was within minutes of the call; otherwise still trading. A fact we could not read makes it not read, never still trading. A coin still on a launch curve with no pool on DexScreener reads as not read.
Why it matters. Anyone can say their check works. This is every call, written before the outcome existed, and scored whatever it says. The same numbers as JSON: /api/record/tx. What the check reads: Safety layer. The launch checker's own record: the record.