Benchmark

What this is

A fixed sample of 200 Robinhood Chain coins: 100 that rugged within 24 h of launch and 100 that were still trading or had graduated. Run on 2026-11-08. Everything below is computed from one published file: benchmark-2026-11-08.json.

Toll put 84.0% of the rugs it answered at against, and 28.0% of the survivors it answered too. Of the rugs, 0 were called clear and 16 careful; every one is listed under Toll’s misses.

Method

• Toll’s answer is the call it filed before the outcome. About 3 minutes after every launch Toll’s launch checker writes down its call (against, careful or clear) and freezes it; 24 hours later the coin is scored (/record). The benchmark reads that frozen call, so nothing in Toll’s column was known after the outcome. Nothing was asked of Toll again.
• The other tools were asked on the run date, after the outcome. No other checker publishes calls frozen at launch, so the only way to ask them is now. A coin that rugged a week ago looks different today (its pool is often empty), so their column is not like for like with Toll’s. Read the two as different questions answered with different information.
• Each tool read two ways. Its stop word only (strict), and its stop or caution word (broad). “Rugs flagged” is the share of rugs it answered that it flagged; “survivors flagged” is the share of survivors it answered that it flagged (false alarms). Coins a tool did not answer are left out of both and counted under “not answered”.

The sample

Robinhood Chain coins from Toll's launch record launched from 2026-10-30T00:00:00Z up to 2026-11-06T00:00:00Z. A rug = outcome rugged 24 h after launch; a survivor = outcome still trading or graduated. Dead coins (no rug, trading stopped) and coins whose outcome could not be read are left out. In each group the coins are ordered by the SHA-256 of the lowercase contract address and the first 100 kept; the first 10 of each group are the comparison set sent to the other tools.

The window held 8,766 scored launches: 5,671 rugs and 705 survivors (the rest died without a rug or could not be read). By launchpad, the sample is: no launchpad 59, unnamed launcher (0x1cba…) 43, Pons 38, unnamed launcher 30, pools.trade 12, unnamed launcher (0x3194…) 7, Long 7, Doppler 2, Bankr 2. Most rugs on Robinhood Chain are coins launched straight into a pool with no launchpad or by an unnamed launcher, and the sample keeps that mix, so the two groups also differ by launchpad.

Results

Toll, the whole sample (200 coins)

No coin in the sample was called clear, so the broad reading (against or careful) flags every coin: it tells rugs and survivors apart no better than flagging everything. The strict reading is the one that separates them.

Every tool, the comparison set (20 coins)

The same 20 coins asked of every tool. Small on purpose: it fits inside each tool’s free calls, and nothing was paid for. Toll before the outcome; the others after it.

Toll’s misses

Rugs whose frozen call was not against: 0 rugs called clear, 16 called careful, 0 not answered (too little to read at 3 minutes).

Who was compared

Every pre-trade checker we found that answers about a Robinhood Chain (4663) coin, and how it was asked. Only free calls were used; no money was spent. Each tool’s answers are its own; the mapping above is ours.

• Actuary answers ALLOW, DENY or REVIEW from rules over contract facts and its own first-hour pool collector. Its default preset actuary-default-v1 was used; under it a fact it could not read gives REVIEW. On this set 17 of its 19 answers were REVIEW because a fact was unknown.
• presign-guard was asked through its free verdict-only tool (token_quick_verdict), which returns green, orange or red and a grade. Its paid full verdict also carries Toll’s launch facts as an information source; the free tool used here does not ask Toll, so its answers are its own.

Not run, or not comparable: why

Re-run it

The script is web/scripts/benchmark.ts. It reads only public HTTPS endpoints: Toll’s launch record at /api/launches and each tool’s free endpoint, at most one request a second per host (presign-guard one every 6 minutes, inside its free hourly calls), and stops at the first refusal. The sample rule is deterministic, so the same window gives the same coins. Download the data: /data/benchmark-2026-11-08.json (every coin, Toll’s frozen call and reason, and every tool’s answer with the time it was read).

cd web
# the sample (from the public launch record) and every tool's answer
npx vite-node --config vitest.config.ts scripts/benchmark.ts -- --out=public/data/benchmark-new.json
# one tool only, resuming where its free calls ran out
npx vite-node --config vitest.config.ts scripts/benchmark.ts -- --out=public/data/benchmark-new.json --tools=presign --retry

What it is not

• Not advice, and not a ranking of tools: the tools answer different questions with different information at different times.
• A sample, not every launch. The comparison set is small, so a single coin moves its numbers by 10 points.
• A rug is the outcome the launch record scores; a coin can still go badly after 24 hours, and a survivor can still rug later.
• Toll’s tiers changed on 29 Oct 2026; the window starts after that, so every call here was made with the tiers in use since then.