Early access — cascade metrics are real (derived from canonical token telemetry); the operator field is a curated seed. Learn more about the data
◈ Score your cascade

How much signal does your token cascade actually compound?

Two paths. Both get you scored. Pick the one that fits.

The four pillars

Every turn, tokens flow through the model in four ways. Two go in (fresh input + reused cache). Two come out (the answer + saved context). The ratio between them is your cascade.

Inputfresh tokens · $3/MCache Readreused context · $0.30/MMODELprocesses tokensOutputthe answer · $15/MCache Writesaved context · $3.75/Mentersreusedgeneratessavesnext turn: saved context becomes reusable context↓ to you

Cache read costs $0.30/M — 10× cheaper than fresh input at $3/M. The cascade is the art of replacing input with cache read.

Why it matters

More tokens doesn't mean better. The operators with the highest yield use almost no fresh input — they compound cached context instead. Volume-based leaderboards reward the wrong thing. SigRank surfaces the operators who actually compound.

487B tokens

ranked #1 by volume

Input33.7%
Cache Read60.9%
Output5.3%
Yield0.02

253M tokens ★

ranked #1,142 by volume

Input0.01%
Cache Read94.2%
Cache Write5.2%
Yield1,825

1,900× less volume. 91,000× more yield. Composition is the only variable that matters.

Fastest path:

$ bunx sigrank

Reads your local logs, scores your cascade, puts you on the board.

How to get scored

Two paths. Both get you scored. Pick the one that fits.

Manual upload

No install. No account. Paste four numbers, get your score. Takes 30 seconds.

1

Open your AI usage dashboard

Go to your platform's usage page. Claude Code: ~/.claude/usage.json. OpenAI: platform.openai.com/usage. Cursor: settings → usage.

2

Find your four token counts

Look for: input tokens, output tokens, cache creation (write) tokens, cache read tokens. These are the four pillars. Copy each number.

IOCWCR
3

Paste them into the calculator

Drop the four numbers into the paste calculator. Instant result: Υ Yield, leverage, velocity, class tier, and your archetype.

paste four numbersΥ=1,825Cache Builder ★

Agent install

Auto-reads your logs. Signed submissions. You're on the board. The real path.

1

Install the SigRank CLI

Run npm install -g sigrank. Bundles ccusage + tokscale + tokendash. Node ≥18, macOS + Linux. One install, everything you need.

$npm install -g sigrankadded 1 package in 2.3s✓ sigrank installed+ ccusage, tokscale, tokendash
2

Enroll your device

Run sigrank enroll. Generates an ed25519 keypair for signed submissions. The agent reads your local logs on-device — never your prompts.

$sigrank enrolled25519 keypairscanning local logs...
3

Submit to the board

Run sigrank submit. The agent counts your four token pillars across all sessions, derives your cascade, and posts a signed snapshot. You're on the board.

$sigrank submit✓ signed (ed25519)✓ submitted to board#42youΥ=1,825
Where do you sit?

The field separates into 10 build archetypes. Yield ranges from 0.05 (Input-Bound) to 469 (Kinetic) — a 9,000× spread. Your cascade shape decides which one you are. Where are you?

Input-Bound6.8%Υ=0.05Priming9.4%Υ=0.32Contextual11.7%Υ=0.84Deep Reader10.7%Υ=1.52Builder17.5%Υ=1.12Archivist11.7%Υ=3.54Recursive8.3%Υ=7.23Amplifier10.2%Υ=27Convergent6.6%Υ=257Kinetic ★7.1%Υ=469log scale · 9,000× spread from worst to best

Paste four numbers and find out which archetype you are. Takes 30 seconds.

Want to understand the cascade model in depth? Learn how it works →

⊙ Data privacy

Zero-paste, on-device read

tokenpull reads local session logs and counts the four token pillars across 7d / 30d / 90d / all-time — no copy-paste, nothing to assemble by hand.

Token counts only

The agent counts tokens. It never reads the content of your prompts or replies. Only the four integers leave your machine.

Signed submissions

Each snapshot is signed with an ed25519 keypair (per-device). The server verifies the signature before accepting. No spoofing, no spoofed ranks.

Read-only by design

The agent is read-only against telemetry. It emits no prompt of its own. It measures without disturbing what it measures.

Token counts only. Never your prompts.