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

Cascade Analysis — Understanding Token Flow

Every AI coding session moves tokens through a cascade. Learn to read the flow — and tell compounding signal from burning tokens.

What a token cascade is

A token cascade is the flow of tokens through an AI coding session. Every turn moves tokens through four stages: you send fresh input to the model, the model generates output back, cache-read tokens are reused from prior context via prompt caching, and cache-write tokens are written to cache for future reuse. The cascade is the full picture of how tokens enter, circulate, and leave the session.

The metaphor is deliberate: tokens cascade like water through a series of pools. Some pools compound — cached context flows back into the next turn, amplifying a small fresh input into large output. Other pools drain — fresh input pours in, nothing is reused, and thin output trickles out. Cascade analysis is the art of reading which kind of cascade you have.

How to read your cascade

Read the four pillars together, never in isolation. A high input number is bad if cache-read is low (you are spending without compounding) but fine if cache-read is also high (you are sending fresh input on top of a large cached foundation). A high output number is good only if it is dense — SNR tells you that. The yield metric Υ = cache_read × output / input² is the synthesis: it rewards high cache reuse and high output while punishing high fresh input. If your yield is low, the four pillars tell you why.

What each pillar reveals

Input → Spend

How many fresh tokens you are sending. The cost side. High input without cache reuse means you are paying full price every turn.

Output ← Return

How many tokens the model generates back. The return side. High output is good — but only if it is dense (check SNR).

Cache-read ↻ Compounding

How much prior context you are reusing for free. The compounding layer. High cache-read is the hallmark of a healthy cascade.

Cache-write ✎ Investment

How much context you are storing for future turns. An investment. It only pays off if you read it back — check cache-read.

Diagnostic patterns

The four pillars combine into recognizable shapes. Here are the four most common — and what each one tells you.

The Compounder

High cache-read, low input, high output

Signal is compounding. You reuse cached context well and send minimal fresh input. This is the TRANSMITTER pattern — the cascade architecture yield rewards.

The Burner

High input, low cache-read, low output

Tokens are burning. You send large fresh prompts without reusing context and get thin output back. Yield is low. Cut input, build cache, reuse context.

The Echo

High input, high output, low cache-read

The model is parroting your input back. Compression ratio may look fine, but without cache reuse the cascade is not compounding. Yield stays flat because input² punishes the spend.

The Hoarder

High cache-write, low cache-read, low output

You are writing context to cache but never reading it back. The investment is not paying off. Either the cached context is not reusable or you are not structuring turns to recall it.

Explore the category

FAQ

What is a token cascade?
The flow of tokens through an AI coding session across four stages: input, output, cache-read, and cache-write. The cascade is the full picture of how tokens enter, circulate, and leave a session.
What is cascade analysis?
The study of token flow through sessions. By reading the four pillars you diagnose whether signal is compounding or tokens are burning. Yield (Υ) summarizes the architecture in one number.
How do I read my cascade?
Read the four pillars together. High cache-read + low input + high output = compounding. High input + low cache-read + low output = burning. The diagnostic patterns on this page map the common shapes.
What does each pillar reveal?
Input reveals spend, output reveals return, cache-read reveals compounding, and cache-write reveals investment. Together they describe the full cascade architecture.
Compounding vs. burning cascade?
A compounding cascade has high cache-read, low input, high output. A burning cascade has high input, low cache-read, low output. Yield is high for the first, low for the second. The difference is architecture, not model.