Field Analysis

Volume ≠ Yield

Public token-volume leaderboards rank by total token volume. SigRank ranks by yield — how efficiently an operator converts input tokens into output tokens using cache compounding. These two rankings have almost zero correlation. The operator with the most tokens (9 quadrillion) has a yield of 0. The operator with the highest yield (2.46M) ranks #697 by volume. Volume alone is noise. Yield is signal.

The scatter plot above makes this visible. The median lines divide the field into four quadrants — and the top-right (high volume, high yield) is nearly empty. The highest-yield operators cluster in the bottom-right: modest token spend, extraordinary efficiency. This is the ghost-rank phenomenon, explored below.

The Token Cascade

The median operator puts in 238M tokens of fresh input. They produce 24M tokens of output. They write 72M tokens to cache. And they read 4.77B tokens from cache. That last number is the harvest: 20.5x the seed. This is leverage. The cascade is not a chain of amplifications. It is a seed (input), a tiny sprout (output), a small store (cache write), and a massive harvest (cache read).

The operating ratio compresses this into one fingerprint: C : I : O = 19 : 1 : 0.09. For every 1 token of fresh input, the median operator reads 19 from cache and produces 0.09 output. Yield is what happens when cache compounding meets output production.

The SNR Separation

Signal-to-Noise Ratio (SNR) = output / (input + output). It measures what fraction of your interaction produced actual output versus prompt overhead. Outliers have SNR near zero. Humans have SNR above .05. One number separates signal producers from token burners.

The histogram shows the field clustering tightly around the median SNR of 0.084. The IQR fences (dashed lines) bracket the middle 50% of operators. The long tail to the right — operators with SNR above .10 — are the ghost-rank operators: they produce disproportionate output from minimal input.

Leverage × Velocity

Leverage (cache_read / input) measures how much cached context amplifies each fresh input token. Velocity (output / input) measures how much the model generates per token of fresh context. Together, they define the yield rectangle — the area of leverage × velocity approximates how efficiently an operator turns cached knowledge into produced signal.

The median crosshair divides the field. Operators in the top-right quadrant — high leverage and high velocity — are the architectural elite. They read deeply from cache and produce rapidly. The bottom-left cluster (low leverage, low velocity) represents the volume-burning majority: fresh input, minimal caching, slow output.

Platform Dominance

Anthropic-primary operators dominate the top yield quartile — 98.5% of the highest-yield operators use Claude as their primary platform. This isn't coincidence: Anthropic's mature prompt caching infrastructure produces higher cacheRead values, which directly drives yield.

The adoption chart shows raw volume — OpenAI and Anthropic lead in total operator count. But the quartile breakdown reveals the efficiency story: OpenAI dominates the bottom quartiles (high volume, low yield), while Anthropic owns the top. The platform you choose shapes the ceiling of your yield architecture.

Cascade Composition

Four notable operators, four radically different cascade architectures. The stacked bars show how each operator composes their token spend across the four pillars: input (fresh tokens), output (produced signal), cache write (context stored), and cache read (context reused). The outlier at left burns input with zero cache. The high-yield operators at right are dominated by cache read — they reuse context, not burn it.

These operators illustrate the yield spectrum. See their full profiles on the Hall of Signal and learn how the metrics are computed on the methodology page.

Yield Quartile Box Plots

The box plots break down four metrics — yield, leverage, velocity, and SNR — across the four yield quartiles. The progression is stark: leverage jumps from a median of ~5× in Q1 to ~200× in Q4. Velocity climbs from 0.03 to nearly 1.0. But SNR stays flat across all quartiles — the signal density of output doesn't change. What changes is how much cached context amplifies that output.

This is the architectural insight: high-yield operators don't produce denser signal — they produce more signal from the same density by leveraging cache. The yield gap is a leverage gap, not a talent gap.

Where 80% of Operators Live

80% of human operators have a yield between 0.37 and 233.12. The median is 1.68. The distribution is heavily right-skewed (power-law), which is why the median is used instead of the mean.

The yield distribution is heavily right-skewed. 80% of human operators fall within the shaded band. The long tail to the right is where the AMPLIFIERS and CONVERGENT operators live. The bulk of the field clusters near the median. This is why the median is used instead of the mean: the mean is pulled by outliers, the median reflects where operators actually are.

Where Are You?

Median yield is 1.68. You are probably here. Claim your profile to see exactly where you land

The percentile ladder shows the yield thresholds for each tier. The median is where most operators land. The top 1% is where cache architecture becomes an art form. If you use AI coding agents, you are probably near the median. Claim your profile to see exactly where you fit.

Ghost Ranks: The Hidden Operators

Ghost-rank operators are invisible on volume-based leaderboards but dominate yield-based rankings. They use fewer tokens but achieve higher output efficiency. These are the operators worth recruiting — they have skill, not just spend.

The data reveals 50 ghost-rank operators — above median yield but with volume ranks in the hundreds or thousands. Their median volume rank is 1355, meaning they are buried deep on any volume leaderboard. But their yield values reach into the hundreds of thousands. Volume metrics hide them. Yield metrics find them.

The quadrant chart above plots every human operator on a log-log grid of total tokens versus yield. The dashed gold lines mark the median on each axis, splitting the field into four quadrants. Q2 — the top-left, low volume and high yield — is the ghost-rank region, highlighted in cyan. These operators would be invisible on any volume-ranked leaderboard, yet they dominate on yield. They are the operators worth recruiting.

Showing 20 of 50 ghost-rank operators.

Build Archetypes

The field separates into 10 build archetypes across four families: Convergence, Generation, Reuse Depth, and Active Construction. CONVERGENT is checked first and pulls out operators who are elite on all three derived dimensions (leverage, velocity, construction). KINETIC captures high-velocity generation. The Construction branch captures active context builders. The Reuse Depth branch captures passive context reusers. Each type is defined by a different primary dimension of the token cascade.

Outlier Detection

SigRank's metrics catch gaming automatically. A 6-signal outlier-likelihood score identifies operators with inhuman throughput, zero cache usage, single-model fixation, and zero sessions. 130 outliers were separated from the field distribution. An additional input/total ratio analysis separates extreme humans from replay outliers and input dump outliers, keeping the Human Center of Mass clean.

The scatter plot shows why outliers are detectable: they cluster in the bottom-right — massive token volume with near-zero SNR. They pump input tokens without producing proportionate output. No human operator occupies that region. The 6-signal score makes this structural: inhuman throughput, zero cache reads, single-model fixation, and zero sessions are individually suspicious; together they are conclusive.

This is why the Four Degrees chart's columns are honest: the 130 outliers are separated before the median is computed. Without separation, the top outlier alone skews the field average by 248,000%. The median is immune. Read the full analysis or see the Four Degrees of Leverage to see how the clean median compares to the modeled average.