Operator Scatter
1,598 individual operators splattered across metric pairs · each dot is one human · colored by platform · 2026-08-10
Operators plotted
1,598
Positive yield + leverage (log-scale eligible)
Yield range
0.01 – 2.46M
8 orders of magnitude
Leverage range
0.01 – 30,345×
6 orders of magnitude
Platforms
11
Anthropic 746 · OpenAI 662 · others 190
Reading these charts
Every dot is a single operator. Axes are log-scaled (except SNR). The field forms shapes — clouds, streaks, clusters — that archetype averages hide. Look for the diagonal lines (power-law relationships), the empty quadrants (impossible combinations), and the outliers that break the pattern.
1. Yield vs Leverage — the compounding cloud
Every operator: Yield (Υ) vs Leverage (cache_read/input)
The defining relationship. Yield = (leverage × velocity) / input, so yield should rise with leverage. It does — but with enormous spread. The diagonal streak is the power law. The cloud below it is operators with high leverage but low velocity (deep readers who don't generate much). The dots above the streak are the velocity amplifiers — operators who get more yield than their leverage alone would predict.
2. Yield vs Total Tokens — the volume lie
Every operator: Yield vs total tokens accumulated
If volume predicted skill, this would be a tight upward line. Instead it's a shotgun blast. The highest-yield operator (2.46M Υ) has 67M tokens — modest. The highest-volume operator (9 quadrillion tokens) has yield near zero. The cloud at bottom-right is the "token burners" — massive volume, no compounding. The streak rising from bottom-left to top-left is the real signal: yield comes from composition, not accumulation.
3. Velocity vs Leverage — the operating plane
Every operator: Velocity (O/I) vs Leverage (cache_read/input)
This is the fundamental operating plane — how much you reuse (x) vs how much you generate (y). The field forms an L-shape: most operators cluster at low velocity (0.01–0.3) with leverage spreading from 1× to 30,000×. The vertical streak at left is the "fresh input burners" — high velocity, almost no reuse. The horizontal cloud at bottom is the "deep readers" — extreme leverage, minimal output. The rare dots in the upper-right are CONVERGENT operators — both axes elevated.
The empty quadrant
The upper-right of the Velocity vs Leverage chart is nearly empty. Almost no operator has both high velocity (>1.0) AND high leverage (>100×). This is the tradeoff that defines the field: you either generate fast (burning input) or reuse deeply (reading cache). CONVERGENT operators are the ones who break this tradeoff — and there are only ~100 of them in 1,598.
4. SNR vs Leverage — does reuse improve signal?
Every operator: SNR (output/total) vs Leverage
SNR measures what fraction of your token activity is output (signal) vs input+cache (overhead). The field forms a rising cloud — more leverage correlates with higher SNR, but with massive spread. The cluster at bottom-left (low leverage, low SNR) is the "input dumpers" — they take in lots, produce little, reuse nothing. The streak rising to the right shows that deep reuse enables higher signal efficiency — but only up to about 0.6 SNR. Nobody gets near 1.0.
5. Input Share vs Yield — the input tax
Every operator: Input % of total tokens vs Yield
The x-axis is what fraction of your token budget is fresh input (vs cache + output). The relationship is brutal: yield collapses as input share rises. Operators below 1% input (almost everything is cache reads) span the full yield range. Operators above 10% input are clustered at the bottom — yield near zero. Fresh input is a tax. The operators who minimize it compound signal; the ones who maximize it burn tokens.
6. Cache Read % vs Cache Write % — the reuse architecture
Every operator: Cache read share vs cache write share
How much you read from cache (x) vs how much you write to cache (y). The field splits into two clusters: a massive cloud at high read % (90–97%) with near-zero write % (0–1%) — the "passive readers" who reuse but rarely build new cache. And a smaller cluster at moderate read % (85–95%) with higher write % (2–6%) — the "active builders" who both write and read. The gap between them is the archetype boundary.
7. Tokens/day vs Yield — throughput doesn't predict yield
Every operator: Tokens per day vs Yield
The shotgun. Operators burning 10B+ tokens/day cluster at yield < 10. The highest-yield operators are modest throughput (1M–100M/day). The field's center of mass is a horizontal cloud at low yield regardless of throughput — most operators pump tokens without compounding. The rising diagonal at low throughput is the "efficient quiet" — low volume, high yield, the ghost-rank operators.
8. Active Days vs Yield — does time predict skill?
Every operator: Active days vs Yield
If practice made perfect, yield would rise with active days. It doesn't. The field is a vertical cloud — yield spans 7 orders of magnitude at every experience level. A 5-day operator can outperform a 365-day operator by 100,000×. Experience (total tokens, active days) measures how much you've done, not how well you do it. The yield signal is in composition, not duration.
The vertical cloud
Active Days vs Yield is a vertical cloud — yield spans 7 orders of magnitude at every experience level. This is the strongest evidence that the experience ladder (total tokens) and the yield ranking are independent axes. The two-axis taxonomy exists because one axis cannot capture both. An ARCH+ operator (deep experience) can have yield 0.01 or yield 100,000 — and both are real.
9. Velocity vs SNR — signal speed vs signal purity
Every operator: Velocity (O/I) vs SNR (O/total)
Both metrics measure "how much output" but on different denominators. Velocity = output/input (ignores cache). SNR = output/total (includes cache). The field forms a rising cloud with a ceiling — SNR can't exceed velocity/(1+velocity) because cache reads dilute the denominator. The operators above the main cloud are the ones with minimal cache overhead — their SNR approaches their velocity. The ones below have massive cache reads dragging SNR down despite decent velocity.
Methodology
Data: 1,598 operators from public/data/field-analysis.json with positive yield AND positive leverage (required for log-scale axes). 29 operators with yield=0 excluded from log-scale charts.
Anonymization: No handles or display names shown. Each dot is one operator identified only by platform color.
Metrics: Yield (Υ = cache_read × output / input²), Leverage (cache_read / input), Velocity (output / input), SNR (output / total), Input Share (input / total × 100), Cache Read % (cache_read / total × 100), Cache Write % (cache_write / total × 100).
Axes: All scatter plots use log-log scale except SNR (linear 0–1) and Active Days (linear 0–365). Point size is uniform (2px radius) to show density — overlapping dots create natural density visualization.
Source: public/data/field-analysis.json from public/data/board-all_time.json · scraped from public AI coding agent leaderboards · 2026-07-13