The Four Degrees of Leverage
Read it as a token cascade: Cache : Input : Output. The AA baseline sits at 3.5 : 1 : 0.50 — the modeled average AI user. The median operator on the all-time board sits at 20 : 1 : 0.10 — the typical operator, the 50th percentile of everyone measured. The median of the top 100 lands at 243 : 1 : 1.04, output traded for cache. The top operator on the live board is 553 : 1 : 1.70: every input returns multiple outputs on a deep cache. Four degrees of leverage, each a real skill, and the distance between them learnable. The last three columns are measured live from the all-time board.
Sources: the AA baseline is a static modeled reference (7:2:1 cache-read : cache-write : input ratio from Artificial Analysis pricing data). The other three columns are measured live from the all-time board (auto-pulled at render). Human Center of Mass = median of all real operators; Power users = median of the top 100 by yield; Top Evals = the single leading operator. 113 outliers separated (see the field analysis). All derived from canonical four-pillar token telemetry. Token counts only. Read the full analysis.
| Metric | AA baseline§ | Human Center of Mass* | Power users† | Top Evals to date‡ |
|---|---|---|---|---|
| Υ Yield | 1.75 | 1.91 | 239.99 | 941.23 |
| SNR | 0.33 | 0.09 | 0.51 | 0.63 |
| Velocity (O/I) | 0.50 | 0.10 | 1.04 | 1.70 |
| Leverage (CR/I) | 3.5× | 20.4× | 242.8× | 553.2× |
| 10xDEV (log₁₀) | 0.54 | 1.31 | 2.39 | 2.74 |
| Efficiency (vs AA 4.0) | 1.00 | 5.26 | 63.64 | 141.90 |
| Operating Ratio (C:I:O) | 3.5 : 1 : 0.50 | 20 : 1 : 0.10 | 243 : 1 : 1.04 | 553 : 1 : 1.70 |
10xDEV read on the log anchor
10xDEV is an exponent, not a multiplier: each whole point is a 10× jump in real cascade amplification (linear = 10^10xDEV).
| Degree | 10xDEV | Linear amplification (10^x) |
|---|---|---|
| AA baseline (7:2:1 modeled reference) | 0.54 | 3.5× |
| Human Center of Mass (median, all human operators)* | 1.31 | 20.4× |
| Power-user median (top 100) | 2.39 | 242.8× |
| Top operator to date | 2.74 | 553.2× |
- Top operator vs AA baseline: +2.20 decades = ~158× more amplification
- Top operator vs median operator: +1.43 decades = ~27× more
- Top operator vs top-100 median: +0.35 decades = ~2× more
10xDEV is an anchor: the telescoping identity (10^10xDEV = cache_read/input) locks the exponent to leverage, so it can't be inflated independently; it has to be earned through the full cascade. Gaining two full points is ~2 orders of magnitude of real amplification, which is why it moves slowly and means a lot.
Sources & provenance
AA baseline (7:2:1 modeled reference) · static
The Artificial Analysis pricing baseline sets a 7:2:1 ratio for cache-read : cache-write : fresh input. This is a modeled reference point, not live data — it represents the “average AI user” operating ratio of 3.5:1:0.5 (cache-read : input : output). Efficiency is 1.00 by definition: all other operators are measured against this. Source: Artificial Analysis pricing data.
Top operator to date · measured live
The top real operator on the live SigRank all-time board (auto-pulled at render via the same operatorTotal path the board uses). Derived from canonical four-pillar token telemetry (input / output / cache_create / cache_read). Token counts only, no prompt content. Source: signalaf.com/board/all.
Power users (top 100 median) · measured live
The median of the top 100 real operators on the all-time board, ranked by Υ Yield. Median, not mean: the top 100 is right-skewed (the top 5 have yields 10-100× higher than the rest), so the median is the honest “typical elite performer” — the operator ranked ~50th out of the top 100. Source: signalaf.com/board/all.
Average users (all operators median) · measured live
The median of ALL real operators on the all-time board, with the 113 outliers separated (they're not deleted — they get their own category — but they don't set the median). Median, not mean: the board is heavily right-skewed (a single IGNITER-class operator with 9 quadrillion input tokens pulls the mean yield to 427 vs the median of 2.5 — a 170× spread). The median is the exact 50th percentile, immune to any outlier no matter how extreme. The trimmed mean (drop 5% each end) still lands at 17.6 vs 2.5 — still pulled by the upper-middle. Median is the cleanest cut. The gap between this column and the AA baseline is the selection effect: real operators who install a token scanner are cache-heavy (leverage ~20× vs AA's 3.5×) but output-light (velocity ~0.09 vs AA's 0.50) — net yield is close to the modeled average, but the composition is very different. Source: signalaf.com/board/all.
Why median, not mean
Token-cascade metrics are not normally distributed. They follow a power law: a few operators compound signal at extreme rates while the long tail burns tokens. The mean is dragged toward the extremes; the median sits at the honest middle. For the “typical operator” column, the median answers “what does the operator at the 50th percentile look like?” — which is the question the chart asks. A mean would answer “what would you get if you pooled everyone's tokens and divided evenly?” — which is a different question, and not a useful one when the distribution is this skewed.
Metric definitions
SNR = O/(I+O) · Velocity = O/I · Leverage = cache_read/I · 10xDEV = log₁₀(transmission × commitment × reuse) · Efficiency = (cache+O)/I ÷ 4.0 (AA baseline 4.0 = (7+1)/2) · Υ = (cache_read × O) / I² · Operating Ratio = cache : input=1 : output. Telescoping identity: (O/I)(C_create/O)(C_read/C_create) = cache_read/input, so 10^10xDEV = Leverage.
§ AA baseline: the Artificial Analysis 7:2:1 modeled reference (3.5:1:0.5 operating ratio). Static — not live data. Efficiency = 1.00 by definition.
* Human Center of Mass: median of all real operators on the all-time board, computed live at render. Excludes staged seeds, The Field, retired/anonymized rows, and the 113 outliers (extreme humans + flagged operators) that would skew the median. See the field analysis for the full outlier separation methodology.
† Power users: median of the top 100 real operators by Υ Yield on the all-time board, computed live at render.
‡ Top Evals to date: the single top real operator on the all-time board, computed live at render.
All signal is monitored. All drift is noted. · SigRank · MO§ES™ · Ello Cello LLC · Token counts only, never prompt content.