The Three Degrees of Leverage
Read it as a token cascade: Cache : Input : Output. Research pegs the average user near 7 : 2 : 1 (~2 input tokens per output, on a ~7 cache); input-normalized that's 3.5 : 1 : 0.5. We surveyed 10 power users (median ~500B total tokens) at about 22 : 1 : 0.08, output traded for cache. The top operator on the live board is 367 : 1 : 1.50: every input returns multiple outputs on a deep cache. Three degrees of leverage, each a real skill, and the distance between them learnable.
Sources: the top operator is measured live from the all-time board (auto-pulled, refreshed daily); the power-user median is a measured survey (n=10). Both derived from canonical four-pillar token telemetry. Token counts only. AA 7:2:1 is a modeled baseline from Artificial Analysis methodology (not measured; a reference floor).
| Metric | Average Users* | Power users† | Top Evals to date‡ |
|---|---|---|---|
| Υ Yield | 1.57 | 1.51 | 552.53 |
| SNR | 0.33 | 0.07 | 0.60 |
| Velocity (O/I) | 0.50 | 0.08 | 1.50 |
| Leverage (CR/I) | 3.2× | 22.3× | 367.3× |
| 10xDEV (log₁₀) | 0.50 | 1.35 | 2.56 |
| Efficiency (vs AA 4.0) | 1.00 | 5.61 | 96.01 |
| Operating Ratio (C:I:O) | 3.5 : 1 : 0.50 | 22 : 1 : 0.08 | 367 : 1 : 1.50 |
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) |
|---|---|---|
| Average users (AA 7:2:1)* | 0.50 | 3.2× |
| Power-user median | 1.35 | 22.4× |
| Top operator to date | 2.56 | 367.3× |
- Top operator vs AA baseline: +2.06 decades = ~115× more amplification
- Top operator vs power-user median: +1.21 decades = ~16× 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
Top operator to date · measured
The top real operator on the live SigRank board so far (MO§ES™ — the owner's own observer-stripped run). Source: signalaf.com/compare (retrieved 2026-06-27, canonical board compute). Derived from canonical four-pillar token telemetry (input / output / cache_create / cache_read). Token counts only, no prompt content.
Power-user median (n=10) · measured
Operators #5–14 of the live board (public tokscale.ai footprints surfaced on SigRank). Median, not mean: n=10 is small and right-skewed (one operator = 64% of the field's total tokens), so the median is the honest “typical operator.” Source: signalaf.com/board/30d (retrieved 2026-06-21).
Average users (AA 7:2:1) · modeled, not measured
A synthetic reference operator built from the Artificial Analysis blended-price ratio: “we calculate a blended price assuming a 7:2:1 ratio of cache hit, input, and output tokens” (Artificial Analysis methodology) (retrieved 2026-06-21). Ratio cache:input:output = 7:2:1 → input-normalized 3.5:1:0.50. The cache term was split create/read (cw=0.7, cr=6.3) to give the reference a defined cascade (a modeling choice). This row is a constructed reference, not telemetry from a real operator. Do not cite as measured.
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.
* Modeled baseline: synthesized from the AA 7:2:1 ratio, not measured. The cache create/read split is a modeling assumption. Treat as a reference floor, not a real operator's telemetry.
‡ Top operator to date: the gold column is the highest real operator measured on the live board so far (MO§ES™, the owner). The claude-mem memory observer (an MCP that auto-prompts memory, low-input/high-output) inflated the raw owner row by ~25% of output; the figure shown here is the observer-stripped read. The inflated vs clean pair is shown openly on the live board, the instrument measuring its own contamination and removing it.
All signal is monitored. All drift is noted. · SigRank · MO§ES™ · Ello Cello LLC · Token counts only, never prompt content.