Metrics
Every comparable operator metric, defined.
Yield (Υ) — Token Cascade Efficiency
Υ = (cache_read × output) / input² — the headline SigRank metric. What cascade yield measures, how to improve it, and class tier mapping.
Cache Hit Rate — Context Reuse Efficiency
cache_read / (cache_read + cache_write) — how well you reuse prompt-cached context. The highest-leverage AI coding metric.
Leverage — Cached Context Amplification
cache_read / input — how much cached context amplifies your fresh input. What token leverage means and how to increase it.
Velocity — Output per Fresh Input
output / input — output efficiency ratio in AI coding. Why it's a secondary metric and how it interacts with yield.
Signal-to-Noise Ratio (SNR) — Signal Density
output / (input + output) — output share of fresh conversational traffic. What signal vs noise means and its link to the Conservation Law.
Efficiency — Operational Amplification
(cache_read + cache_write + output) / (4 * input) — total operational amplification versus the AA 7:2:1 baseline. The only public metric that includes cache_write.
More metric definitions
Shorter definitions for the remaining metrics. Each stays on this page because it has no separate reference page.
Scale V — Recorded Token Activity
The base-10 logarithm of the sum of recorded input, output, cache-write, and cache-read token counts over the selected measurement window. Every increase of 1 corresponds to ten times the recorded activity volume. The raw Total Tokens figure remains separately available.
Scale V = log₁₀(I + O + W + R), for positive recorded totals. A total of one million tokens gives Scale V = 6; ten million gives 7.
Inputs: Input (I), Output (O), Cache-write (W), Cache-read (R), as nonnegative recorded token counts from the same window.
What it measures: Scale V measures order-of-magnitude token activity, not whether tokens are unique, meaningful, relevant, newly generated, or successfully retrieved. Cache-read tokens can be reused context; these still count as recorded activity, not distinct content.
Limitations: A large token total can reflect repeated context rather than new information. Scale V is not Yield, Leverage, efficiency, unique-content volume, a quality measure, or a validated “needle in a haystack” detection metric. Unavailable or zero totals cannot produce a meaningful positive logarithmic scale.
Test: For a positive total, sum the four raw counters and take log₁₀ of their total; multiplying all counters by ten must increase Scale V by one. Compare against the raw Total Tokens column and CSV export.
Observable: Scale V and raw Total Tokens in the leaderboard’s selected time window.
Falsifiers: If Scale V differs from log₁₀ of the four recorded pillars (for a positive total) or is presented as a retrieval-success signal, the presentation is incorrect.
Evidence: Deterministic formula over token-only telemetry; leaderboard export retains exact recorded total. No assessment of underlying prompt or response content is performed.
Evidence level: Production evidence. Spec version: TTEOP (Token Telemetry Evaluation Operator Protocol) v0.1-draft. Last updated: 2026-10-12.
Lineage: Four-pillar token telemetry, derived measurement. The distinct efficiency and reuse metrics are explained separately.
Related: Yield (Υ) · Leverage · Velocity · Leaderboard
Context Activity Ratio
The ratio of cache activity (creation + read) to fresh input. Measures how much an operator invests in context vs fresh prompting.
context_activity_ratio = (cache_creation + cache_read) / max(input, 1)
Inputs: cache_creation, cache_read, input.
What it measures: Context activity ratio measures the operator’s investment in building and reusing context relative to fresh input.
Limitations: High ratio may indicate efficient context reuse OR may indicate excessive caching overhead.
Test: Verify the formula for known inputs.
Observable: The numeric context activity ratio value.
Falsifiers: If the ratio does not equal (cache_creation + cache_read) / max(input, 1), the implementation is wrong.
Evidence: Computed from field data. Used in workflow friction analysis.
Evidence level: Repeated experiment. Spec version: TTEOP (Token Telemetry Evaluation Operator Protocol) v0.1-draft. Last updated: 2026-09-01.
Lineage: Upsilon workflow friction analysis. Architecture: mos2es.com/architecture.
Related: Output Flow Share · Active Output Share
Compression Ratio — word-era (historical)
Compression Ratio (ticker COMP, canonical id M.01) is a word-era SigRank metric. It is retained as background for the planned evolution to word-based ranking, and the board still shows it as a legacy Compression view. It is not one of the token-era metrics.
Compression Ratio was the first of SigRank’s original Core 5 metrics, alongside Prompt Complexity and the others. It was a bounded 0 to 1 score rather than a count: the SigRank badge criteria for structural badges such as Compression Forge are written as thresholds on that 0 to 1 scale (for example, a Compression score of 0.85 or higher). This site does not publish a token formula for Compression Ratio.
The token-era output-per-fresh-input ratio is Velocity: Velocity = output / input, a secondary metric and one of the two terms of Yield (Υ), the headline: Υ = (cache_read / input) × (output / input), which is leverage × velocity. For how the board is built, see the methodology.