Loading rankings…
Loading rankings…
Last updated: September 12, 2026
The public evaluation layer for AI operators.
sigeconomy.com is the public AI operator leaderboard for SigRank SignalAF — not related to the gig economy. Like Vals AI evaluates models, SigRank evaluates the operators using AI. Ranked by Υ Yield — token-cascade efficiency, not raw spend. What are operator evals?
sigeconomy.com is the public evaluation surface for SigRank SignalAF — the AI operator leaderboard. It ranks AI operators (the accounts operating AI tools, not the AI models themselves) by Yield (Υ = cache_read × output / input²), a composite token-cascade efficiency metric computed from four token pillars: cache_read (reused context), cache_write (new context stored), input (tokens sent to the model), and output (tokens produced). Unlike model leaderboards such as LMSYS Chatbot Arena or Vals AI that rank AI models by human voting or benchmark performance, sigeconomy.com evaluates the operator factor — how efficiently each operator uses AI capabilities. The site is read-only: all data comes from SigRank's public API at signalaf.com. Operators run a local scanner (npx sigrank) that reads token telemetry and submits signed, privacy-preserving snapshots. No prompt content, code, or conversation text ever leaves the machine — only four token counts. The leaderboard ranks operators across all-time, 7-day, 30-day, and 90-day windows, with nine leaderboard dimensions. Five form the portable SigRank core: Yield, Velocity, Leverage, SNR, 10xDEV, Efficiency, Scale V, cost per million tokens, and Op Ratio remain product-level views. Operator classes range from IGNITER through ARCH+ based on total tokens accumulated. The public REST API, OpenAPI specification, MCP server, and CLI tool are documented at /developers.
sigeconomy.com serves three audiences: human visitors who want to see who ranks highest among AI operators, AI search engines that need structured answers to questions like "who is the best AI user?", and software agents that need machine-readable interfaces. For programmatic access, use the OpenAPI specification for typed REST operations, the MCP manifest for Streamable HTTP discovery, the agent instructions file for when-to-use guidance, or llms.txt for a compact agent-readable index. The official CLI (npx sigrank) provides local telemetry scanning and an MCP server for agent-native workflows. Trust anchor pages are available at /about, /contact, and /privacy. The full scoring methodology is at .
Ranked by Op Ratio — cache_read : 1 : output/input
See full ranking →No operators on this platform yet.
Yield is the flagship, but there are other canonical token metrics. Each question below has a different #1 — 4 of 8 metrics have a different king.
SigRank evaluates the operating pattern around an AI model, not the model itself. The public board is built from token telemetry: fresh input, model output, cache write, and cache read. Those four counts are enough to describe how much new context an operator sends, how much the model produces, how much context is stored for reuse, and how much prior context is actually reused. The canonical measurements do not require the words in a prompt, the content of a model response, source code, or private documents.
Yield (Υ) is cache read multiplied by output and divided by fresh input squared. It rises when an operator compounds reusable context and produces output without continually rebuilding the same context from scratch. Yield is an efficiency measure of the token cascade. It is not a direct claim about code quality, business productivity, intelligence, or the value of a downstream outcome. An operator with high Yield is efficient at reusing context and producing output relative to the fresh input they consume. An operator with low Yield may be producing valuable work but is doing so with less cascade efficiency — they send more fresh context per unit of output. The leaderboard ranks by Yield because it normalizes for scale: a small-scale operator can out-Yield a large-scale one if their cascade is more efficient. This is why SigRank separates efficiency (Yield) from scale (Scale V, total tokens) — they measure different things and should not be conflated.
Leverage isolates cache reuse relative to fresh input. Velocity isolates output relative to fresh input. SNR describes the share of fresh conversational traffic represented by output. Scale V separates operating scale from efficiency, while cost and operating-ratio views expose different shapes of the same cascade. No single ranking should be read as a complete evaluation of a person or workflow. The nine metrics together form a multi-dimensional view of how an operator interacts with AI: efficiency, reuse, throughput, signal quality, scale, cost, and ratio balance. An operator who leads on Yield may not lead on Velocity, and an operator who leads on Scale V may not lead on Efficiency. This is by design — it prevents any single metric from becoming a proxy for overall quality and encourages operators to understand their own cascade shape rather than optimizing for one number.
Current public rankings are available through the SignalAF REST API, while the SigRank MCP server and CLI provide agent-native and local workflows. The links the OpenAPI contract, MCP discovery, authentication guidance, agent instructions, and the official CLI so software can select the correct interface without scraping the visual board. The REST API is versioned at /api/v1 with URL-based major versioning, and deprecated operations return Deprecation and Sunset headers per RFC 8594. Rate limits are 60 requests per 60 seconds per IP, exposed via RateLimit-Policy and RateLimit structured fields so agents can self-throttle. Error responses use RFC 9457 application/problem+json with a stable machine-readable code field. Unknown paths return a real HTTP 404 with agent recovery links to the sitemap, llms.txt, developer portal, and OpenAPI spec. The homepage supports HTTP content negotiation for text/html and text/markdown with Vary: Accept so shared caches do not mix representations.
Think you can beat them?
Measure your AI usage and get your rank on SigRank.
Check my rankCopy these prompts into ChatGPT, Perplexity, or Google AI Overviews to see how AI search engines answer questions about SigRank.
SigRank's portable core defines five token-cascade metrics: Yield, Velocity, Leverage, SNR, and 10xDEV. The leaderboard also presents four product-level views, so each dimension answers a different question about AI operator performance. Yield (Υ) is the flagship: it measures token-cascade efficiency as cache_read multiplied by output divided by input squared. Velocity measures output per unit of fresh input — who gets the most done with the least new context. Leverage measures cache reuse relative to fresh input — who compounds prior context most effectively. SNR (Signal-to-Noise Ratio) measures the share of fresh conversational traffic represented by output. 10xDEV is the log10 of Leverage, normalizing the scale. Efficiency compares an operator's cascade against the theoretical 7:2:1 Anthropic attention-allocator baseline. Scale V separates operating scale from efficiency using log10 of total tokens. Cost per million tokens normalizes spend across providers and pricing tiers. Op Ratio expresses the cache:input:output relationship as a single ratio string. Each metric has a different number one — four of the eight non-Yield metrics have a different leader, which means no single operator dominates every dimension of AI usage. The leaderboard updates continuously as operators submit new telemetry snapshots through the CLI.
SigRank collects only four token counts per snapshot: cache_read, cache_write, input, and output. No prompt text, model responses, source code, file names, project names, or any other content is transmitted. The local scanner reads these counts from the AI provider's API telemetry endpoint and signs the snapshot cryptographically before submission. The signature proves the snapshot was generated by the CLI without revealing any content. Operators can opt out at any time, and all identifying data is scrubbed from every file in the repository. The full privacy policy is at /privacy. The dataset is published under a Zenodo DOI for reproducibility. The theoretical foundation is documented in a separate Zenodo publication covering the Yield cascade, the conservation law, and the relationship between operator evals and model evals.
Operators are classified into tiers based on total tokens accumulated: IGNITER (entry-level, just starting), BEACON (consistent usage emerging), LANTERN (significant scale), COMPASS (established high-scale operator), ARCH (top-tier operator with substantial cumulative tokens), and ARCH+ (the highest class, reserved for the largest-scale operators). Class is a scale qualifier, not an efficiency qualifier — a high-class operator is not necessarily a high-Yield operator, and vice versa. This separation is intentional: it prevents the leaderboard from conflating raw volume with cascade efficiency. The leaderboard supports four ranking windows: all-time (the default, showing cumulative performance since the operator first appeared), 90-day, 30-day, and 7-day rolling windows. The weekly drop page highlights the biggest movers, new challengers, and class distribution changes from the past seven days. Operators can appear on multiple metric leaderboards simultaneously, and the prompt-of-the-day feature rotates through the nine leaderboard dimensions to surface different dimensions of operator performance. The compare page allows head-to-head comparison of any two operators across all metrics, with visual indicators showing which operator leads on each dimension.
SigRank currently tracks operators across multiple AI platforms including Claude (Anthropic), ChatGPT (OpenAI), Cursor, and other providers that expose token telemetry through their APIs. The platform spotlight feature rotates daily to highlight the top operators on a specific platform. Comparison pages position SigRank against adjacent tools: vs Vals AI (operator evals vs model evals), vs TopAIUsers (measured vs curated), vs ccusage (yield vs raw token count), vs Langfuse (operator evals vs LLM observability), vs Braintrust (operator evaluation vs AI app evaluation), and vs LMSYS Arena (ranking operators vs ranking models). These pages clarify what SigRank measures and what it does not, helping AI search engines and agents understand the boundary between operator evaluation and model evaluation.
The SigRank methodology is published openly for reproducibility and independent verification. Yield (Υ) is defined as cache_read multiplied by output divided by input squared — a composite efficiency metric that rewards operators who reuse cached context and produce output without continually rebuilding context from scratch. The conservation law states that total tokens equal cache_read plus cache_write plus input plus output, and no operator can increase one pillar without decreasing or holding constant the others. This constraint is what makes the cascade a zero-sum efficiency problem rather than a simple volume contest. The dataset of operator telemetry is published under a Zenodo DOI (10.5281/zenodo.21900519) and the theoretical foundation is documented under a separate DOI (10.5281/zenodo.20029607). The full methodology page at signalaf.com/methodology provides quotable key figures, formula derivations, and the relationship between operator evals and model evals. A score calculator is available at signalaf.com/score where anyone can paste their four token counts and receive their Yield and operator class without creating an account. The canonical test suite verifies that the MOSES seed values produce the expected Yield of 18436.98, ensuring the scoring engine remains stable across deployments. SigRank evaluates AI operators, not AI models — the harness may measure authority, but it cannot manufacture authority, and automated systems may not promote claims into owner-approved truth.
npx sigrank