Early access — cascade metrics are real (derived from canonical token telemetry); the operator field is a curated seed. Learn more about the data

The evaluation platform for AI operators

Models are benchmarked constantly. The people operating them are not.

§
powered by MO§ES™
Token counts only. Never your prompts.

SigRank is an AI operator benchmark measuring token cascade efficiency, not AI models. It turns privacy-preserving token telemetry into a repeatable performance evaluation: your Yield, workflow signature, benchmark, and progress over time. The leaderboard is proof, not the product. The product is the operator-evaluation standard.

Fair warning: the blade cuts both ways.

Measure how you operate AI. Token counts only — never your prompts. Or paste your numbers →

Identifying Burners, Builders, and 10×ers.

signal AF

1,498operators ranked
18.2Qtokens analyzed
17platforms tracked
3,304models measured
1.68median Yield (Υ)

Last updated: August 14, 2026

SigRank is the evaluation platform for AI operators — the humans using AI tools, not the AI models themselves. It ranks operators by Yield (Υ = cache_read × output / input²), a composite 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 that rank AI models by human voting, SigRank measures the human factor — how efficiently each operator uses AI capabilities. The platform supports 17 AI tools including Claude Code, ChatGPT, Cursor, Copilot, Windsurf, and Codex. Operators run a local scanner 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 7-day, 30-day, 90-day, and all-time windows, with operator classes from IGNITER through ARCH+ based on total tokens accumulated. TRANSMITTER is a separate peak-activity badge, not a tier. The public REST API, OpenAPI specification, MCP server, and CLI tool are documented at /developers. Pricing is free during the build stage at /pricing.

The dataset spans 1,498 human operators across 17 platforms and 3,304 models, with 18.2Q tokens analyzed. The median Yield is 1.68, and the top operator achieves a Yield of 88. The full methodology, metric definitions, and evidence boundary are documented at /methodology. The anonymized research dataset is available on Zenodo (DOI: 10.5281/zenodo.21900519) under a CC-BY-4.0 license. The theoretical foundation, the Conservation Law of Commitment, is published separately (Zenodo: 10.5281/zenodo.20029607). Operators can check their efficiency without installing anything using the score calculator, or run the full local scanner with npx sigrank to submit a signed snapshot to the leaderboard. The CLI also runs a local MCP server exposing leaderboard, operator profile, and rank-paste tools to AI agents via the Model Context Protocol.

⊙ The four degrees of leverage
MetricAA baseline§Human Center of Mass*Power users†Top Evals to date‡
Υ Yield1.751.90239.61941.23
SNR0.330.090.510.63
Velocity (O/I)0.500.101.041.70
Leverage (CR/I)3.5×20.3×241.6×553.2×
10xDEV (log₁₀)0.541.312.382.74
Efficiency (vs AA 4.0)1.005.2562.83141.90
Operating Ratio (C:I:O)3.5 : 1 : 0.5020 : 1 : 0.10242 : 1 : 1.04553 : 1 : 1.70

† Power users: median of the top 100 real operators by Υ Yield.

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. 130 outliers separated (see the field analysis). All derived from canonical four-pillar token telemetry. Token counts only. Read the full analysis.

The baseline builds. The field caches. A few compound.

Read it as a token cascade: Cache : Input : Output. The AA baseline at 3.5 : 1 : 0.50 is the modeled reference — the average AI user. The median operator on the board sits at 20 : 1 : 0.10 (the second column) — the typical operator, the 50th percentile of everyone measured.

The median of the top 100 operators lands at 242 : 1 : 1.04. What they give up in output they bank in cache — the typical elite performer.

The top operator on the live board sits at 553 : 1 : 1.70: every input token returns multiple outputs while carrying a deep cache. That's the eval to beat.

Full description, the 10xDEV log read & full provenance →
⊙ Live board

Real operators. Real cascades.

Full board →
Live activitycoming soon
1.7K
Operators ranked
Active in the last hour
328
Operators in TRANSMITTER class
132
Comparisons ran
18158.68T
Total tokens measured
⊙ Live telemetry — coming soon

The leaderboard is proof the measurement works — scored live by the same engine that scores you.

See where your operating efficiency actually stands.npx sigrank→ scan, submit, measure
⊙ How it works

Three commands. That's it.

The SigRank agent reads your local AI session logs on-device, derives your token cascade, and publishes to the board. No paste, no prompts read — only the four token counts leave your machine. ccusage, tokscale, and tokendash are bundled — no separate installs.

Step 1

Install

npm install -g sigrank

Pulls the agent + ccusage + tokscale + tokendash in one install. Node ≥18, macOS + Linux.

Step 2

Sign in

sigrank enroll

Paste a connect code from signalaf.com → Settings → "New key". Binds your device to your operator identity.

Step 3

Submit to the board

sigrank submit

Signs + publishes your cascade. Your rank updates live on signalaf.com.

Or let your AI agent do it

Don't want to leave your agent? Just tell it to run npx sigrank to see your cascade, or npx sigrank submit to publish (sign in once first with npx sigrank enroll). It reads your logs, derives the cascade, and submits — you don't paste anything. For direct tool calls, wire it as an MCP server — see the local agent wiki page.

⊙ The IP boundary

What we open. What we keep.

Trust requires transparency. A leaderboard requires a moat. We open everything that proves the system is honest — and keep everything that prevents it from being gamed or cloned.

Open source

What we publish

Everything that makes the privacy claim verifiable. Read the code. Inspect the schema. Trust by audit, not promise.

  • The Υ Yield formula(cache_read · output) / input² — the rank metric, in the open
  • Local agent source codeEvery line auditable · See exactly what gets sent
  • Snapshot payload schemaThe exact four token pillars that cross the network
  • Cascade metric definitionsSNR, Leverage, Velocity, 10xDEV — every formula published
  • Class taxonomy and tier names8 experience tiers from ARCH+ down to IGNITER + TRANSMITTER badge
  • Privacy guaranteesNo raw transcripts · signed snapshots · verifiable counts
  • Adapter SDK + public REST APIBuild integrations · Read boards, profiles, snapshots
  • Ruleset version historyEvery change documented and replayable
Proprietary

What we keep

The moat. The math that prevents gaming and cloning. Without these, anyone could rebuild a clone with our data and undercut the leaderboard.

  • Class threshold breakpointsThe exact SNR / 10xDEV cuts. Why "rare" stays rare.
  • Species classifier weightsThe velocity / leverage quadrant boundaries
  • Promotion stickiness rulesHow a class is held vs. demoted across windows
  • Recency modifier curvesHow live rankings decay with inactivity
  • Anti-gaming detectionPattern matching against spam, redundancy, synthetic inflation
  • Reader-robustness normalizationHolding rank stable across token readers (RS.xx)
  • The verification batteryThe deeper signal-integrity checks behind the audit tier
  • Corpus + MO§ES anchorThe verified seed that makes the field hard to clone
⊙ Tiers

Free for ranking. Back the build for what's next.

The leaderboard is free and stays free — your cascade metrics, your class, your rank, no paywall. If you find SigRank useful, consider a one-time contribution to support the build.

FREE

Operator

$0 · always free

  • Submit token telemetry, get ranked instantly
  • Full cascade layer — Υ Yield, SNR, Leverage, Velocity, 10xDEV
  • Cascade species + class assignment
  • Public leaderboard with platform filters
  • Head-to-head compare on the cascade metrics
  • Operator profile with the cascade fingerprint
  • MCP server for Claude Code, Cursor, and any MCP-compatible client
Start free
SUPPORT

Back the build

Pay what helps · one-time, no subscription

  • Support the ongoing build if you find SigRank useful
  • Help cover server costs and data infrastructure
  • Keep the corpus verified + independent
  • No subscriptions, no tiers, no paywalls
  • The free board stays free — this funds the build
Support the build

Ask AI about us

Copy these prompts into ChatGPT, Perplexity, or Google AI Overviews to see how AI search engines answer questions about SigRank.

⊙ Measure your operating efficiency

Four integers in, full evaluation out.

Architecture is the only variable that matters. Run the local agent or just paste your numbers — see your Υ Yield, your class, and where your operating efficiency actually stands in under a minute.

Building against SignalAF? Developer portal · OpenAPI · MCP

Learn more

Best AI Users · Benchmark Your AI Usage · AI Operator Evaluation · AI Benchmarking · AI Coding Metrics · AI Power User Statistics · Human vs Model Performance · Privacy-Preserving AI Measurement · Operator Performance · Cascade Analysis · Token Telemetry · Cache Hit Rate