The SigRank Index

Key Figures

What “token-cascade efficiency” means

Yield (Υ) = cache_read × output / input². It measures the architecture of an operator’s token cascade — whether signal is compounding (high cache reuse × high output per fresh input) or tokens are burned (low cache, low output). Volume is noise; yield is signal.

The 10 build archetypes

Every operator’s cascade falls into one of 10 build archetypes, classified by leverage, velocity, and construction. Archetypes are distinct from class (scale/qualification) and rank (a leaderboard position by Yield).

• CONVERGENT (Convergence) — Deep reuse, active construction, and high generation rise together. A rare composition where all three operating axes are elevated without the usual tradeoffs. Defined by: P80 on all 3 axes (leverage + velocity + construction).
• KINETIC (Generation) — Generation has broken out. Output approaches or exceeds fresh input, making transmission the defining feature of the composition. Defined by: velocity >= 0.80.
• INPUT-BOUND (Reuse Depth) — Fresh input still carries most of the workload. Little prior context is returning, so each cycle depends heavily on new input. Defined by: leverage < 5.
• PRIMING (Reuse Depth) — Reuse is beginning to form. Prior context is returning, but the system has not yet developed deep leverage. Defined by: leverage 5–10.
• CONTEXTUAL (Reuse Depth) — Retained context is now materially supporting the workflow. Reuse is established, while active construction remains limited. Defined by: leverage 10–15, passive.
• DEEP READER (Reuse Depth) — Strong accumulated context is carrying the workflow. The operator draws deeply from retained context while creating relatively little new context. Defined by: leverage 15–23, passive.
• ARCHIVIST (Reuse Depth) — Extreme reuse of accumulated context. A deep context library carries the system while new construction remains limited. Defined by: leverage >= 23, passive.
• BUILDER (Active Construction) — Active context construction has begun. The system is creating material for future reuse while leverage is still developing. Defined by: construction >= 0.02, leverage < 30.
• RECURSIVE (Active Construction) — New context is being built on top of an already substantial reusable base. Construction and reuse are now feeding the same operating loop. Defined by: construction >= 0.02, leverage 30–50.
• AMPLIFIER (Active Construction) — Deep reuse and active construction are operating together at scale. Existing context produces new work that expands the context available for future cycles. Defined by: construction >= 0.02, leverage >= 50.

Methodology

• Inputs: on-device token counts (fresh input, output, cache_read, cache_write) per session per platform.
• Verification: each snapshot is ed25519-signed and verified server-side; replay and plausibility guards apply.
• Windows: operators are ranked over 7-day, 30-day, 90-day, and all-time cohorts.
• Privacy: token counts only — message content is never transmitted, read, or stored.
• Scoring: the yield metric Υ is computed from the four token pillars via a cascade model. The composite SIGNA rate blends Υ with signal-force and drift components. Proprietary weights (RS.xx) govern the composite and remain server-side.

Derived metrics

Yield is the headline metric, but the cascade produces four derived metrics that describe different aspects of operator behavior:

Normalization

Raw pillars are made comparable through ratios rather than direct token totals. The cascade clamps the input denominator to at least one to avoid division by zero. Volume is separately represented as log10(total pillars) rather than folded into every ratio. The scoring engine normalizes selected Core 5 inputs to [0,100]: compression is multiplied by 100, throughput uses min(100, 20 × log10(x + 1)), and session depth uses a server-side bucket table.

Normalization improves comparability but cannot remove differences in tools, models, workloads, windows, or reporting practices.

Limitations

Upsilon measures a token-flow relationship: cache reads and output relative to fresh input. It does not measure correctness, novelty, user satisfaction, economic value, code quality, safety, talent, effort, or intelligence.

Yield is especially sensitive to small input denominators. The app uses a denominator floor and ingest plausibility checks, but neither turns a ratio into proof. Cache behavior also depends on provider, model, prompt structure, session design, and collection tooling.

Scores and classes are conditional on submitted data, verification tier, scoring version, and selection into the field. Missing telemetry, unverified submissions, changing tools, and uneven sampling limit comparisons. Treat rank and signals as observations to investigate, not final judgments.

Gaming threat model

SigRank defends against leaderboard gaming through multiple layers:

• Cryptographic signing: every snapshot is ed25519-signed by the operator's device key. The server verifies the signature before accepting.
• Server-side rescoring: the server recomputes all derived metrics from the four raw token pillars. Operators cannot submit inflated yield or leverage values directly.
• Replay protection: nonce tracking blocks resubmission of identical snapshots.
• Plausibility gates: implausible token ratios are flagged for review before appearing on the public board.
• Benford's Law check: aggregate leading-digit analysis flags statistical anomalies in token counts. A Benford flag is evidence for review, not proof of fabrication — it is considered alongside data coverage, independence, and collection behavior.
• Proprietary weights: the composite SIGNA rate uses server-only weights (RS.xx) that are not published, preventing targeted optimization against the composite.

The leaderboard is discovery, not judgment. No single metric can fully prevent gaming — the goal is to make gaming harder and more detectable than honest participation.

Protocol authority

The canonical metric definitions and conformance suite are governed by the TTEOP specification (Token Telemetry Evaluation Operator Protocol, v0.1.5-draft). SigRank is the public reference implementation; Upsilon is the enterprise measurement engine. Both implement TTEOP. The governance framework (MO§ES™, patent pending) enforces submission integrity.

The scoring is grounded in published science — the Conservation Law of Commitment (DOI: 10.5281/zenodo.20029607).

How the data updates

The Index updates continuously as operators submit signed snapshots. Public top-N data is available at /api/v1/leaderboard.

How does SigRank verify submissions?

SigRank verifies submissions using ed25519 cryptographic signatures, server-side rescoring, plausibility gates, and replay controls. Each device generates a keypair during enrollment; every snapshot is signed and the server verifies the signature before accepting. The server recomputes all derived metrics from the four raw token pillars — operators cannot submit inflated yield or leverage values directly. Replay attacks are blocked by nonce tracking, and implausible token ratios are flagged for review.

License & citation

The SigRank Index dataset is licensed under Creative Commons Attribution 4.0 (CC-BY-4.0). Attribution is the citation mechanism — reuse requires credit, which turns reuse into citations. The source code is separately licensed under MIT.

Cite as: “SigRank Index, September 2026. signalaf.com/methodology.”

FAQ

What is the SigRank Index?

A continuously-updated leaderboard that ranks AI operators by token-cascade efficiency (Υ = cache_read × output / input²), computed from privacy-preserving, on-device, cryptographically-signed token-telemetry snapshots.

How is operator efficiency measured?

Each operator runs an on-device scanner that reads four token pillars locally. The yield metric Υ measures the architecture of the token cascade — whether signal is compounding or tokens are burned. No message content is ever read or transmitted.

Is the data private?

Yes. The scanner reads token counts and content lengths only — never the words of your prompts. Only the resulting numeric scores, signed with ed25519, leave your device.

How do I get listed?

Install the SigRank MCP server (npm i -g sigrank), enroll, and submit a snapshot. Visit /score to get started.