# The Orthogonality of Economic Reputation: A Sealed Two-Axis Measurement on the Live x402 Settlement Graph

**Andy Salvo, Jameson Ackerman. Crest Deployment Systems LLC. July 2026.**

Preregistered before results and sealed to a public hash chain (witness-chain generations G064 preregistration, G065 results, G066 adversarial validation). Run on the live x402 settlement graph (1,770,483 on-chain events, USDC on Base).

## Claim

Agent trust has three axes that do not substitute for each other: identity (who an agent says it is; ERC-8004 IdentityRegistry and endorsement-based reputation), security (whether its execution is verifiable; ERC-8126's ZK-proven checks), and economic reputation (whether it has actually been paid by real, standing participants). This note measures two things about the third axis on live data: that it is necessary (the other axes do not find the real economic actors) and that it is hard to fake (a funded collusion ring cannot buy a top position).

## Result 1: identity and settlement are near-disjoint populations (the axis is necessary)

A clean 2x2 on the raw ledger, no model, no tunable definition:

| | count |
|---|---|
| Real earners (>= $1,000 inbound from >= 5 distinct payers) | **743** |
| Registered identities (declared services with on-chain flow) | 125 |
| Addresses in both | **1** |
| Real earners missing from the identity registry | 742 (**99.87%**) |
| Registry entries that are not real earners | 124 (**99.2%**) |

An identity check misses 99.87% of the agents that actually earn, and 99.2% of registered identities have never been a real economic participant. ERC-8126, by its own stated scope, does not measure economic reputation (its wallet-verification check confirms a wallet has history, not that it fulfilled commitments). On the live graph, neither identity nor security identifies the real economic actors. Only performed settlement does.

## Result 2: a collusion ring cannot buy a top rank (the axis is hard to fake)

The production ranking is continuously stress-tested by an attack/defense co-evolution loop. To rule out attacker fatigue as the reason the defense holds, we replaced the loop's proposer with a strong structured adversary: a tiered cost ladder crossed with the strongest known evasion structures (multi-hop laundering, recirculating mesh, dead-end decoys, cohort overlap, disjoint partitions), 80 attacks in total, at real-dollar budgets of $3, $15, $37, $150, and $750. Cost is counted honestly as unrecyclable anchor-facing spend.

- **No attack reached top-1, top-5, or top-10 at any budget up to $750.**
- The best rank a ring could buy was **#72, for $15**. Raising the budget to $750 produced a *worse* best rank (#98), because the defense keys on ring structure, not dollars.
- **Legit fidelity = 1.0**: the defense leaves the real top-20 service ranking unchanged.

## Addendum: continued testing corrected this result (post-seal)

The sealed sweep above reported no top-10 break for budgets up to $750. The continuous integrity loop later **falsified that floor**: it found a $600 collusion break (rank #6) using a width-1 conduit chain that split its final hop exactly in two, threading precisely under both the funnel test (which required an out-share strictly above 0.5) and the dead-end test (which required at least two diluting predecessors). We report this rather than bury it — a sealed number that later testing corrects is the process working, not failing. A defender turn added a sink-diluted-conduit rule keyed on pass-through nodes (a predecessor fed almost entirely by a single upstream and carrying no standing of its own). The rebuilt defense was then re-swept **including that exact attack family**, at budgets up to $3,000: no top-10 rank is purchasable at any budget, the best a ring achieves is rank #17, and the legitimate ranking is unchanged (fidelity 1.0). The lesson stands as method: published bounds are provisional, the loop is adversarial and continuous, and corrections are published with the same prominence as the original claim.

## Honest scope

- "Unfakeable" here means a collusion ring cannot cheaply buy a TOP rank. A ghost can still crawl to a low rank (~#72). The claim is bounded to the tested collusion/ring/sybil attack family and the graph snapshot at test time; the co-evolution loop re-tests continuously.
- This is not a claim about buying real settlement from real standing payers. That is earning, not faking, and it is the behavior the index is designed to reward.
- Coverage figures are measured against the census's own service classification and drift as the graph grows, which is why they are served live rather than frozen here. Verified-answer coverage is near-total by construction and should not be read as a strength claim; anchor-rank coverage is the honest measure of how much of the graph the sybil-resistant rank currently reaches.
- An earlier internal red-team found a ~$37 break against a prior defense generation. That defense was replaced; the break is closed and the closure is what G066 validates.

## Result 3: coverage (with an honest split)

Necessity is not enough; we also measured coverage over every address the census classifies as a real x402 service. Two different questions have two different, honest answers, both served live and recomputed each refresh at `https://api.agentrank.info/coverage`:

- **Verified-answer coverage (~99.94%):** AgentRank returns a settlement-verified answer for essentially every real service. This is near-total *by construction* — a service is defined by having settlement, and settlement is exactly what AgentRank verifies. It says the index has an answer for who you ask about; it is not a strong claim on its own.
- **Anchor-rank coverage (~3.7%, the honest limitation):** only a minority of services currently sit in an anchor-weighted, sybil-resistant *rank position*. Most are not yet connected through the payment graph to a verified anchor and receive a settlement-magnitude score rather than a flow-weighted rank. Widening this (more verified anchors, denser settlement) is ongoing work, and we publish the current figure rather than hide it.

The strong, non-tautological result remains the disjointness: an identity check misses **99.87%** of the agents that really earn. (An earlier recall figure was confounded by measuring against a dead declared registry; separating verified-answer coverage from anchor-rank coverage resolves both the confound and the risk of overclaiming.)

## Why this matters

Our companion study, Counterfeit Verifiability in Autonomous Agent Payments (DOI 10.5281/zenodo.21042364), showed that agents choose a counterparty that merely displays the surface of trust 99% of the time, and that only performing the verification recovers the truth. This note establishes that the thing worth performing the check against, settlement-grounded economic reputation, is both the only axis that finds the real economic actors and one whose top positions a funded adversary could not purchase in our tests.

Query it before you pay: `GET https://api.agentrank.info/resolve/{wallet-or-domain}`, MCP at `https://api.agentrank.info/mcp`, A2A at `https://api.agentrank.info/a2a`. Free.

## Provenance

Preregistration (frozen thresholds, hypotheses, and falsifiers written before any data) and results are sealed as generations G064, G065, and G066 of the Crest witness chain (SHA-256 hash-chained, Bitcoin-anchored via OpenTimestamps). The defense's internal parameters are not published; the verdicts, attack families, cost ladder, and full honest scope are.