Agentic Resource Radar

Governance

The rules that keep the measurements honest, published before anyone asks. No paid placement, ever. Versioned public methodology. Hidden held-out test splits. Anonymous accounts paying standard rates. Judges audited against humans with the numbers published. Public corrections. These commitments bind us from day one, while the rankings are still small enough that nobody would bother attacking them. That is the point.

Independence and No Paid Placement

  • No provider can pay for results, methodology changes, placement, or listing. Ever. Inclusion is by published selection criteria only.
  • Rankings and scores are never influenced by any commercial relationship. If ARR ever offers vendors paid services (private benchmarking, monitoring), those never touch public scores, and any vendor with such a relationship is disclosed on its provider page.
  • We do not publish usage-based or popularity-based rankings. Usage signals are gameable; measurement is not. Our founding story is having watched a usage leaderboard get gamed.
  • Telemetry tables order alphabetically by default and sort only on measured columns. Ranked leaderboards rank only on the measured Agentic Search Index, with confidence intervals and rank ranges displayed.

What Providers Cannot See or Train On

  • The scored task splits are private and rotate every release. The freshness generator and calibration labels are private.
  • Benchmark accounts are registered off-domain under neutral identities and pay standard self-serve rates, so providers cannot identify benchmark traffic to special-case it (mystery-shopper discipline).
  • Provider-visible artifacts (heads-up scorecards) are sent shortly before publication, never with an option to alter results.

Judges Are Audited Instruments

  • Automated judges never operate uncalibrated. Before a release, dual human labeling establishes the human-human agreement ceiling, and the judge's agreement (Cohen's kappa) and false-positive rate against those labels are published with every release.
  • Judge model and prompt versions are pinned per release. Changing a judge is a logged, re-calibrated event. Judge changes that increase false positives are rolled back.

Money and Measurement Are Separated

  • The continuous monitor never pays, never authenticates, never sends payment-like headers. It observes public surfaces only, politely (per-host limits, backoff on any 429/403).
  • Paid probing exists only inside the evaluation harness, on eval cadence, with spend logged. Buying is separate from watching.

Corrections, Disputes, and Versioning

  • Every release is versioned with a public changelog. Our own errors get logged in the same changelog. Corrections are public, never silent.
  • Disputes: any provider can challenge a result. We respond publicly and itemized. Methodology critiques are engaged on the merits; results are never negotiated privately.
  • Score-affecting methodology changes ship in versioned releases with rationale, never retroactively.

Contact for disputes: contact@agenticresourceradar.com.

Provenance Disclosures

  • The judge methodology builds on the published Universal Verifier work (Microsoft Research and Browserbase, arXiv:2604.06240). Because Browserbase operates in a category ARR may later benchmark (browser infrastructure), that category will launch only after our judge-calibration numbers are public, with this provenance disclosed on the category's methodology page.
  • Task-bank seeds and their licenses are disclosed in the methodology: SimpleQA-Verified (MIT), FRAMES (Apache-2.0), WebWalkerQA (Apache-2.0).

Probing Ethics

  • Unauthenticated telemetry respects robots-level politeness: never more than one in-flight request per host, exponential backoff, a browser-normal user agent, and no bypassing of bot-walls. A block is recorded as blocked, not circumvented.
  • Paid benchmark calls are ordinary customer traffic at list prices.
  • We do not exploit misconfigured or unintentionally open endpoints. Anything that looks unintentionally free is tested through normal paid or keyed access instead, and the observation is noted.

Conflicts

Agentic Resource Radar operates its own paid API, which appears in the x402 lane like any other monitored service. Its rows receive no special treatment: the same probes, the same columns, the same ordering rules.

What Happened to Reviews and Claims

Earlier versions of this site carried wallet-verified reviews and a provider claim flow. That functionality is removed. Existing review and claim data is frozen in storage, not deleted, and no longer renders or affects anything. The site now shows our independent measurements only.