Verging Labs

Methodology

One pinned reference agent runs every task against every provider identically; LLM judges grade the answers and are themselves audited against human labels before anything publishes. Nothing is blended into one number: every index is a pure quality construct, and cost and latency are separate published axes.

Principles

  • A task passes only if it addresses the question, is correct, is current where that applies, and is supported by what the tool actually returned. An unsupported or fabricated answer is a scored failure.
  • The scored task splits are private and rotate per release, so providers cannot train on the test.
  • Judges never run uncalibrated: labels from at least two humans set the agreement ceiling, and the judge's agreement against them is published with every release.
  • No provider can pay for results, methodology changes, placement, or listing. Paid engagements never touch public scores; the full rules live on the governance page.

Memory

The Agentic Memory Index applies the same measurement to agent memory tools. Eight memory tools measured on simulated multi-week company brain operations. This section summarizes the method; the full data lives on the index page.

Recall Retain Onboard Store Ready Store Ready Store Ready session 1 session 2 session 56
One simulated working relationship. The agent is onboarded once, then stores and recalls across 56 sessions, and facts stored early are asked for again dozens of sessions later.

Task Bank

Task typeWhat it tests
Direct RecallFacts stored once and asked for directly.
Updated FactsFacts that changed after storage; the newest value must win.
Thread GrowthThreads that grew across sessions; the accumulated state must come back.
SynthesisAnswers assembled from several stored memories.
Long TermFacts stored weeks of simulated time before the probe.
False memory checkProbes about things never stored; the honest answer is that nothing is known.

The bank holds 272 scored probes per tool: 200 positive probes and 72 never-stored honesty probes. The store every tool answers from is built through 42 simulated multi-week operator relationships plus two long arcs, and early facts are re-probed at five checkpoints across 56 working sessions.

Judging and Calibration

The verdict enum is shared with search, with two relabels: a fabricated memory is an assertion about something never stored, and a provider infrastructure fault voids the affected runs instead of scoring them.

Two independent humans labeled the same 66-answer calibration set and agreed 91% of the time before adjudication (Cohen's kappa 0.63); every disagreement traced to one rubric clause. The judge pipeline matched the primary labeler 98.5% of the time (kappa 0.95) across the full set.

Judge vs primary labeler
98.5%
kappa 0.95 · n=66
Human-human, pre-adjudication
91%
kappa 0.63 · n=66

Claude Code built-in memory runs the same bank as an unranked reference and scored 67.7.

Cost and Speed

Cost per 1,000 successful answers counts agent token spend to store and answer (CLI-metered at API-equivalent rates), the entry paid tier amortized at its capacity through measured usage, and embedding spend where billed separately.

Speed is end-to-end answer time with the same pinned agent model on every tool, published alongside one-time setup, indexing lag, and per-operation latency. The full charts, per-tool numbers, and the leaderboard live on the Agentic Memory Index page.

Versioning

  • Kind weights and the judge (model and prompt) are pinned per index version; any change ships as a new, recalibrated version.
  • Task splits are private and rotate per release; saturated sources are retired with published rationale.
  • Changes and corrections are logged in the changelog, never silent.

Telemetry Semantics

  • Liveness is not "returns 200": every watched surface carries an expected status set, and a 200 with the wrong content records as a soft-404, not presence.
  • Blocked is not down: bot walls are recorded as blocked and never counted as downtime.
  • Percentiles come from rollups only; a single-sample latency is never displayed. Speed and liveness are measured with unpaid probes from a single vantage point (our probe region). Treat them as our view, not global ground truth.
  • The watcher never pays and never authenticates: unpaid, unauthenticated requests only.

Lineage

The judge design adapts the Universal Verifier's published calibration principles (Rosset & González Fernández et al., arXiv:2604.06240); hurdle and signed grounding follow Mercor's ACE.