AI metrics without a fake universal score.
Artificial Index evaluates dimensions independently. Performance metrics require benchmark or operational evidence. Structural metrics measure documented breadth, access or controls. Missing evidence remains Pending and is never converted to zero.
Headline dimensions
Intelligence
Benchmark-backed reasoning/knowledge performance. No benchmark evidence means Pending.
Capability breadth
Breadth of indexed capabilities, modalities, tasks and integrations. This is breadth, not model intelligence.
Speed
Latency, time-to-first-token and throughput evidence when measured.
Reliability
Uptime, error rate and structured/tool-call success when measured.
Cost efficiency
Relative token-price efficiency only where comparable USD/1M-token pricing exists. It is not performance-per-dollar.
Access breadth
Availability across API, web/app, integrations and local/self-host channels.
Deployment flexibility
Breadth of API/cloud/private/local/self-host deployment options supported by indexed evidence.
Privacy/security controls
Documented privacy/security control coverage, not an absolute security guarantee.
Openness
Open-weight/licensing accessibility. This is not a quality score.
Evidence confidence
Artificial Index verification depth, freshness and source coverage. It measures our confidence in the profile, not AI quality.
Grade scale
| Grade | Range | Interpretation |
|---|---|---|
| A+/A/A- | 85–100 | High score within that dimension and its stated methodology. |
| B range | 70–84.9 | Strong dimension result. |
| C range | 55–69.9 | Moderate dimension result. |
| D/E/F | 0–54.9 | Lower result within that metric; interpretation depends on the metric. |
| — | Pending | Insufficient comparable evidence. Pending is not zero. |
Important distinctions
Having more modalities or tools does not prove stronger reasoning. Intelligence stays Pending until compatible benchmark evidence exists.
R13 compares token price only where units are compatible. Successful-task economics require task success data later.
The metric measures documented control coverage, not a promise that one vendor is “secure.”
It measures how strongly Artificial Index can support the profile’s claims.
R14 performance engine
Every result is attached to the exact model, benchmark version, protocol, source and date.
TTFT, throughput, uptime, error rate, structured-output success and task economics have dedicated protocols.
Cross-benchmark comparisons use within-benchmark percentiles only after at least five compatible peers exist.
A model-family result does not automatically become the current product score.