BENCHMARK PERFORMANCE
Performance by product and language.
See held-out benchmark results for each analyzer with published evidence, and supported language coverage for the rest - pending analyzers are marked rather than attributed results. Every published result is tied to a documented evaluation dataset and detection configuration.
How we measure ranking performance
Top-3 Hit Rate
The percentage of evaluation cases where the expected indicator appears anywhere among the three highest-ranked results.
Per-language evaluation
Each supported language has its own held-out evaluation set, reported independently - never averaged or pooled. The full language-level matrix is available to customers and evaluators after sign-in.
Example: If the expected indicator appears in the top three for 92 of 100 cases, Top-3 Hit Rate = 92%.
This metric measures ranking performance. It does not measure false-positive rate, confidence calibration, or whether a downstream policy decision is correct.
Top-3 hit rate by analyzer
Per-analyzer ranges and medians across all evaluated languages. The complete language matrix is available to customers and evaluators - sign in to view it.
| Analyzer | Languages | Median Top-3 | Mean Top-3 | Range | Samples |
|---|---|---|---|---|---|
| Trust & Safety | 16 | 98.0% | 97.9% | 95.7–100.0% | 1,544 |
| Cybersecurity | 15 | 89.1% | 90.7% | 85.3–96.6% | 2,237 |
| Discrimination Analysis | 16 | 92.4% | 92.3% | 88.0–98.7% | 2,399 |
| Dating Safety pending | - | - | - | - | - |
| Fraud & Scam Detection pending | - | - | - | - | - |
| Crisis Intelligence | 21 | 91.3% | 91.2% | 86.7–96.7% | 3,117 |
| Disinformation Analysis | 16 | 91.3% | 91.8% | 87.3–100.0% | 2,400 |
| Market Intelligence | 8 | 88.3% | 88.7% | 85.4–93.2% | 1,081 |
| Insurance Intelligence pending | - | - | - | - | - |
| Emotion Analysis | 16 | 93.7% | 92.8% | 87.3–95.2% | 1,008 |
| Customer Service Intelligence | 18 | 91.7% | 91.2% | 87.3–96.7% | 2,700 |
| Wellness Intelligence pending | - | - | - | - | - |
Same metric and dataset version for every analyzer. Per-language detail and evaluation timestamps are available after sign-in.
Per-analyzer detail
16 languages · 210 indicators
Representative languages shown. The complete 16-language matrix is available to customers and evaluators after sign-in.
| Language | Top-3 Hit Rate |
|---|---|
| English en bench-ts-2026-09-en | 98.0% |
| German de bench-ts-2026-09-de | 98.0% |
| French fr bench-ts-2026-09-fr | 98.0% |
sqcfg_ZZW77NM9
·
Evaluation dataset: generated-natural
·
Evidence: bench-ts-2026-09
How the benchmark was built
These results are from internal Sentiqra testing. For each published benchmark, the following information is documented:
Evaluation set
Number of examples overall and per language. Each language has a documented case count.
Gold labels
How expected indicators were selected and reviewed for each evaluation case.
Language sourcing
Whether examples were native-written, translated, synthetic, licensed, or drawn from another documented source.
Dataset separation
Development, validation, and evaluation sets are kept distinct. Development data does not appear in the evaluation set.
Versions
The exact versioned configuration used for each evaluation.
Negative controls
Examples designed not to match target indicators are included when measuring false positives.
Evaluation date
Each benchmark has a documented evaluation date. Results are tied to the detection configuration active at that time.
More than one metric matters.
Top-3 Hit Rate answers: Did the expected indicator appear near the top? It does not answer every production question.
False positives
How often does the analyzer produce high-confidence matches on neutral or difficult negative examples? Negative-example evaluation sets are being expanded.
Score consistency
Do confidence values behave consistently at different score ranges? Consistency metrics will be added to future reports.
Per-group performance
Are results evenly distributed across the taxonomy, or are some groups materially stronger than others? Per-group breakdowns are planned for future publications.
These results measure ranking performance on Sentiqra's internal evaluation set. They do not measure your policy accuracy, downstream decision quality, or performance on your production distribution.
Evaluate Sentiqra on your own labeled data before setting production thresholds.
Test on your own distribution.
Internal benchmarks tell you how Sentiqra performs on our evaluation data. Your production content may differ. Use the free tier to build a labeled evaluation set from your own workload and measure Top-3, false positives, false negatives, threshold behavior, and language-specific performance.
5,000 analyzer runs/month · No card required