Provable honesty
Accuracy & test corpora
Every result badge across Everyday Tools Hub comes from a single typed registry and reflects how the calculation actually works — never marketing optimism.
Deterministic calculations verified directly against math, checksums, RFC parsers, open BPE tables (o200k_base / cl100k_base), or generated file byte/dimension headers.
Approximations where an exact proprietary specification (such as closed Anthropic or Google tokenizers) is not bundled client-side. Ranges show expected variance.
Pattern-based inference (such as PDF near-white fills, truncation detection, or optional prompt-instruction phrase flags) that can have false positives or bypasses.
Safe-to-Paste Redactor — Synthetic Corpus Benchmark
Detector set version 2026.11.1 • Computed directly from the shared production detector engine across 6 synthetic cases with decoys.
| Detector ID | True Positives | False Positives | False Negatives | Precision | Recall |
|---|---|---|---|---|---|
| 1 | 0 | 0 | 100.0% | 100.0% | |
| phone | 1 | 0 | 0 | 100.0% | 100.0% |
| credit_card | 1 | 0 | 0 | 100.0% | 100.0% |
| iban | 1 | 0 | 0 | 100.0% | 100.0% |
| credential_url | 1 | 0 | 0 | 100.0% | 100.0% |
| api_key_token | 1 | 0 | 1 | 100.0% | 50.0% |
| ipv4 | 1 | 0 | 0 | 100.0% | 100.0% |
| in_pan | 1 | 0 | 0 | 100.0% | 100.0% |
| us_ssn | 1 | 0 | 0 | 100.0% | 100.0% |
Note: Synthetic benchmark scores measure known pattern rules (email, Luhn cards, mod-97 IBANs, API prefixes, IPs, SSN/PAN, phones). Unstructured human names and street addresses are intentionally omitted from automated claims until an audited on-device NER gate passes.
Why We Deliberately Do Not Build an AI-Text Detector
Published studies demonstrate high false-positive rates against non-native English writers on automated “AI text detectors.” Because accuracy cannot be defended responsibly and false accusations cause real harm to students and applicants, we will not ship an AI-writing detector.