Detection continuum · Rung 1 of 5
1 · Screening
“Which items need a closer look?” Platforms, schools and HR at volume.
- Evidence standard
- A known false positive rate, a confidence threshold and an appeal path.
- Cost of a wrong call
- A wrongly flagged person loses a grade, a job lead or an account.
What it can and cannot claim
Can claim
- This item scored above a stated threshold on a screen with a stated false positive rate.
- Items below the threshold were not reviewed further, and that is a policy choice, not a clean bill.
Cannot claim
- That any single flagged item is AI-generated. A screen ranks attention; it does not decide.
- A false positive rate measured on a different population, language or content type.
Error rates
Silvector publishes no measured detection error rate for this rung yet. Any number you see elsewhere on this site is labeled with its source, and none of them are ours. The identity fusion module refuses to show a confidence without a calibration report attached, but its corpus today is a seven-sample synthetic test fixture, so it supports no published rate. For scale, a third party's screening tool: OpenAI reported its 2023 text classifier caught 26% of AI-written text and flagged 9% of human-written text, then withdrew it for low accuracy.
A valid error-rate statement at this rung includes:
- False positive and false negative rates at the threshold in use, each with its denominator.
- The dataset, its date and whether it was LIVE, REPLAY or SIMULATED.
- An appeal path a flagged person can use.
Bias risks
- Base rates matter more than accuracy: at a 1% false positive rate, a screen of 10,000 human-written essays flags about 100 people who did nothing.
- Detector bias against non-native writers carries straight into screening queues.
What Silvector has at this rung
| Capability | State | Where |
|---|---|---|
| Calibration that attaches FAR, FRR and EER to any confidence (identity claims of agent traffic) | behind a flag | packages/identity/src/identity-fusion.ts, exp.rd3_identity_fusion |
| Agent catalog used to screen declared crawlers and agents (1,706 entries, refreshed 2026-10-29) | shipped | packages/detect/src/catalog-meta.ts |
How to escalate
To reach Rung 2 · Operational:
- Make a deterministic decision under a named policy and keep its signed receipt (rung 2).
Sources
- OpenAI, New AI classifier for indicating AI-written text (updated July 2023)
- Liang et al., Patterns 2023
Previous: Rung 0 · Intuition · Next: Rung 2 · Operational · The whole continuum