Sensitive-character recall on the named Nemotron evaluation sample.
Early Access evidence
Coverage, latency
and limitations.
Reviewed benchmark results for a named candidate and corpus—not a guarantee for every prompt.
Phone
Sensitive-character recall on the named Nemotron evaluation sample.
IP address
Sensitive-character recall on the named Nemotron evaluation sample.
Contextual entity coverage
Person detection
Sensitive-character recall for PERSON on the same independent sample. Contextual-model results are reported separately from deterministic structured patterns.
Coverage by entity
Sensitive-character recall
EMAIL100.00%
PHONE100.00%
IP ADDRESS97.62%
PERSON88.97%
LOCATION61.92%
CREDIT CARD8.57%
ORGANIZATION6.50%
Overall benchmark and residual risk
Complete results remain visible
The overall values include strong and weak categories. They must not be inferred from the headline structured results alone.
What residual risk looks like
Preserved context is intentional. A missed entity is exposure.
These synthetic examples illustrate failure classes observed in the reviewed benchmark. They are not predictions that the exact sentence will always succeed or fail.
Original Ramesh moved from Madurai to Austin.
Protected Ramesh may be identified as PERSON.
Could remain visible Madurai or Austin when a LOCATION is missed.
“moved from” and “to” remain visible by design because they are ordinary connecting context.Original Carolina joined Hartford & Co. last spring.
Protected Carolina may be identified as PERSON.
Could remain visible Hartford & Co. when ORGANIZATION detection misses.
“joined” and “last spring” remain visible by design as useful sentence context.Latency definition
p50 2.836 ms · p95 3.438 ms · p99 4.177 ms
This measures warm, in-process detection of short rotating samples. It excludes HTTP, gateway, tokenizer, network, LLM provider, cold-start, and long-document latency. It is not an end-to-end SLA.
Targeted regression suite
97.17% F1 across 66 cases
This small positional suite checks known behaviors. It is not overall product accuracy and is not interchangeable with the independent coverage result.
Unmeasured categories
Implemented, independent coverage not yet measured
Age, religion, ethnicity, health/disability, biometric IDs, bank accounts, routing numbers, IBAN, CVV, PIN, passcodes, and verification codes remain experimental until a representative holdout is reviewed.
Marai reduces sensitive-data exposure; it does not claim to be leak-proof. Results depend on the corpus, labels, model, rules, threshold, and input distribution. Snapshot: 2026-08-11-reviewed-v1.