Compliance
Redaction policy
What gets removed, and when. Strictness is configuration rather than hardcoded behaviour, because the rule driving it is internal policy — not a PDPA requirement.
Worth knowing.Singapore's PDPA imposes no data-localisation requirement, and only NRIC is genuinely restricted by statute — names and addresses are not. The stricter internal rule is honoured here by default, but it is expressed as settings so it can be relaxed without re-engineering if legal reviews it.
When redaction happens
Identifier formats
Singapore NRIC (S1234567A) and Malaysian MyKad (YYMMDD-PB-###G) share nothing structurally. A market left off here is silently not detected.
Spoken digits are matched in English, Malay, Mandarin and Tamil regardless of market, since callers code-switch freely.
What gets removed
Retention
Live tester
Paste anything. Runs the production engine under the policy above.
Agent: Can I have the last four of your NRIC? Customer: [NRIC]. My mobile is [PHONE] and I live at [ADDRESS], [ADDRESS].
Known coverage gaps
Stated openly. A redaction claim is only credible if its limits are on the same screen.
- NAME: only detected after an explicit cue ("my name is X"). Third-party mentions and bare names are missed — needs a NER pass.
- ADDRESS: tuned to Singapore blk/unit formats. Malaysian and free-form addresses are largely missed.
- Spoken digits cover English, Malay, Mandarin and Tamil, but only as contiguous runs — an identifier broken up by filler words ("satu… erm… dua") may slip through.
- ASR errors can split or mangle a number so no pattern matches. Redaction cannot recover what transcription mangled — this is the largest real-world risk on 8kHz telephony audio.
- Bare 12-digit runs are treated as possible MyKad and redacted; long account or order references will therefore sometimes be over-redacted.
The multilingual gap is material for Singapore, where code-switching is routine. An NRIC dictated as “yi er san si wu liu qi” currently survives redaction. This must be closed before live data.