What broke, what changed and what is fixed now.

Critical PII was missing: SSNs, IBANs and CVV codes.

The original detector set covered emails, phones and payment card numbers but missed government-issued identity numbers, bank account IBANs and card security codes. Those are high-value targets for identity theft and financial fraud.

Fixed by adding three critical-severity detectors. The SSN detector excludes invalid area codes (000, 666 and 900-999). The IBAN detector runs a full mod-97 checksum to reject false positives. The CVV detector is keyword-triggered and requires a 3-4 digit number adjacent to CVV/CVC terminology.

Dates of birth and IP addresses were not flagged.

Issue #3 was previously assigned to an external contributor. Both detectors are now part of the core engine. Date of birth handles natural speech patterns ("born on March 15, 1988") and multiple date formats. IP address detection excludes the broadcast and zero addresses.

The tool had no way to detect emotionally vulnerable users or coercion.

A voice assistant may receive transcripts where the user is in emotional distress or is being pressured by a third party. Neither state was detectable and both require very different assistant responses than a straightforward PII redaction.

Fixed by adding two new intent categories. The emotional distress signal (high, 70% confidence) catches hopelessness, worthlessness and giving-up language. The coercion or pressure signal (high, 65% confidence) catches phrases that suggest the user is being forced to say or do something. Both trigger specific guidance to the assistant.

Detection had no confidence values. Every finding looked equally certain.

The original engine returned binary findings with no indication of how confident the detection was. A phone number match near a card number was treated with the same certainty as a Luhn-verified card. This made the risk score misleading and offered no way to tune sensitivity.

Fixed by adding confidence: float to every Finding. The risk score is now computed as severity_points × confidence, so a 70%-confidence high finding contributes 25 points rather than 36. The web app shows a CSS confidence bar and percentage on every finding card.

The web app had no way to see which words in the original transcript were flagged.

The existing "Protect" view showed a cleaned transcript, but there was no visual way to understand which specific words or phrases triggered each finding. Users had to read the finding descriptions and mentally match them back to the original text.

Fixed by adding an Inspect tab to the Check panel. It renders the original transcript with inline <mark> highlights: amber for medium, red for high, brighter red for critical. Each mark is hoverable and shows the finding label, confidence and explanation in a native tooltip.

Analyzing a multi-turn conversation required calling the API once per turn with no aggregation.

Real voice assistant interactions span multiple turns. The old API had no way to analyze a sequence of utterances together or understand the cumulative risk across a session.

Fixed by adding analyze_conversation(utterances) to the Python API and a Conversation mode toggle to the web app. Each turn is analyzed independently. The result includes a peak decision, cumulative score and the union of all entity kinds seen across the session. A conversation timeline view shows all turns with their individual findings.

There was no way to copy or export the analysis result.

The safe transcript could only be read on screen. Developers building voice pipelines had no way to programmatically access the full analysis output from the web app.

Fixed by adding a Copy button (clipboard, Cmd/Ctrl+Enter shortcut) and an Export JSON button that downloads a structured file with the decision, score, findings with confidence, redaction map and guidance.

Saved exports could keep sensitive values in plain JSON.

VoiceSafeKit intentionally keeps the transcript readable in memory while filtering because the detector rules need plain text to find emails, SSNs, IBANs, passwords and safety signals. The saved JSON file was different. Once it landed on disk, it could keep the redaction map and other sensitive details readable.

Fixed by encrypting saved analysis files with AES-256-GCM and a key made from the export passphrase. Browser export now uses the Web Crypto API and a local export key. The CLI --out path encrypts by default with VOICESAFEKIT_EXPORT_KEY. There is also voicesafekit decrypt for opening an encrypted export again. TLS/HTTPS still protects network traffic. It does not encrypt local files on disk.

The encryption path had to keep the local-first design intact.

The detector still works on readable transcript text in memory, because that is the only way to find and redact sensitive values. The storage boundary is where the new protection starts: before a result is downloaded or written to disk, the full analysis payload is wrapped into an encrypted export object.

The browser path derives an AES-GCM key from the local export key with Web Crypto, then stores only salt, nonce, iterations and ciphertext in the JSON file. The Python CLI mirrors the same idea with the cryptography package, encrypted --out by default, --plain-out for an explicit local debug file and voicesafekit decrypt for intentional recovery.