Skip to Content
PlatformData HandlingRedaction evidence

Redaction evidence

Redaction replaces sensitive values with a tag such as [REDACTED:email] before a span is stored. Redaction evidence is the count of those replacements, so you can see that it is happening and where, without reading any payload.

Two counts

Every span gets two counts, each broken down by tag:

  • By your SDK. The Morse SDK redacts in your process, before a span leaves it, and leaves a [REDACTED:<tag>] marker behind. Morse counts the markers that are already in the payload when it arrives.
  • By Morse. Morse then applies its own redaction on arrival: the always-on patterns, the optional ones your workspace has enabled, your custom patterns, and trace-level redaction. Morse counts its own substitutions.

A value redacted in your process is counted once, under SDK. A value that got past the SDK and was caught on arrival is counted once, under Morse. A span that arrives with nothing sensitive in it has no counts.

The tags are api_key, jwt, credit_card, email, phone, ssn, ip, secret (the value after a key such as password), custom (one of your own patterns) and base64. A marker with any other tag is counted under other.

The counts are written by Morse. A span cannot supply its own counts: an incoming metadata key named morse.redactions or redaction_counts is dropped on arrival.

Where it shows

  • A trace. When a trace has any redactions, the trace header shows a neutral “N redacted” chip. Each span row in the waterfall shows its own chip, and the span detail lists the tags under two headings, “By your SDK” and “By Morse”. A span with no redactions shows no chip.
  • Data handling. The Redactions section totals the last 30 days by tag for each side, and lists the agents with the most redactions.
  • A share link. The same totals appear in a shared summary, with agent names but no trace ids.

What the numbers do and do not say

  • They count replacements, not people or records. One email address that appears in an input and again in an output is two.
  • They show what was redacted, not what was missed. Redaction is pattern matching on text, so a value that matches no pattern is neither redacted nor counted. See What is not redacted.
  • A trace has counts only from the time this feature was switched on. Older traces are not counted, which is different from none being redacted: they never show as zero.

Before you have data

With no counted traces in the window, the Redactions section reads Not measured. With counted traces but no redactions it reads “0 redactions in 30 days”. When counting began inside the window, the page says since when, for example “since Oct 23”.

Last updated on
Guardrail verdictsTools first seen