Transcript Lake

Documentation / Operate

Walkthrough: audit synthetic masking and derived output

This walkthrough uses the same isolated 2026-08-24 synthetic fixture as the live-append walkthrough. All apparent credentials were invented specifically to exercise the three masker classes.

This walkthrough uses the same isolated 2026-08-24 synthetic fixture as the live-append walkthrough. All apparent credentials were invented specifically to exercise the three masker classes.

Stored evidence#

The initial user event became:

json
{"ts":"2026-08-24T09:00:00.000Z","runtime":"claude","machine":"demo-host.local","session_id":"0a1b2c3d-1111-4222-8333-444455556666","project":"/tmp/demo-app","event_type":"user","text":"Deploy fails. I set [masked:assignment:37:5be474f8] and the provider token is [masked:token:36:e3bebba1] - can you check the config?"}

The tool result retained only assignment markers:

json
{"event_type":"tool_result","text":"[masked:assignment:37:5be474f8]\n[masked:assignment:54:b3bc1b09]\n"}

The live append became:

json
{"ts":"2026-08-24T09:01:00.000Z","event_type":"user","text":"Also scrub this pasted blob before we archive: [masked:entropy:44:d0fa7c09]"}

The repeated 5be474f8 demonstrates stable equality correlation. The markers disclose class and original character count, not plaintext.

Plaintext absence check#

The retained audit searched for all three invented values under the isolated lake/. Counts were zero in:

text
lake/stream-status.json
lake/exports/oko/runtime=claude/<session-hash>.jsonl
lake/cursors.json
lake/events/runtime=claude/date=2026-08-24/part-b5939b619a3d.ndjson

It concluded NO PLAINTEXT SECRETS UNDER lake/. That conclusion is scoped to these exact synthetic literals and files, not to arbitrary secret shapes.

Derived mirror#

The already-ingested partition was compacted with DuckDB 1.5.5:

text
claude: ndjson 3352 bytes -> parquet 5122 bytes (/private/tmp/lake-demo/lake/parquet/runtime=claude/events.parquet)

Because compaction reads canonical NDJSON, the Parquet mirror receives masked text. NDJSON remains authoritative and the mirror can be removed with clean --target parquet --apply after preview.

Operator conclusions#

  • Whole detected hits did not cross the durable boundary in this run.
  • Oko export and Parquet were downstream of canonical masking.
  • A repeated fingerprint is intentionally correlatable.
  • Machine/project/session metadata and ordinary text remained and must be treated as sensitive.
  • Real credentials should be rotated, not merely archived behind markers.

See the synthetic masking example and full guarantee boundary.