Every batch in Philter Workspace answers three questions before a redacted document leaves it: which detections a person has to decide, which documents a person opens at all, and how many people have to approve a document before it can be finalized. This guide covers the batch settings behind each answer. Set them before the first real batch. The PII threshold and audit sampling rate are applied when a document is redacted, and changing them later does not revisit documents already in the queue.
1. Choose the PII threshold
The PII threshold is a per-span confidence floor, from 0 to 1. When Philter returns a detection whose confidence is at or above it, Philter Workspace accepts that span. Anything below it arrives unreviewed. A document with even one unreviewed span is REVIEW_REQUIRED, and a reviewer has to accept or refuse that span before the document can be approved.
The default is 0.80. Lowering it accepts more spans without a person looking at them, and raising it sends more to reviewers. The threshold only sorts the detections Philter made. A value Philter did not detect produces no span at all, so no threshold setting sends it to a reviewer. Reviewers who read the Original pane, and not only the PII Navigator, are how missed values get found and added by hand.
2. Set the audit sampling rate
A document where every span cleared the PII threshold has nothing left to decide. The audit sampling rate decides whether it goes to a reviewer for a spot-check anyway. Each such document gets one random roll when it is redacted. If the roll falls under the rate, the document is AUDIT_REQUIRED and goes to a reviewer. Otherwise it is AUTO_APPROVED, which skips the review queue but still needs an approval before it can be finalized (see step 6).
The default is 0.10, so roughly one in ten of those documents is spot-checked. At 0 none are, and at 1.0 all of them are, which means a person approves every document in the batch.
An administrator can set the defaults for new batches, and a floor that no batch may go below, in the deployment’s configuration:
philter-workspace.batch.defaults.pii-threshold=0.80
philter-workspace.batch.defaults.document-threshold=0.25
philter-workspace.batch.defaults.audit-sampling-rate=0.10
philter-workspace.batch.minimum-audit-sampling-rate=1.0
Each property can also be set as an environment variable, such as PHILTER_WORKSPACE_BATCH_MINIMUM_AUDIT_SAMPLING_RATE. Every value must be from 0 to 1, and a default rate below the minimum stops Philter Workspace at startup. With a minimum set, saving a batch with a lower rate is refused and recorded in the audit log as BATCH_AUDIT_SAMPLING_RATE_REFUSED. A minimum of 1.0 is how a deployment requires a person on every document.
3. Treat the Document Threshold as a label
The Document Threshold (default 0.25) is easy to misread. A document whose risk score is at or below it gets a Low risk badge in the queue, and that is all. It approves nothing and does not change the document’s status. A REVIEW_REQUIRED document can carry the badge and still have spans waiting. Use it, with the queue’s default sort by risk score, to decide where reviewers start.
4. Apply the settings to a batch
Thresholds and the sampling rate are set in the Create Batch dialog on the Batches page, and changed later with Settings on the batch’s row. Both are open to administrators and to team leads of the batch’s group. Each change is recorded in the audit log as BATCH_THRESHOLDS_CHANGE, with the old and new values.
5. Decide who signs off
Approve on the review page records an approval. It is refused while any span is unreviewed or awaiting a second opinion, and the same person can never approve a document twice. Without approval rule sets, one approval moves the document to APPROVED.
Approval rule sets, under Admin → Approval Rules, make some documents need a second approval from a different reviewer. Conditions inside one rule set must all hold, and any one rule set on the batch is enough. The conditions available are:
| Condition | Default cutoff |
|---|---|
| Document contains an SSN, credit card, or passport number | none |
| Risk score above the cutoff | 0.9 |
| Reviewer refused a span with confidence above the cutoff | 0.95 |
| Approving reviewer has fewer reviews than the cutoff | 100 |
| Reviewer added more than the cutoff number of spans by hand | 5 |
| Document contains a listed keyword (case-insensitive) | Classified, Proprietary, Secret |
| Document is randomly sampled for dual approval | 0.02 |
A rule set can also require that the second approval come from an approver of the batch’s group. For example, one rule set with the sensitive-identifier condition and another with a 0.03 sampling rate sends every document with a detected SSN, plus about 3% of the rest, to two people.
Approvals cover the spans as they were approved. Any change to a span afterwards, including an approver’s adjudication, clears the approvals recorded so far and logs DOCUMENT_APPROVALS_CLEARED, so a dual-approval document needs both approvals again on its final spans.
If you need a measure of how consistently reviewers decide, Blind Double Review on the Create Batch dialog sends a share of documents (default 10%) to an independent second review. It can only be set when the batch is created. A selected document cannot be finalized until both reviews are recorded and an approver reconciles them.
6. Finalize
Finalize appears in the queue once a document is APPROVED with all its approvals. It renders the redacted text, issues a certificate of redaction, applies the batch’s finalization policy to the original, and queues delivery to the batch’s destination if one is set. FINALIZED cannot be reopened. Finalize accepts only APPROVED documents, so an AUTO_APPROVED document has to be opened and approved by someone before it can be finalized. A batch can be closed only when every document in it is rejected or finalized.
What to do next
Before you lower a batch’s audit sampling rate or PII threshold, open a few of the AUDIT_REQUIRED documents it has already sent to reviewers and see what they changed. If reviewers are refusing accepted spans or adding missed ones, the threshold is accepting more than it should, and lowering either setting will hide that. The Philter Workspace documentation covers the risk score formula, PII type weights, and finalization policies in detail, and the support page links the issue tracker. Review reduces what gets through. It does not make output complete, so validate redacted documents against your own data before you rely on them.