Proposed vendor input

Keep buyer knowledge inside.

Scan locally before text crosses into an AI product.

0 / 1,500NOT SENT

Local boundary: text is processed in this browser. Model files are fetched on first use; your input is not uploaded by this page.

Review the membrane

Exact spans, not a vague score.

Run the bundled sample or paste a draft. Each finding will show its text, type, rule-tagger provenance, character offsets, and detection engine.

Found0

Selected0

Coverage—

Reviewed output

Only sanitized text exits.

No scan yet.

Sanitized promptHASH —

Your reviewed result will appear here before any download.

Interpretation: a clean scan is not proof of safety. Rule-based NLP can miss codenames and specialized trade secrets; review business context before sharing.

Knowledge loss begins before “send.”

The useful decision happens at the boundary: which exact words need to cross, and which can be replaced without changing the task?

Local NLP tags semantic entities.The bundled rule tagger identifies person and organization names with exact offsets and provenance, without sending text to a model host.

Patterns catch exact structured disclosures.Email, phone, money, project, and account-like spans remain available if the NLP module cannot load.

You decide every crossing.Selected replacements update the visible prompt immediately; export only preserves evidence of work already completed.

Understanding the local redaction pipeline

Your browser does not support video.

Read the explanation

Knowledge Airlock scans browser text using exact patterns and, when loaded, a local Compromise rule tagger. Its model mode is not a generative model or a remote model request. If the external rule tagger cannot load, the scan keeps the exact pattern fallback and labels reduced coverage. Findings are review aids, not proof that every secret or business-sensitive phrase was found. The pipeline ends in a user selection, not automatic permission to share. In the example Pay one hundred dollars to a at b dot example, the twenty four character input has a money span from offset four through eight, excluding the endpoint, and an email span from twelve through twenty three. The sanitizer sorts selected spans from right to left. Replacing the email first preserves the earlier money offsets. It then replaces the money, producing twenty three characters of sanitized text. This order avoids shifting a later span before it has been replaced. The merge routine orders candidates by start offset, length, and score. When an exact pattern overlaps an existing rule-tagger entry, it replaces that entry. Otherwise the routine keeps non-overlapping findings independently reviewable. Pattern scores of one are rule tags, not a calibrated probability that the disclosure is sensitive. Deselecting a finding restores that original span in the reviewed output. This preserves user choice but can also leave sensitive context in the text. The scanner rejects empty inputs and lengths above fifteen hundred, even though the textarea accepts eighteen hundred characters. Export requires a completed scan and contains original finding text, offsets, selections, runtime details, and sanitized output. The evidence file itself may therefore contain sensitive material. Typing after a scan updates the character count but does not clear the previous findings or completed run. Rescan the edited text before relying on those offsets or exporting. No clean scan or replacement guarantees safe vendor use.