DocsOpen appApp

THE PRIVACY COMPILER / PRODUCT

ONE SOURCE.
SIX WAYS
TO SHRED.

SHREDDAR does not apply one blunt redaction rule. It decides what the task needs, transforms each sensitive fact, and controls where the result can travel.

OPEN THE APP ↗READ THE DOCS →
INPUT Document + task + policyOUTPUT Minimum useful context
A private document enters a carbon privacy compiler and exits through six distinct transformation lanes, including anonymous tokens, ranges, abstraction, removal, blocking, and a local-only return path
TRANSFORMATION MAP / 01SIX ACTIONS. ONE POLICY DECISION.
01 / BEFORE THE MODEL

Shredding is a decision.
Not a black marker.

Traditional redaction asks what looks sensitive. SHREDDAR also asks what is necessary for this specific job. The same fact can be pseudonymized for one task, generalized for another, removed from a third, or held for local inference.

01

RELEVANCE

Does the stated task need this fact at all?

02

PRECISION

Would a range or category preserve enough utility?

03

LINKABILITY

Can this detail identify or reconnect the subject?

04

ROUTE

May the protected representation leave the trusted boundary?

02 / THE ACTIONS

Different data.
Different destruction.

Each action makes a different tradeoff between utility, precision, reversibility, and exposure. Policies choose by entity class; task minimization can remove facts the job does not require.

01PSEUDONYMIZE

Keep the relationship

Replace a direct identifier with a stable typed placeholder. The model can follow who did what without receiving the original identity.

PRIVATE INPUTMira Chen↓PROTECTED VIEW<PERSON_01>

The reversible mapping stays in the current local Vault session.

02GENERALIZE

Reduce the precision

Turn an exact value into the coarsest useful range. The task keeps scale or timing while losing unnecessary precision.

PRIVATE INPUT€82,471↓PROTECTED VIEW€80K–€90K

Use when magnitude matters but the exact value does not.

03CONCEPTUALIZE

Keep the function

Replace a named entity with a task-relevant category or role. Meaning survives while the link to a specific organization is weakened.

PRIVATE INPUTRijksmuseum Twenthe↓PROTECTED VIEWregional cultural museum

Use when function matters more than which exact entity it is.

04REMOVE

Delete the irrelevant fact

Drop a field or complete line when the stated task does not need it. This is task minimization, not cosmetic masking.

PRIVATE INPUTmira@company.example↓PROTECTED VIEW— removed —

Removed content is absent from the protected representation.

05BLOCK

Stop the secret value

Replace credential-like material with a typed presence marker so a task may know a secret exists without exposing the secret itself.

PRIVATE INPUTdemo_key_••••••••↓PROTECTED VIEW<API_SECRET_PRESENT>

The original credential value is never eligible for forwarding.

06LOCAL ONLY

Do not take this route

Mark sensitive classes for local inference. In gateway mode, a local-only route stops forwarding to a configured cloud model.

PRIVATE INPUThealth information↓PROTECTED VIEW<HEALTH_INFORMATION_LOCAL_ONLY>

The route—not only the wording—is constrained by policy.

One private document enters a local boundary and three separated routes carry different small subsets of protected context to isolated processing modules
SEMANTIC COMPARTMENTS / SHARD

03 / SHARD

Split by meaning.
Route by risk.

After transformation, SHARD can split a complex task into coherent compartments—for example obligations, financial terms, and termination. Each route receives only the facts its subtask requires.

  • Less complete exposureNo single route needs the entire source.
  • Independent routingSome compartments can remain local while others use configured models.
  • Local synthesisThe final task view is assembled back inside the boundary.

PRECISE CLAIM SHARD reduces exposure; it does not create cryptographic secrecy. Correlation can remain possible when providers, metadata, or fragments overlap.

04 / RETURN PATH

Inspect. Stitch.
Leave evidence.

The answer returns protected. SHREDDAR inspects it, reconstructs only valid current-session placeholders inside the trusted boundary, and emits a receipt that describes the operation without copying the content.

Protected context lines return to a local identity vault where authorized mappings remain separated
STITCHAuthorized references resolve locally.
A privacy boundary releases sparse protected context while a separate abstract receipt records transformation metadata without source content
RECEIPTPolicy and integrity evidence, not the prompt.
POLICYID + HASH
ACTIONSCOUNTS ONLY
ROUTEPROVIDER CLASS
EXPOSURECHARACTERS / TOKENS
TIMINGLATENCY
INTEGRITYRECEIPT HASH
05 / END TO END

The complete
request path.

Transformation is one layer. The product boundary also governs routing, output inspection, reconstruction, and privacy-safe observability.

  1. 01

    DETECT

    Find sensitive entities, secrets, exact values, and linkable facts.

  2. 02

    DECIDE

    Apply task relevance and the selected policy to every fact.

  3. 03

    TRANSFORM

    Choose the least revealing representation that still supports the task.

  4. 04

    ROUTE

    Keep local-only classes local and split bounded work when useful.

  5. 05

    INFER

    Send only policy-approved protected context to the chosen model.

  6. 06

    INSPECT + STITCH

    Guard the return and resolve only authorized current-session placeholders locally.

  7. 07

    RECEIPT

    Record policy, counts, route, timing, and integrity metadata—never raw content.

WHAT WORKS IN THE PUBLIC ALPHA

See the boundary
before trusting it.

BROWSER-LOCAL APP

Text and supported file previews run in this tab. Use synthetic or authorized content.

SELF-HOSTED GATEWAY

The OpenAI-compatible gateway is designed to run inside the deployer's trust domain.

AUTHENTICATED HOSTED API

The deployed gateway is reserved for operator verification, not anonymous private-data processing.

NO ANONYMITY CLAIM

Pseudonymized data may remain personal data. SHREDDAR reduces unnecessary disclosure.

OPEN SHREDDAR APP ↗READ DOCUMENTATION →