RELEVANCE
Does the stated task need this fact at all?
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.
INPUT Document + task + policyOUTPUT Minimum useful context
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.
PRECISION
Would a range or category preserve enough utility?
LINKABILITY
Can this detail identify or reconnect the subject?
ROUTE
May the protected representation leave the trusted boundary?
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 INPUT
Mira 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–€90KUse 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 INPUT
Rijksmuseum Twenthe↓PROTECTED VIEWregional cultural museumUse 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 INPUT
mira@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 INPUT
demo_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 INPUT
health information↓PROTECTED VIEW<HEALTH_INFORMATION_LOCAL_ONLY>The route—not only the wording—is constrained by policy.
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.
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.
POLICYID + HASH
ACTIONSCOUNTS ONLY
ROUTEPROVIDER CLASS
EXPOSURECHARACTERS / TOKENS
TIMINGLATENCY
INTEGRITYRECEIPT HASH
The complete
request path.
Transformation is one layer. The product boundary also governs routing, output inspection, reconstruction, and privacy-safe observability.
ACTIVE STAGE01 / 07
SOURCE ANALYSIS
Find what should not travel.
SHREDDAR identifies sensitive entities, secrets, exact values, and linkable facts before any model route is opened.
- OPERATION
- ENTITY + SECRET SCAN
- BOUNDARY
- RUNS INSIDE YOUR TRUST BOUNDARY
ACTIVE STAGE02 / 07
TASK MINIMIZATION
Give every fact a reason to exist.
Task relevance and policy determine whether each fact should be kept, reduced, removed, blocked, or reserved for local inference.
- OPERATION
- TASK + POLICY EVALUATION
- BOUNDARY
- NO EXTERNAL ROUTE IS OPEN YET
ACTIVE STAGE03 / 07
PRIVACY COMPILATION
Reveal less. Keep the utility.
The compiler chooses the least revealing representation that still supports the task: pseudonymize, generalize, conceptualize, remove, block, or keep local.
- OPERATION
- 6 PRIVACY ACTIONS
- BOUNDARY
- ORIGINAL VALUES STAY IN THE VAULT
ACTIVE STAGE04 / 07
EXPOSURE CONTROL
Split the work. Constrain the route.
SHARD can separate coherent subtasks and send each route only the protected facts it needs, while local-only classes remain inside the boundary.
- OPERATION
- SHARD + LOCAL-ONLY
- BOUNDARY
- ONLY APPROVED COMPARTMENTS LEAVE
ACTIVE STAGE05 / 07
MODEL EXECUTION
Send the protected view—not the source.
Only policy-approved protected context reaches the configured model. The external route receives enough meaning to work, not the complete private document.
- OPERATION
- CONFIGURED MODEL ROUTE
- BOUNDARY
- PROVIDER SEES PROTECTED CONTEXT
ACTIVE STAGE06 / 07
TRUSTED RETURN
Guard the return. Resolve locally.
The protected answer is inspected, then only authorized current-session placeholders are resolved against mappings held inside the local Vault.
- OPERATION
- OUTPUT GUARD + STITCH
- BOUNDARY
- IDENTITY MAPPINGS NEVER LEAVE
ACTIVE STAGE07 / 07
PRIVACY-SAFE EVIDENCE
Leave evidence, not a copy.
A Privacy Receipt records policy, action counts, route, timing, exposure, and integrity metadata without storing the raw prompt or response.
- OPERATION
- METADATA-ONLY RECEIPT
- BOUNDARY
- NO RAW CONTENT IN THE RECEIPT
Stage 01 of 07: DETECT. Find what should not travel.
WHAT WORKS IN THE PUBLIC ALPHA
See the boundary
before trusting it.
Text and supported file previews run in this tab. Use synthetic or authorized content.
The OpenAI-compatible gateway is designed to run inside the deployer's trust domain.
The deployed gateway is reserved for operator verification, not anonymous private-data processing.
Pseudonymized data may remain personal data. SHREDDAR reduces unnecessary disclosure.