SHREDDARDOCS
/
DOCS/CORE CONCEPTS
LIVE8 MIN READUPDATED 03 SEP 2026

CORE CONCEPTS / 01

A complete request path you can inspect.

Eight bounded stages protect context before, during, and after inference.
SHREDDAR places policy decisions before external inference and output inspection after it. A future local model may improve recognition, but deterministic secret controls, the Vault boundary, and the output guard remain independent layers.
A carbon privacy machine holds coral sensitive fragments locally while four lime protected-context lanes travel to separate processors and a paper receipt leaves on an independent rail
SHREDDAR / DOCUMENTATIONLOCAL IDENTITY / BOUNDED ROUTES / SEPARATE EVIDENCE
REQUEST LIFECYCLE

Eight stages, two hard boundaries

  1. 01

    VALIDATE

    Strict schemas, authentication, field bounds, body-size checks, and policy lookup reject malformed or unauthorized input.

  2. 02

    DETECT

    Priority-ordered patterns, checksums, IP validation, and a small local gazetteer produce non-overlapping sensitive spans.

  3. 03

    MINIMIZE

    Task keywords identify the fact groups needed for the stated job; unrelated sensitive lines can be removed in stricter modes.

  4. 04

    TRANSFORM

    Policy applies keep, remove, pseudonymize, generalize, conceptualize, block, or local_only to each span.

  5. 05

    ROUTE OR SHARD

    Risk selects local-only or transformed routing. SHARD can split paragraphs into two to eight bounded purpose groups.

  6. 06

    INFER

    Demo mode returns synthetic output. A self-hosted operator may configure a non-streaming OpenAI-compatible upstream.

  7. 07

    GUARD + STITCH

    Credential-like output hard-fails. New direct identifiers are replaced; only authorized current-session placeholders may reconstruct.

  8. 08

    RECEIPT

    A SHA-256-linked record stores policy, counts, route, timing, reduction, and integrity metadata—not source or mappings.

DATA MOVEMENT

What crosses each line

LINEMAY CROSSMUST NOT CROSS
Source → compilerAuthorized raw source and exact taskUnbounded unrelated datasets
Compiler → upstreamProtected messages approved by routeRaw identifiers, credentials, Vault mappings
Upstream → output guardModel resultA bypass around output inspection
Application → VaultAuthorized placeholders plus session credentialsCredentials inside agent or model context
Operation → receipt storeCounts, hashes, timing, route, policyPrompts, protected text, source values, mappings
FAIL-CLOSED RULES

Some outcomes should stop the workflow

  • A missing or materially ambiguous purpose for sensitive external processing.
  • A local_only route for the selected policy and data class.
  • A detected credential, private key, seed phrase, or other blocked secret.
  • An unavailable gateway or an MCP tool error.
  • Output that triggers the credential guard or cannot be safely inspected.
SHREDDAR DOCSYOUR DOC. YOUR DATA. YOUR CONTROL.

Public alpha documentation. Capability labels describe implementation state, not certification or guaranteed suitability.