Architecture
How BatchGuard turns a recipe and an SOP into a list of disagreements, where the models run, and what the public demo does and does not include today.
From two documents to a list of findings
- Read. Both documents are converted to text and held in memory for the check, unchanged; every quote later points back into this text. Nothing is stored.
- Plan. The recipe and the SOP are split into pairs of parts that fit the model's context, so that every character of both documents is covered.
- Compare. An open model compares each pair and reports where the recipe and the SOP disagree, quoting both.
- Verify. Each quote is checked against the source text; a finding whose quote is not found is not reported as evidence.
- Reconcile. Findings from all parts are brought together, something "missing" in one part is checked against the rest of the recipe, and duplicates are merged with every citation kept.
- Report. A short summary and a complete report, each finding with its quotes from both documents, for the reviewer to decide on.
Where the models run
Inference runs on open models (GLM-5.3) through Tinfoil. Before any text is sent, our server verifies the attestation of Tinfoil's router and encrypts to it; the router verifies the model and document-conversion enclaves it forwards to. We do not verify those enclaves directly. Tinfoil is US-based.
The demo today, and phase 2
- Phase 1, live now: the public demo runs on ordinary cloud hosting with synthetic example documents only. Uploads are switched off.
- Phase 2: confidential hosting for the service itself, with uploads of your own documents.
What BatchGuard is not
BatchGuard is decision support for a reviewer. It does not release batches, approve recipes or make a compliance statement, and a short list of findings is not a certificate that the recipe is right.