Bounded AI Architecture (BAA) is how we build private AI.
- Paddock (Knowledge Plane)
- Binds your Data Modules directly to the reasoning layer.
- Ram (Reasoning Layer)
- Deploys full-capability frontier models that rely strictly on your Data Modules for knowledge.
- Watchman (GRC Layer)
- Unifies governance, risk, and compliance into a single, complete audit log.
Nothing leaves your boundary.
- Bounded AI
- Paddock
- Ram
- Watchman
- Shepherd
- Merino
- Neural Pod
- RAI Framework
BOUNDED AI ARCHITECTURE
Three layers, each bound to something you control
Bounded AI Architecture is how Black Sheep AI builds private AI. Paddock binds your Data Modules directly to the reasoning layer. Ram deploys full-capability frontier models that rely strictly on those modules for knowledge, and you swap them as better models ship. Watchman unifies governance, risk and compliance into a single, complete audit log. Nothing leaves your boundary.
- 01 · The Reasoning Layer
Ram
Full-capability frontier models, shrunk to fit smaller GPUs with reasoning kept and certified. They rely strictly on your Data Modules for knowledge, so you swap the model as better ones ship.
How Ram works - 02 · The Knowledge Plane
Paddock
Binds your Data Modules directly to the reasoning layer: one per topic, period, person, permission or tenant. Permission, freeze, hand back or delete each one on its own.
How modules work - 03 · The GRC Layer
Watchman
Governance, risk and compliance in a single, complete audit log: every module permission, every bind, every model's provenance. Five certificates before anything goes live.
What Watchman governs
Frontier capability without the frontier bill
Built for data that
cannot leave the building
WHY BLACK SHEEP AI
Cited Answers
Retrieval accuracy on generated test questions, September 2026.
Cited Answers
Ram: Smaller, Same Reasoning
TARGET : 12.6 GB
12.6 GB
MODEL ON DISK
32K
CONTEXT HELD
16 GB
GPU MEMORY
HARDWARE : ONE 16 GB GPU, THE CLASS EVERY CLOUD RENTS BY THE HOUR
Ram: Smaller, Same Reasoning
Certified Builds
TERMINAL
Certified Builds
A Fraction of the Cost
A 27B model: 54 GB at full precision, 12.6 GB with Ram.
A Fraction of the Cost
Air-gapped by design
Runs on-premise and air-gapped. Zero cloud calls, so neither your data nor the audit ever leaves your infrastructure.
Data Modules
Partition by topic, time, person or tenant. Each Data Module stands alone, permissioned on its own, so GDPR and data-residency rules are met by construction.
Regulatory evidence
CycloneDX AI-BOM attestations mapped to the EU AI Act, NDAA/DFARS and OMB requirements. Audit-ready output, not screenshots.
CI-native gate
Watchman returns deterministic exit codes for your pipeline: 0 clean, 2 modification detected, 1 indeterminate.
What you can build
Cited answers from your documents
Exact values from manuals and policies with the page cited every time.
Data Modules
One Data Module per topic, person, permission or tenant, each permissioned on its own.
Point-in-time recall
Ask what a manual required on the day of the incident.
Research Edition
A lab that never forgets, negative results first.
Private AI that holds up
across regulated industries
One module per patient, removable in one operation the day the obligation ends
Read the use case
Matters kept apart by structure, so a query outside the grant has nothing to search
Read the use case
Compartments that live on hardware you control, with zero cloud calls
Read the use case
Answers from the manual revision that was current on the day
Read the use case
Desks separated by module, not by policy, with every access logged
Read the use case
Point-in-time recall of the policy wording a claim was written under
Read the use case
CAPABILITIES
