Architecture decision • October 24, 2026
Build online.
Run on your own hardware.
Recommended direction: one portable CaseSignal application, a replaceable local AI engine, and separate public-record search tools. Develop online now; validate the same release on supported agency hardware before offering standalone installations.
Download full research and implementation plan Download development source
Three deployment modes
| Mode | Evidence and AI | Public searches |
|---|---|---|
| Hosted pilot | Private servers and private model endpoint | Enabled internet sources |
| Agency machine with internet | Local storage, processing and AI | Same permitted connectors; queries leave the machine |
| Disconnected agency machine | Preloaded runtime, model and saved evidence | Previously imported records only |
Standalone software can analyze evidence offline. Fresh web searches require internet access. Hardware profiles need separate validation; containers do not remove GPU, driver or architecture requirements.
The AI recommendation
Test Ministral 3 8B Instruct first against Qwen3.5-9B. Both publish Apache-2.0 model licenses. Neither has been benchmarked for CaseSignal. Use one production model after measuring citation accuracy, false connections, speed and memory.
Start with llama.cpp for portable inference. Keep the application behind a model adapter so runtime or provider changes do not rewrite the investigation workflow. Fine-tuning is deferred; case facts remain in permission-scoped evidence retrieval.
Search first, explain with evidence
Use free official APIs/downloads, approved government-site retrieval and guided import. Evaluate private SearXNG discovery as an optional connector. It forwards queries to external engines and can be blocked; it is not unlimited free access or a complete records database.
Keep exact identifier matching separate from semantic search and AI interpretation. Each suggested association needs its source passage, document hash, date and page/timestamp. Shared numbers, registered agents, reused identifiers and contradictory records remain separate until detective review.
Practical build order
- Prepare the portable core and a production hosted runtime with agency/case isolation.
- Add quarantined intake, malware scanning and isolated PDF/OCR extraction.
- Add private model inference and source-cited conversations within a case.
- Connect permitted government retrieval; record source coverage and failures.
- Add reviewed relationships and test on labeled fictional cases.
- Install the exact release on a fresh machine and verify disconnected operation.
Planning target: 64 GB RAM, 24 GB GPU memory and 1–2 TB NVMe for pilot headroom. This is an engineering estimate, not measured performance or a purchase recommendation. Size hardware after the model evaluation.
What changed in this release
The source now includes this architecture decision and a read-only bundle checker. It checks application/runtime/model/license files against a supplied SHA-256 manifest and rejects missing, changed or unsafe paths. It never downloads or runs files. Checksums require a trusted manifest and do not certify hardware or production readiness.
Current status: this online Site remains a private preview/download surface. The functional Python development backend runs separately. Protected intake, ClamAV integration and PDF page extraction are now implemented in the downloadable backend; live scanner/sandbox activation remains pending on a supported host. AI, automatic retrieval, encrypted offsite backups and an offline production installer are not active. Existing full backups remain local and unencrypted. Use fictional test data only.