Two surfaces handle data differently: the local runtime deployed inside your platform, NGO, or workstation processes sensitive material, while the hub we host accepts only public-source facts, vetted pack metadata, anonymized aggregate signals, hash receipts, and consented contact metadata. Every card below opens its canonical page.
The harness, model, and packs that live on the platform's own infrastructure or the caseworker's own laptop. Raw chats, case files, IDs, and private documents stay there unless an authorized user creates a sanitized submission.
Per-role deployment paths. Each one ends with a working local runtime.
The runtime that wraps the model: grep rules first, retrieval next, tool calls mediated, and citation-critical answers refused or narrowed when they cannot cite.
Why the runtime targets Gemma 4: open weights, on-device, native tool calling, fine-tunable, ~10× cheaper inference.
Per-corridor LoRA adapters: how they are trained, where they live, how they are pinned to a base model and a pack range.
Thin wrappers around the public API. Anything an SDK does, curl can do.
The append-only public hub: vetted packs, the tools registry, anonymized signals, and the audit feed. Everything here is curator-vetted and reproducible from public inputs.
Canonical JSON schemas. Inline reference. view, copy, validate against your own payloads in the browser.
How the eight components fit together: what stays local, what is anonymized, and what the hub is allowed to store.
The contract: what stays local, how local Gemma 4 anonymization works before submission, and how the server rejects or redacts PII before storage and display.
How releases are evaluated before publication: the six suites, latest run metrics, regression policy.
Reference pages the architecture overview links into. not entry points. Open these when a section above sends you here.
Pattern-matching rules that run before the model is invoked. Cheap, deterministic, fail-closed.
Per-tool descriptors and behaviour: allow-listed lookups, external-API guards, audit hooks.
How packs are retrieved, scored, and stitched into the harness prompt. RAG citation rules.
How the research monitor proposes pack diffs from public sources without ever auto-publishing.
How the harness composes detection signals from grep rules, retrieval hits, and tool outputs.
Browse the published knowledge-pack catalog with filters by corridor and sector.
Every vetted pack release is reproducible from public inputs alone. The reproduce script lives in /tools/reproduce; it pulls the pack at a given hash, validates the curator signature, and re-runs the six evaluation suites against a clean Gemma 4 build. Same hash → same answer, every time.