Q01Can we run the entire crew on a local model so client data never leaves the firewall?
Yes, this is the default for privileged work. Point each persona at a local Ollama or LM Studio instance running on your own hardware. ContractReviewer, AMLSpecialist, KYCLead, and PrivacyOfficer can run end-to-end with zero outbound traffic. Cloud models (Claude, GPT, Gemini) stay an option per persona for non-privileged matter work. The per-agent model picker is the privacy lever.
Q02Where does our data physically sit when the crew runs?
On your hardware when you run a local model. Every prompt, every retrieved document, every draft, and every audit log stays on the machine running the runner. The per-workflow vector store is yours to host. Cloud models only see the prompts the agent sends them, and you decide which personas can call them. EU teams can deploy entirely on EU-hosted Mistral or a local model to meet MiCA, GDPR, and DORA residency obligations.
Q03How does this handle attorney-client privilege and litigation hold?
Privileged matters route to a separate workspace with scoped tools and an explicit hold flag. On hold, the agent cannot call external search, web_fetch, third-party tools, or cloud models, full stop. The local-model lane is the default for hold work. Every read is logged with timestamp, persona, model id, and matter anchor so privilege is defensible line by line.
Q04Will the agents file SARs or send regulator responses on their own?
No. Every SAR draft, regulator reply, redline, and licence renewal email is staged for a human signer. The AML, KYC, and litigation personas ship with HITL on every write so the lawyer who signs is the human, not the agent.
Q05Can the agents reason over our policies, contracts, and regulator letters?
Three layers, all per-workflow and tenant-isolated. Static context attaches your AML policy, contract playbook, and licence conditions to specific personas on every run. The rag_retrieve tool lets ContractReviewer and RegulatoryAffairs pull from prior matters, regulator correspondence, and your DPA library on demand. Cross-run memory carries yesterday's litigation hold and DSR queue into today's run. None of those layers leave your tenant or the local runner you point them at.
Q06Does this replace Ironclad, Harvey, Robin AI, or Spellbook?
No, it sits next to them. Melaya is the orchestration and audit layer across nine legal personas, with the option to run every one on a private model. If you already pay for Ironclad as your CLM or Harvey for litigation research, ContractReviewer hands drafts to them through tool calls and keeps the replay log on your side.
Q07How do we keep the output from sounding like AI?
Every redline, SAR, and DPIA cites the specific clause, transaction hash, or regulator paragraph it relies on, pulled from your own corpus loaded into the knowledge store. Compliance officers can require a citation on every paragraph as a HITL pre-check before signing.
Q08How fast can a legal team get the first pipeline running?
With your DMS or shared drive connected, a contract playbook loaded into the knowledge store, and Ollama running locally, the redline workflow is a 4-node canvas: ingest, retrieve precedent, draft redline, route for approval. Most teams ship the first version in a working session and never send a single token to a cloud vendor.
Q09Can I audit exactly what each agent did and why?
Yes. Every run logs every tool call, every model invocation (with model id), every retrieved document, and every approval decision. Replay any matter at any time. Examiners get the full evidence pack in under five business days, which is the ComplianceMonitor target.