Agent, harness, and platform
| Layer | Responsibility | Clarivy surface |
|---|---|---|
| Agent | Perceive authorized case data, plan within policy, call approved tools, observe results. | Case AI dock and specialist team roles |
| Harness | Governed runtime for state, tool calls, memory scopes, citations, evals, and human review. | ClarivyLLMHarness and /cases/{id}/ai/* |
| Platform | Multi-tenant identity, durable execution, cost controls, and tenant-filtered retrieval. | Tenancy, Postgres/SQLite audit, object storage, Qdrant filters |
Decision authority
Controlled AI review
Document intelligence pipeline
Governed specialist team
| Specialist | Bounded responsibility | Control boundary |
|---|---|---|
| Citation verifier | Extract supported claims and verify exact source quotes. | Cannot use another tenant or application. |
| Gap planner | Identify missing, stale, rejected or contradictory material. | Creates review work; cannot change policy or request data alone. |
| Packet composer | Draft a cited review summary and decision-packet narrative. | Cannot approve, reject or close the application. |
| Quality sentinel | Check validation failures, citations and signed activity lineage. | Blocks unsafe output; never repairs it silently. |
Server-side orchestration
Seven control gates
- 1
Authorize the context
Load the case using the authenticated tenant and limit selection to evidence already attached to that case.
- 2
Minimize identifiers
Replace common direct identifiers in the temporary model context without mutating authoritative evidence.
- 3
Use a versioned task contract
Bound the task, context, output length, prohibited actions and required human review.
- 4
Route to an approved provider
Use a local, private or customer-approved managed provider. Remote processing is disabled unless explicitly permitted.
- 5
Validate the structure
Reject unknown fields, missing fields, placeholders or output that does not match the strict schema.
- 6
Verify every citation
Confirm that evidence IDs belong to the case and exact quotes occur in the supplied source.
- 7
Require human review
A named user accepts or rejects the assistance with a reason. Acceptance does not silently create evidence.
What the model may and may not do
| Allowed assistance | Prohibited action |
|---|---|
| Extract source-supported fields | Invent or silently correct a material value |
| Summarize approved case evidence | Browse beyond the authorized case boundary |
| Identify possible gaps or contradictions | Turn uncertainty into a customer risk score |
| Prioritize review tasks | Approve, reject or close a material gate |
| Draft cited reviewer text | File an STR/SAR or trigger recovery action |
Machine-learning capabilities
| Capability | Model signal | Reviewer control |
|---|---|---|
| Multilingual entity resolution | Name embeddings or deterministic multilingual similarity, official identifiers, address and country | Shows the weighted contribution of every feature; conflicting identifiers block a match and records are never merged automatically |
| Document intelligence | OCR text, page-linked field candidates, document class and extraction confidence | The original file and page remain visible; a reviewer confirms or corrects each material field |
| Contradiction and gap prioritization | Differences across cited case evidence and policy-required proof | Creates review work and keeps both claims visible; it does not turn uncertainty into customer risk |
| Controlled case assistance | Schema-bounded extraction, summaries and gap analysis over authorized case evidence | Every quote is checked against a case source and the output needs a named acceptance or rejection |
Compare two legal-entity records
Entity resolution is designed for candidate matching across a case record, company registry, sanctions or PEP screening result, CRM record, and internal risk signal. A score is an investigation aid—not a customer decision.
POST /ml/entity-resolution/compare
{
"left": {
"name": "Al Noor Trading LLC",
"identifiers": { "trade_license": "CN-12345" },
"country": "AE"
},
"right": {
"name": "AL NOOR TRADING L.L.C.",
"aliases": ["شركة النور للتجارة ذ.م.م"],
"identifiers": { "trade_license": "CN-12345" },
"address": "Business Bay, Dubai",
"country": "AE"
}
}Explainability contract
Document processing lifecycle
- 1
Preserve the original
Record filename, media type, uploader, time, file hash and quarantine state.
- 2
Normalize and extract
Run bounded text extraction or OCR with page, character and execution limits.
- 3
Propose fields
Produce page-linked candidates such as legal name, licence number, expiry date or address.
- 4
Compare and correct
A reviewer checks each material value against the original page and records any correction.
- 5
Promote deliberately
Only reviewed fields can support a policy gate; low confidence becomes review work, never customer risk.
Run assistance through the authenticated workspace API
POST /cases/{case_id}/ai/agent-team-runs
{
"objective": "Prepare a policy-bound onboarding review for a named human decision."
}POST /cases/{case_id}/ai/runs
{
"task": "claim_extraction",
"evidence_ids": ["evidence-uuid"],
"focus": "Keep contradictions between payment evidence visible."
}POST /cases/{case_id}/ai/runs/{run_id}/review
{
"status": "accepted",
"note": "Exact citations and stated limitations were checked."
}Safe degradation
Continue reading
Security model
Review tenant, identity, evidence and AI boundaries together.