Vector — Verification & Trust

Status: Available · Spec: openapi/vector.yaml · Base URL: https://vector.dev.apex.reisiger.org

Vector is the trust layer: verify claims against evidence, score entity risk, run investigations, execute methodology-pack assessments, and maintain company knowledge graphs. Wherever your platform asserts something to a user — a generated proposal, an agent's answer, a supplier profile — Vector is how you back it with a verdict and calibrated confidence.

Auth & tenancy: Authorization: Bearer <tenant API key> — the Forge-issued key itself (not a JWT), centrally verified. Tenant carriage uses two patterns: query parameter ?tenant_id=… on verification, investigations, search, compliance, and tools; URL path /api/v1/{tenant}/… on assessments, company graphs, and the evidence graph. The spec marks each operation.

Claim verification

POST /verify — {claim, context?, evidence?[], depth: quick|standard|comprehensive, evidence_source: auto|zenith|web|graph} returns a VerificationResult:

• verdict: supported / contradicted / insufficient / mixed
• confidence: calibrated 0–1 — carry it into your UI rather than collapsing to a boolean
• supporting_evidence[] / contradicting_evidence[] + reasoning — the audit trail
• cost_zar, model_used, hallucination_flagged

evidence_source: zenith grounds verification in your own corpus — the strongest pattern for generated content. POST /verify/batch takes 1–100 claims and returns per-claim results with aggregate counts and total cost.

Risk & entity screening

• POST /score-risk — multi-dimensional entity risk (financial, reputational, regulatory, operational, political, environmental) with an overall score/level, per-dimension detail, and flags.
• POST /check-entity — jurisdiction-aware entity checks (default ZA); higher-tier feature (403 on entry tiers).
• POST /investigations — asynchronous deep investigation of an entity; poll GET /investigations/{id}, and render …/{id}/graph (Cytoscape format) for the relationship picture.

Assessments (methodology packs)

POST /api/v1/{tenant}/assessments runs a structured methodology pack against collected evidence (e.g. a Microsoft 365 estate via the microsoft_graph collector), producing scored dimensions and a gap register — GET …/gaps filters by gap class (visibility, control, policy_reality), dimension, and severity. GET /api/v1/{tenant}/methodology/packs lists packs available to your tenant.

Two graphs — do not conflate them

1. Company knowledge graphs (/api/v1/{tenant}/company-graphs…) — entity/relationship intelligence you build: create graphs, add entities, assert relationships (asserted edges start at confidence 0.0 until verified), ingest from sources, request dossiers, run gap and risk analysis, export (JSON/GraphML). Verification endpoints upgrade edge confidence one relationship or a batch at a time.
2. Evidence graph (/api/v1/{tenant}/graph/…) — read-only provenance of verification activity: claims, evidence, chains, contradictions. It is written by verification runs, not by you.

Also on the surface

Federated POST /api/search across Vector's stores; compliance helpers (frameworks, POPIA checks, assessment templates); a no-auth models catalog; and tool-status endpoints. Heavy operations (/verify, /score-risk, /check-entity) carry per-tenant hourly caps (429 + Retry-After).

Integration notes

• Verify at the boundary. The high-leverage pattern: everything your platform publishes passes through /verify with evidence_source: zenith; store verification_id with the published artifact.
• Confidence is a product feature. Calibrated confidence displayed honestly builds more trust than suppressed uncertainty.
• Batch what you can. Proposal-style documents produce dozens of claims — one /verify/batch call, not a loop.