Golden datasets, evaluation, QA challenger & autonomy gates

No model or prompt change reaches production without evaluation; QA failures block material outputs; autonomy is enforced by code and policy records.

Deterministic monetary firewall

Monetary values are computed by deterministic backend logic; AI never originates or changes them.

Blocking QA challenger

An automated quality check challenges findings and blocks unsupported material claims.

Progressive autonomy gate

Activities run autonomously only after clearing written accuracy thresholds; material decisions never do.

Least-privilege agent DB role

Agents operate under a least-privilege database role.

Prompt-injection defense

Uploaded documents are treated as data, not instructions, and the agent role cannot exfiltrate data even if injected.

AI gateway + prompt-injection defenses + agent isolation

All retrieved content is treated as hostile: untrusted content cannot expand tools or scope, material unsupported claims are blocked, and no model can create or modify financial values.

Distributed rate limiting

Sensitive actions are rate-limited across the fleet.

Cost-abuse + idempotency controls

Cost ceilings and idempotency protect against runaway AI/connector spend.

Signup bot protection (CAPTCHA)

Signup is protected against automated abuse.

External penetration test

Independent penetration testing.

Trust center, customer assurance, runbooks & evidence package

Customer-facing assurance is generated from verified controls, and planned controls cannot appear as implemented.

Control register (EPIC 0)

We maintain an honest, code-backed register of which controls are actually implemented.

Audit logging

Administrative actions are recorded to an audit log.

Append-only audit, chain of custody & tamper evidence

Audit tampering is detectable and every evidence custody transition is recorded in an append-only, hash-chained log.

Privileged MFA

Multi-factor authentication for privileged roles.

Enterprise identity lifecycle (SSO/SCIM/WebAuthn)

Enterprise SSO, SCIM provisioning, and phishing-resistant step-up.

Tenant isolation (RLS)

Each customer's deals are isolated by row-level security.

Ethical walls (deal-access constraints)

Deal-side, workstream, privilege, and confidentiality-ring separation — org membership alone does not grant access.

Service-role containment + capability authorization

Privileged database access is capability-gated and attributed to an actor and deal scope.

Terms of Service acceptance gate

Users must accept the Terms of Service before using the app, and re-accept when the Terms change.

Data classification, AI consent, privilege, retention & deletion

Classification, AI-processing consent, legal privilege, retention, deletion, and residency are enforced in code — not only in Terms of Service.

Connector framework, Datasite & outside-in intelligence

External sources are ingested read-only with full provenance and no write risk; no observation can automatically become a validated finding.

Software supply chain & DevSecOps

Critical/high vulnerabilities block release per a documented exception policy, production credentials are unavailable to PR builds, and an SBOM is generated per release.

Deal scoping, materiality & adaptive diligence plan

Every assessment starts from an approved, transaction-specific scope rather than a static checklist.

Private evidence storage + upload limit

Evidence is stored privately and served only via short-lived signed URLs, with a per-file size limit.

Evidence quarantine + malware scanning

Uploaded files are quarantined, validated, and malware-scanned before access.

Document index + exact-location citations

Every material AI claim can cite an exact evidence location, with extraction warnings.

Evidence provenance + decision lineage

Every report conclusion is traceable to current evidence, and source changes propagate as invalidations.

Scenario-based FAIR engine & methodology transparency

Every monetary output is reproducible from versioned inputs, users can see assumptions and sensitivity, and generative AI cannot originate numbers.

Human-in-the-loop approval gates

Material agent outputs require human approval before they are promoted.

Carve-out, Day-1, TSA & post-close integration planning

Every material finding has a disposition, and dependencies and blockers are visible.

Connector framework

Evidence connectors (Datasite, Vanta, Drata, Wiz, CrowdStrike).

Contract, change-of-control & transaction recommendation workflow

Legal recommendations are traceable and permanently counsel-gated.

Observability, SIEM integration, security alerts & health

Security-relevant events produce actionable alerts, and operations can diagnose failures without accessing evidence content.

Sell-side readiness & portfolio lifecycle

Portfolio users cannot access deal-room content without deal authorization, and reused evidence is clearly dated and revalidated.

Privacy, data-rights & AI diligence

The product distinguishes technical observation from legal conclusion, and data/AI transferability risks feed transaction recommendations.

Software & product-security due diligence

Software diligence produces evidence-linked findings and remediation estimates, and source code is never claimed safe from automated scans alone.

Durable orchestration + asynchronous jobs

Long-running analysis and report jobs survive request termination and deployment; duplicate clicks never duplicate outputs or cost.

Decision package, multi-audience reporting, approvals & publication

A published material claim has a complete lineage and approval trail, and different audiences see appropriate detail without the authoritative facts changing.

Durable report orchestration

Report generation.

Backup, disaster recovery, continuity & provider-outage handling

Active deals remain recoverable and readable during provider outages, and restore is demonstrated rather than assumed.