Responsible AI & AI Usage

1. The valuation engine is deterministic

Verus's fair-value engine is deterministic — same inputs and same engine version produce the same outputs, no Monte Carlo drift, no learned weights in the pricing path. Generative AI is never used to compute a valuation, a spread, a recovery rate, or any number that appears on a mark. This is a design invariant, not a policy preference, and it is the reason our outputs are reproducible for auditors and regulators.

2. Where we use AI

AI models are used only in narrow, optional, human-reviewed extraction and summarisation steps:

• Deal extraction from documents. When enabled, an LLM reads a credit memo, term sheet, or prospectus you upload and returns a structured draft of facility terms that you confirm or edit before it touches the valuation engine.
• Describe-with-AI prompt flow. A free-text description of a borrower or deal can be turned into a candidate facility set, with optional web-search grounding. The user always confirms the structured output before a deal is created.
• Deal summarisation for review packets. A deterministic report is generated first; an optional LLM pass can produce a plain-English executive summary that sits alongside — never replacing — the quantitative output.

3. Models we use

The model identifier, version, and backend are recorded on every AI-assisted call and surface in the run log. The defaults below describe the shipped configuration; customers can pin a specific model at contract time and we will not swap it without written notice.

• Anthropic Claude — Opus / Sonnet (the latest GA model of the Enterprise API family, pinned per tenant). Selected via ANTHROPIC_MODEL_EXTRACTin the deployment. Under the Enterprise API terms, Anthropic does not train models on customer inputs or outputs and applies a short input-retention window. We record the exact model string returned by the API on every call so the audit trail pins the version.
• Locally hosted private LLM where customers require fully on-premise AI. An open-weights model is served from inside the customer network, pinned per tenant at contract time, and runs deterministically at temperature=0 with a seeded sampler. No data leaves the customer network; the model identifier is recorded on every call so the audit trail pins the exact build in use.

Determinism and prompt-contract versioning. AI calls in Verus run at temperature=0 against versioned prompt templates. Each prompt template carries a SHA-256 hash recorded alongside the model identifier on the audit log. A reviewer can reconstruct exactly which prompt + model + version produced any extracted field. Non-deterministic sampling (e.g. temperature > 0 for summarisation) is only used where the customer explicitly opts in.

4. Human-in-the-loop

Every AI-assisted step produces a structured draft that the user must confirm. A deal is never created, a valuation never run, and a report never finalised without explicit user action. The deterministic engine treats AI-extracted inputs identically to hand-entered ones — the source does not affect the price.

5. Training and retention

• We do not use customer content (deal documents, financial statements, assumption sets, chat transcripts) to train models of any kind.
• Prompts sent to Anthropic are processed under the enterprise API terms, which prohibit training on customer data and delete inputs within a short retention window.
• Audit logs of AI usage (who ran what, when, on which deal) are retained for the life of your contract plus up to six years.

6. Data residency for AI calls

Anthropic's API is hosted in the United States. Cross-border transfers rely on SCCs and the UK IDTA Addendum (see our Data Protection page and the full regional matrix at /legal/data-residency). For customers with strict residency requirements, the locally hosted private-LLM path keeps inference inside your infrastructure.

Our position under the EU AI Act is documented separately at /legal/eu-ai-act: markst is a deployer of a GPAI model (not a provider), and Verus is not a high-risk AI system under Annex III.

7. Opt-out

AI features are a self-service toggle. Go to Settings → AI features inside the app and switch AI-assisted intake off; all /intake/* and /llm/* calls immediately return a 403 and the UI surfaces stop offering AI steps. The core deterministic valuation pipeline is unaffected. No email to support required; the change takes effect on the next request.

The toggle revokes AI on the account that flips it. Organisation- wide default-off provisioning at contract time is still available — contact support@markst.comif you want AI off before your users log in.

8. Transparency and bias

Because AI is confined to extraction and summarisation (not pricing), model bias cannot influence the mark. The extraction layer is nevertheless audited: every AI call logs the model version, prompt, and parsed output, and each extracted field displays its provenance in the UI so reviewers can spot inconsistencies.

9. Questions

AI governance queries: ai@markst.com. General support: support@markst.com.