Beta Product brief
How OrbaLM works
OrbaLM is not a foundation model and not a public chatbot. It is a governed specialist runtime that sits in front of a language model you approve — so treasury and FX desks get usable answers without invented capital or PnL.
One sentence
A model-agnostic control plane and Treasury/FX domain pack for private specialist review: policy, deterministic tools, and provenance around your hosted or private-cloud model.
What happens when you ask
- You ask in private chat or from the Excel task pane (selection, sheet, or full workbook).
- Policy runs first — the runtime refuses to invent official capital, live PnL, or fabricated market figures.
- Governed tools calculate when the question matches — FX options, bonds, forwards, IRS, repo, caps/floors, NDFs, money markets, FRAs, FRNs, CCS, amortising debt, commodity hedges, and hedge-effectiveness diagnostics where the pack enables them. Numbers come from tools, not from model prose.
- The model explains conventions and context around those results. Every answer can carry pack version, tool status, and evidence pointers.
What you see
- Private chat — local runtime UI with thread history
- Excel task pane — ask from the workbook you already use
- Word and PowerPoint task panes — select-and-ask prototypes moving through pilot packaging
- Tool reports — when a calc tool runs, the answer is labelled as assembled from tool output (not free-form invention)
What Beta is — and is not
Included
- Desk runtime + Treasury/FX pack
- Chat and Excel workbook flows
- Governed calc tools (as enabled by the pack)
- Local provenance on answers
Not included
- Online purchase / self-serve checkout
- Enterprise multi-user ops
- Live terminal connectors by default
- Pack updates after the 3-month founding window
How it treats numbers
Every figure comes from a deterministic tool, and the runtime refuses to fabricate the ones it cannot compute. Alongside each result you get discrete scenario tables (rate and volatility shifts), sensitivities (option Greeks), and data-quality halts — for example, an FX-forward MTM stops rather than guessing when the interest-rate parity gap exceeds tolerance.
For FX options, forwards, bonds, and interest-rate swaps, the report also includes a bounded scenario envelope: a full reprice under documented one-factor input shocks, with optional independent multi-factor corner checks. You see low / base / high and which factor dominates the width.
Still not a confidence interval: the envelope is a deterministic reprice band (one-factor plus independent ± corners), not a probabilistic, correlated, or model-error interval. FX-forward MTM can still halt on data-quality gaps.
How evidence stays current
First-party desk literature ships with the estate: the final FX Options book (788 pages, 7×10) is indexed for section-aware, retrieval-only evidence. Separately, OrbaLM checks topic-specific allowlists of official financial and regulatory sources. New findings are deduplicated and staged with their source URL, retrieval date, content digest, and rights status.
Staging is not publication. Proposed additions remain retrieval-only and enter a review branch before they can become trusted evidence. The process does not silently retrain a model, alter governed calculations, or change the live domain pack.
- Owner-controlled FX Options PDF indexed on the private estate (not for training)
- Coverage includes treasury, markets, risk, regulation, and financial technology
- Official-domain restrictions apply before a harvested result is accepted
- Repeated sources are suppressed by URL and content digest
- Every proposed harvest update remains subject to human review
Where it runs
Designed for your estate: laptop / air-gap Ollama pilots, or a VPC container pointed at your approved model endpoint. The marketing site is only for Beta inquiry — the product itself is not a multi-tenant public SaaS by default.