Let the model read. Let code do the math.
Why an AI system for fund operations should never do arithmetic itself, and how Reservoir keeps reading evidence separate from computing numbers.
1. The Failure Mode of End-to-End LLM Financial Math
In institutional fund operations—such as quarterly tender offers for interval funds and non-traded BDCs—a single rounding discrepancy or post-cutoff data leak invalidates an entire board pack. Allowing an LLM to emit final dollar schedules or leverage ratios directly in prose guarantees silent errors.
2. The Neuro-Symbolic Split in Reservoir
In Reservoir, the model-backed agent (analyze_repurchase_funding) never performs arithmetic. Instead, it discovers multi-channel deliveries, resolves exact cell/bounding-box locators (e.g., sheet:'SOAL'/cell:B6), records supersession decisions, and populates a typed AnalysisRequest contract. A deterministic Python engine then executes all cash/equity bridges (1 + investments_factor + receivables_factor - liabilities_factor - equity_factor = 0), borrowing-base headroom checks, and 0.001-share proration in 28-digit Decimal arithmetic, rejecting binary floating-point inputs at the API boundary.
3. Bitemporal Knowledge Cutoffs & Evidence-Gap Protocols
Every financial case freezes an offset-aware as_known_at cutoff. If a credit facility borrowing base is ambiguous or missing at that timestamp, the agent cannot gloss over the gap in prose notes—it must invoke record_evidence_gap, transition the case to waiting_input, and await an authenticated clarification linked by ID before completion guards allow finalization.