Auditable Ratios.
A real recovery, every number from the actual engine run.
Larson Industries is the fixture client that runs through every demo in the CFO Operating System. This page shows what the ratio engine surfaced about Larson’s recovery from FY2024 to FY2025 — what it computed, what it flagged as pending, and how it turned the diagnostic into a recommendation a banker could read.
Watch the score move: a six-month trajectory + a 30-day forecast
The video above walks the methodology. This one puts it in motion on a single company: six months of QuickBooks, the Survival Score falling into a Q1 cash squeeze and recovering — and a 30-day cash forecast fitted from the same books. Metrics and forecast, one auditable pipeline. (Forecast shown on synthetic demo data.)
| Larson Industries | FY 2024 | FY 2025 |
|---|---|---|
| Revenue | $806,000 | $1,007,882 ↑ |
| Total Equity | −$165,083 | +$208,807 ↑ |
| Current Ratio | 0.79 | 1.33 ↑ |
| Gross Margin | 70.0% | 63.5% ↓ |
| Survival Score (Z″) | 2.30 · caution | 4.90 · very strong |
The four tiers, named honestly.
The engine separates work that should be deterministic from work that benefits from AI synthesis. The line between them is named explicitly on every screen — there’s no LLM in the numbers, and the LLM that summarizes is bounded by a hallucination gate.
What the engine reads.
Roughly twenty inputs from the balance sheet and income statement, pulled from a real QuickBooks Online run. Cash, accounts receivable, inventory, accounts payable, owner’s equity, revenue, gross profit, operating income. Every input is source-hashed — the engine cites both the rule that fired and the specific report row that triggered it.
Defined formulas. Defined thresholds.
Eighteen indicators drawn from David Duryee’s Sixty Minute CFO methodology, with a few additions from our forward-looking work. Each ratio has a defined formula, a defined denominator convention, a defined day-count basis, and a defined band threshold — all of it inspectable in the config file. The headline number is the Survival Score (Altman Z″): 6.56·(WC/TA) + 3.26·(Eq/TA) + 6.72·(EBIT/TA) + 1.05·(Eq/TL).
Endsley’s three levels, bounded by a hallucination gate.
Mica Endsley’s three-level framework applied to financial data. L1 — Perception: the facts. L2 — Comprehension: what the combination means. L3 — Projection: where this is heading. The LLM (Sonnet 4.6) restates the facts grounded in the ratios; a hallucination gate checks that every figure mentioned was present in the ratio output. If a figure wasn’t, the narrative is rejected and regenerated.
The recommendation is the practitioner’s.
Klein’s recognition-primed decision pattern says expert decision-makers don’t generate every option and evaluate — they match the current situation against a library of patterns they’ve seen before, and run the first play that fits. The play library is a versioned YAML file. An OPA policy gate then validates the selected play against the client’s tier, prerequisites, and any standing firm policy.
The audit-trail commitment.
Every screen in the video can be reconstructed from the event log of a single engine run, identified by a trace ID. That’s what makes a methodology engine bankable instead of just impressive.
Deterministic math
Every ratio value comes from a Python module that runs the defined formula on the defined inputs. No LLM is in the path that produces a number a client relies on. Run the engine twice on the same inputs, you get byte-identical outputs.
Source-hashed inputs
Each input is hashed at extraction. Findings cite the rule that fired and the report row that triggered it. If the books change, the hash changes; if the hash changes, the lineage is visible.
Replayable from events
The materialized outputs are projections of an event log. Drop the materialized output, the replay endpoint rebuilds it from the events with no data loss. Audit on demand, months after the engagement.
Want to see this on your own books?
Same engine, same audit trail. The engagement is fixed-scope, fixed-fee, and most clients start with a Balance Sheet Integrity Scan — if the findings warrant the deeper work, the scan fee credits toward the engagement.