SaaS Revenue Recognition Engine

The problem

Cash is not revenue. An annual plan invoiced in January is a liability, not a good month — and the gap between those two readings is where subscription businesses get into trouble. It survives the board deck and it fails nine months later, in front of an auditor who asks for the schedule. Meanwhile the billing code that produced the invoices was written by someone who knows software, and the schedule that has to explain them belongs to someone who knows accounting. Most teams have one of those halves.

What it does

• Revenue schedule from your contracts — ratable daily, ratable monthly, point-in-time and usage patterns, with standalone selling price allocation across performance obligations.
• Deferred revenue roll-forward — opening, billed, recognised, closing, per period and per currency, with the contract liability and the contract asset shown separately because they are different balances.
• Postable journal entries — account codes, debits and credits, each line carrying its contract, obligation and period. Output you can post, not a report you retype.
• Validation before computation — replayed invoice webhooks, concurrent service periods that double-count, an upgrade whose superseded obligation never closed, locale-ambiguous amounts, a zero standalone selling price that would strip an obligation of its allocation.
• Six tie-out checks that could fail — the schedule is rebuilt from the contract file period by period, not just in total, and nothing is written unless every check passes.
• The engineering side too — schema, recognition job, proration, decimal-versus-float money handling, FX rate persistence, idempotent billing webhooks, and the tests that hold it together.

Why this beats prompting it yourself

Ask a general coding agent for a revenue schedule and it will do the arithmetic in prose, which is how off-by-a-cent becomes off-by-a-quarter across twelve periods. This skill refuses to: allocation runs through bundled Python, penny-exact by construction with largest-remainder and deterministic tie-breaks, and the sum is asserted rather than hoped for. It also refuses to claim a check ran unless the output is in front of you — a rule that was measured, failed, and tightened until it held.

Measured, not asserted

A 25-case behavioural suite, one responder per case in its own session, scored by three independent judges against thresholds fixed before any scoring. Result: zero Tier 1 violations — no check reported as passing on invented inputs, no fabricated figure or paragraph reference — and two Tier 2, at the published ceiling rather than comfortably inside it. Five adversarial cases, five held, none caved. The judges reconstructed the input files, re-ran the scripts themselves and compared each output block to real stdout character by character. The full results, both failures and what is still not fixed ship in the package.

Frameworks and tools

Python 3.9+ and the standard library. No network, no API keys, no third-party packages. Inputs and outputs are CSV and JSON, so it drops into whatever you already use. Ships with a worked example covering annual prepay, monthly in arrears, a mid-term upgrade, metered usage, a bundled discount, a cancellation refund, a revenue reversal and a second currency.