AI Projects

Epic Mountain

Site link →

Objective: build a financial reasoning system that catches its own mistakes and flags its own exposure before a lender or investor does, not a repository of last month's numbers.

A monthly review that re-argues the same points from scratch has no memory. I built one that logs every claim, forces a prediction before the actuals are pulled, and never lets a wrong number quietly disappear.

Built inside a $5M–$10M operating business I co-own and help run, across five revenue segments closing monthly with an outsourced bookkeeper. A reasoning layer on top of Claude and Notion: standing context, a ledger of claims that have to be tested rather than re-argued, and a repeatable review procedure, adapted from a published CFO-brain framework and rescaled to the company's real close cadence.

The problem

What a monthly review forgets by the next cycle

01
No memory across cycles
Standing facts get re-derived from scratch every month instead of read once, and challenged only when something changes.
02
Claims go untested
A note in a deck isn't falsifiable. Nothing forces a prediction before the actuals get pulled, so nothing can actually be wrong.
03
Corrections vanish
A wrong number gets quietly fixed, with no trace it was ever stated, tested, or withdrawn.
04
External claims go stale
A figure quoted once to a lender or investor gets repeated indefinitely, never re-checked against what the data actually shows now.
The system

Three parts: standing context, a claims ledger, a repeatable loop

Structure adapted from a published framework for building a CFO brain with Claude and Notion, then rescaled from that piece's daily cadence to this company's actual monthly close with its outsourced bookkeeper.

Standing context
Company Context
The facts that don't change month to month, read before any review starts, not derived from the raw data each time. Corrections get marked inline, not silently edited away.
The claims ledger
Knowledge Ledger
Every observation, assumption, risk, and decision logged as a typed, sourced, dated claim with a status and a next-review date. Nothing sits unexamined by default.
The procedure
Monthly Review Loop
The repeatable sequence run every month once the books close: check the ledger, predict, pull the numbers, score the prediction, log what's known and what isn't, carry forward.
The mechanism

State the prediction before you pull the actuals

The single rule that makes the ledger more than a note-taking habit. Before the closed month's numbers get opened, the system writes down what's expected, by segment and direction, based on what's already logged. Then the actuals confirm it or refute it. A note can't be wrong. A prediction can.

A live entry on the ledger right now: a revenue range for the coming season, logged and locked in August, against a scenario band built from historical seasonal drivers. Nobody touches those numbers again until the season closes, when the actual result scores the prediction as confirmed or refuted, not restated as something that was known all along.

The same discipline runs on smaller claims too. A margin figure gets logged, then challenged the next cycle against fresher data, and either holds or gets marked withdrawn, in view, with the reason attached.

What makes it durable

Not just built. Built to keep being right, or say so when it isn't

The same three constraints that make Alice's version of this durable, applied here to financial claims instead of relationship data.

Nothing gets deleted, only its status changes
A wrong claim stays visible with the correction attached, not erased. The record of being wrong is part of what makes the record trustworthy.
Every claim carries a next-review date
Nothing stays open indefinitely by default. A claim without a review date is a claim nobody owns.
Evidence is required to move to Confirmed
An opinion or a note doesn't earn Confirmed status. Only a number or a quote behind the claim does.
What it's caught

Three findings the ledger surfaced, not the other way around

No dollar figures here, by design, since this is a live operating business. The shape of each catch is the point.

Withdrawn
A margin claim
Logged, then pulled eleven days later when the underlying cost data proved unreliable. Marked withdrawn on the page, not deleted.
Restated
A rate calculation
Corrected from a misleading line average to a properly weighted figure, and the conclusion it supported reversed.
Flagged
Two external claims
Checked against transaction data and found unsupportable, surfaced before either reached a lender or an investor.

Each of these is the system catching its own error and showing the correction, not a feature that made the numbers look better.

Scale

A live operating business, not a demo

$5M–$10M

Five revenue segments, a monthly close with an outsourced bookkeeper, a founder-operator and a board of investors who see the output. The system runs on the same close cadence the business actually has, not a cadence built for a demo.

Companies outgrow their decision systems before they outgrow their markets.
Matt Boone AI Projects