Close Commenter is a month-end close review tool where a person has to read the numbers and write the commentary in their own words. The AI finds what to look at and polishes the writing afterwards. It never approves anything and it never gets to introduce a number.

Close tools can already classify, match and accrue. The part nobody wants to automate away is the human sign-off: someone who understood what happened and can explain it to an auditor or a board. Close Commenter is built around that part.

How it works

• Review: 409 synthetic transactions for September 2026, each with code-computed flags and a fast AI anomaly score.
• Batch approve with a gate: one click approves only what passes a safety gate (no flags, under the cap, known vendor, normal range, preparer is not approver). The server re-runs the gate, so the browser cannot override it. Held items come back with reasons.
• Route and decide: send items to someone else with a note, cancel an accounting-only entry, request a refund on a duplicate payment, or approve with a resolution such as a 12-month prepayment. The ledger export is never edited; every insight shows the number as booked and as adjusted.
• Two ways to review: a table with an evidence panel, or cards, one held item at a time with its reasons in words.
• Insights, then your words: the app proposes a few board-level insights. The Enhance button stays locked until you have written at least 25 words yourself.
• Number guard: every figure in AI text is checked against the numbers the code computed and the numbers in your own draft. Anything else is highlighted as unverified and cannot be pinned.
• Deck: a 16:9 close deck with P&L against budget and prior month, an operating-expense bridge and one slide per insight.
• Anchor on Hedera: finalizing hashes the close package, encrypts it and writes it to a Hedera Consensus Service topic (testnet). Anyone can verify integrity from the public Mirror Node, and the key holder can restore the approved close without the accounting system.

Who does what

• Code computes every number, every flag, the batch gate and the close package. Nothing is delegated to a model.
• Jev (TypeSafe System One) gives an anomaly probability per transaction in about 150 ms. It is one signal next to the rules and cannot approve anything.
• Claude writes the narrative from the facts and polishes your draft. It may not introduce a figure that is not in the facts or your draft.
• The human approves, routes, writes the commentary and finalizes.

Measured on the planted data: the seed plants 8 anomalies. Rules find 8 of 8; Jev at a 0.5 cut-off flags 5 of 8 and 5.5% of the other transactions. This is a small synthetic test, not a detection rate.

Built in one day by Salvador Aveldaño. All data is synthetic (“Northgate Labs”), Hedera usage is testnet only, and the README carries the full AI and tool disclosure.