Markdown, text, csv, Word or Excel.

# CREDIT MEMO: [[BORROWER_2]] o/a [[BORROWER_1]]

Recommendation: approve subject to conditions.

The borrower operates from [[PROPERTY_1]] in [[PLACE_2]]. Revenue has grown from
$22,400,000 in FY2023 to $31,100,000 in FY2025, and the appraisal supports an
as is value of $6,200,000 at a 6.25% capitalization rate.

Debt service coverage of 1.42x on FY2025 supports the $1,850,000 equipment
loan for two CNC press brakes from [[VENDOR_1]], advanced at 85% of invoice cost.

GUARANTOR_1 provides a personal guarantee. A search of [[BORROWER_3]] should
be completed prior to closing.

Conditions: confirmation of CRA account standing under [[BN_1]], and Royal Bank
of Canada to confirm its postponement on the equipment.

# CREDIT MEMO: 2419887 Ontario Inc. o/a Northline Fabricating

Recommendation: approve subject to conditions.

The borrower operates from 118 Passmore Avenue in Toronto. Revenue has grown from
$22,400,000 in FY2023 to $31,100,000 in FY2025, and the appraisal supports an
as is value of $6,200,000 at a 6.25% capitalization rate.

Debt service coverage of 1.42x on FY2025 supports the $1,850,000 equipment
loan for two CNC press brakes from Halton Machine Tools Ltd., advanced at 85% of invoice cost.

Peter Vasilenko provides a personal guarantee. A search of [[BORROWER_3]] should
be completed prior to closing.

Conditions: confirmation of CRA account standing under 812345673RC0001, and Royal Bank
of Canada to confirm its postponement on the equipment.

# CREDIT MEMO: 2419887 Ontario Inc. o/a Northline Fabricating

Recommendation: approve subject to conditions.

The borrower operates from 118 Passmore Avenue in Toronto. Revenue has grown from
$22,400,000 in FY2023 to $31,100,000 in FY2025, and the appraisal supports an
as is value of $6,200,000 at a 6.25% capitalization rate.

Debt service coverage of 1.42x on FY2025 supports the $1,850,000 equipment
loan for two CNC press brakes from Halton Machine Tools Ltd., advanced at 85% of invoice cost.

Peter Vasilenko provides a personal guarantee. A search of 2419887 Ontario Inc. should
be completed prior to closing.

Conditions: confirmation of CRA account standing under 812345673RC0001, and Royal Bank
of Canada to confirm its postponement on the equipment.

The four things banks now ask for, kept on this deal. Ava records what it did. You attest what only you can know.

Download deal record (PDF) Save it to this deal’s 05_Records folder

Saved to /Users/matt/Desktop/Ava_Deals/TC-26-0014/05_Records/TC-26-0014_Deal_Record_2026-10-16.pdf

Available any time. Open pillars show as open on the record. It holds no client content.

1. Set the Deal ID. An existing Deal ID picks up its token map from the deal folder automatically, so the same party keeps the same token across days.
2. Seed the names. Type in the borrower, guarantors, property address and any other party you already know. Include nicknames on separate rows.
3. Scrub. Upload the deal documents. The token map and the scrubbed files are written to the deal folder for you. Nothing to download, nothing to file.
4. Analyze. Upload the scrubbed files to Claude or Gemini. Paste the prompt preamble above your instructions so the model returns the tokens intact.
5. Restore. Bring the finished document back here with the same Deal ID in the sidebar. Check the validation report before it goes to a client.

The coloured message gives the verdict. The table under it lists every token with the name it was restored as, worst first. Read the “Restored as” column: it is where a wrong name shows up.

• No orphans. Every token came back and was replaced. Ship it.
• Tokens that do not appear in this document. The token map covers the whole deal, so a document about part of it will not use every token. Check that no party the document should name is on the list: that is how a model that dropped or paraphrased a party shows up.
• Not in the deal map. The model invented a token or mangled one past recognition. Fix this before the file leaves the office.
• Restored from bare form. The model dropped the brackets, for example BORROWER_1. The name was put back; read that passage before sending.

Dollar amounts, rates, ratios, dates and fiscal periods are never touched. That is the point: the model needs the real numbers to reason, and the numbers are not what identifies your client.

Standard credit vocabulary is also left alone. LTV, DSCR and NOI stay as written. If a real company trades under an acronym, seed it and it will be tokenized like any other party.

Every scrub and restore adds one line to ava_access_record.csv at the vault root: when, who, which deal, which file, a fingerprint of the file, counts and the outcome. Never any document text, names or tokens.

The deal’s token map on screen, read only: each token, the real value behind it, its role and how it was found. Press Show to see it and Hide when done; each view is noted in the access record. Nothing on it is saved or downloaded, and no table in Ava offers a download. Close it before anyone else can see the screen.

The four things banks now ask for, kept on each deal: which AI tools touched the file, who used them and when, whether client data was kept out of the AI, and whether a person approved the work before a lender saw it.

• AVA lines are written by Ava from its access record. Nothing to fill in.
• MATT lines are yours: choose the AI from the approved list, give the date of the AI session, say how the scrubbed files were checked. Pillar 04, your approval, opens once 01 to 03 are complete, and reopens if the work changes after it.
• Download deal record (PDF) gives the one page for the lender, at any stage. Open pillars print as open. It holds no client content: Ava checks the page against the deal’s token map and will not print it if a name or value from the deal would appear on it.
• Choose your AI once, at the top of the tab, under Which AI do you use?: tap each AI you use and its plan, then Save. On each deal, pillar 01 shows them as chips; tap the ones used (more than one is fine). A business plan is rated Low and a personal plan Medium.

Everything runs on this Mac: reading files, reading scanned pages (Apple Vision, part of macOS), detection and substitution. No document leaves the Mac to be read. Only tokenized text is ever uploaded to an AI provider, and only by you.

A scanned page is read on this Mac. A page the reader is not confident of is refused whole. A line it is unsure of is withheld from the AI as [[UNCLEAR_n]]. A total that does not add up is listed, because a scanner can misread a digit. The scrubbed file tells the AI its figures came from a scan.