Ava 1.0
Torcal Credit Inc. — strip identifiers before analysis. Put them back after.
1. Name the parties you already know
Seeded names are always tokenized, whatever the model thinks. This is the single biggest accuracy gain available, so fill it in before uploading. Add nicknames and short forms on their own rows.
| Name in document | Role | |
|---|---|---|
| 0 | 2419887 Ontario Inc. | BORROWER |
| 1 | Northline Fabricating | BORROWER |
| 2 | Peter Vasilenko | GUARANTOR |
| 3 | 118 Passmore Avenue | PROPERTY |
| 4 | Add a name | BORROWER |
2. Upload the deal documents
Supported: pdf, docx, xlsx, pptx, csv, tsv, txt, md, html, eml, msg, jpg, png, tif
Drag and drop files here · 200MB per file
Northline_CreditSummary_2026.pdf×
Northline_FinancialStatements_FY2025.docx×
Passmore_Appraisal_Feb2026.pdf×
Writes to the deal vault
Loading language model…
2
documents scrubbed
1
blocked
14
identifiers tokenized
0
figures altered
Northline_CreditSummary_2026.pdf — scrubbed. 9 identifiers tokenized across 6 pages.
Read with PyMuPDF (local text layer). 6 pages, no OCR required.
| Token | Value | Role | Source | Score |
|---|---|---|---|---|
| [[BORROWER_1]] | 2419887 Ontario Inc. | BORROWER | seeded | 1.00 |
| [[BORROWER_2]] | Northline Fabricating | BORROWER | seeded | 1.00 |
| [[GUARANTOR_1]] | Peter Vasilenko | GUARANTOR | seeded | 1.00 |
| [[PROPERTY_1]] | 118 Passmore Avenue | PROPERTY | seeded | 1.00 |
| [[ADDRESS_1]] | 44 Meadowvale Street, Scarborough ON | ADDRESS | detected | 0.90 |
| [[POSTAL_1]] | M1L 3T2 | POSTAL | detected | 0.75 |
| [[SIN_1]] | 046 454 286 | SIN | detected | 0.95 |
| [[BN_1]] | 812345678 RC0001 | BN | detected | 0.90 |
| [[ACCOUNT_1]] | 00432-003-1234567890 | ACCOUNT | detected | 0.85 |
Scrubbed output
# COMMERCIAL CREDIT SUMMARY Borrower [[BORROWER_1]] o/a [[BORROWER_2]] Guarantor [[GUARANTOR_1]], [[ADDRESS_1]] [[POSTAL_1]] SIN [[SIN_1]] CRA business no. [[BN_1]] Banking [[ACCOUNT_1]] Subject property [[PROPERTY_1]], Toronto ## Facility Loan amount $4,250,000 Appraised value $6,200,000 LTV 68.5% DSCR 1.42x Rate 6.85% Amortization 25 years Closing March 14, 2026 ## Operating results Revenue 2023 $22,400,000 Revenue 2024 $28,900,000 Revenue 2025 $31,100,000 NOI $512,400 Covenant DSCR minimum 1.25x, tested quarterly
Save the token map to /Users/matt/Torcal/Deals/TC-26-0014 — beside the client documents, not in Downloads. It is the only file that contains real names.
Northline_FinancialStatements_FY2025.docx — scrubbed. 5 identifiers tokenized.
Coverage check: page 3 carried a large image region and was flagged for review.
Read with python-docx. Tables, headers and footers included.
| Token | Value | Role | Source | Score |
|---|---|---|---|---|
| [[BORROWER_1]] | 2419887 Ontario Inc. | BORROWER | seeded | 1.00 |
| [[BORROWER_2]] | Northline Fabricating | BORROWER | seeded | 1.00 |
| [[ENTITY_1]] | Bevan & Cole LLP | ENTITY | detected | 0.85 |
| [[EMAIL_1]] | peter@northlinefab.ca | detected | 1.00 | |
| [[PHONE_1]] | 416-555-0142 | PHONE | detected | 0.85 |
Tokens continue from the same registry. [[BORROWER_1]] is the same party here as in the credit summary, because the deal was continued rather than started fresh.
Passmore_Appraisal_Feb2026.pdf was NOT scrubbed. Cannot be read safely: page(s) 12, 13, 14 appear scanned; page(s) 4 carry a large image region that returned no text. Identifiers in those regions would pass through un-tokenized. Configure Document Intelligence, then scrub this document again.
No partial output is produced for a blocked document. There is no override.
3. Take it to the model
Paste this above your own instructions. It tells the model to return the tokens intact.
This document has had identifying details replaced with tokens of the form [[ROLE_N]], for example [[BORROWER_1]]. Treat each token as a stable name for that party. Reproduce every token exactly as written, including both pairs of square brackets. Do not expand, guess at, or invent names for them, and do not create new tokens. All figures, rates, ratios and dates are real and unmodified.
Next: do the analysis in Claude or Gemini as you normally would, then bring the finished document back to the Restore tab.
1. Load the token map
The map produced when the deal was scrubbed. It is the only file that can resolve the tokens.
TC-26-0014_tokenmap.json · 14 tokens loaded
2. Paste the model’s output
Whatever came back from Claude or Gemini. Markdown, Word or Excel.
Nothing is written until you download
Matching tokens…
12
substitutions made
1
token never appeared
1
invented token
Do not send this file yet. The model produced a token that was never issued. Review the validation report below before this leaves the office.
Validation report
| Token | Outcome | What it means |
|---|---|---|
| [[BORROWER_1]] | 4 substitutions | Resolved to the registry value. |
| [[BORROWER_2]] | 2 substitutions | Resolved to the registry value. |
| [[GUARANTOR_1]] | 3 substitutions | Resolved to the registry value. |
| [[PROPERTY_1]] | 2 substitutions | Resolved to the registry value. |
| [[BN_1]] | 1 substitution | Resolved to the registry value. |
| [[POSTAL_1]] | Never appeared | The model dropped or paraphrased it. Usually harmless. Confirm the party was not described in an identifying way. |
| [[BORROWER_3]] | Not in registry | The model invented a token. It resolves to nothing and has been left in place rather than guessed at. Fix before the file leaves the office. |
Restored document
# 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 2023 to $31,100,000 in 2025, a compound rate of approximately 17.9%. NOI of $512,400 supports the requested facility of $4,250,000 at 68.5% LTV against an appraised value of $6,200,000. Debt service coverage of 1.42x provides headroom against the 1.25x covenant tested quarterly. At the contract rate of 6.85% over a 25 year amortization, coverage remains above covenant on a 200bp rate shock. Peter Vasilenko provides a personal guarantee. A search of [[BORROWER_3]] should be completed prior to closing. ← unresolved Conditions: confirmation of the CRA account standing under 812345678 RC0001, and evidence of fire insurance naming the lender as loss payee.
The unresolved token is reported rather than silently removed or guessed at. A tool that quietly cleaned this up would hide the fact that the model asserted something about a party that does not exist.
The workflow
Six steps. The working day barely changes.
- Set the Deal ID. New deal, take the default and change the number. Adding to a deal scrubbed earlier, also load its existing token map so the same party keeps the same token.
- Seed the names first. Borrower legal name, trade name, guarantors, property address, any individuals. Short forms and nicknames on their own rows. This is the highest-value minute in the workflow, because seeded names are tokenized unconditionally and everything else depends on the model’s judgment.
- Upload and scrub. The first scrub of a session takes about a minute while the language model loads. Later ones are fast.
- Review the detection table. The step people skip and should not. Two questions: did anything real get missed, and did anything ordinary get tokenized that should have been left alone.
- Analyse. Take the scrubbed file into Claude or Gemini and work exactly as you normally would, with full use of everything you already pay for. Paste the preamble above your own instructions.
- Restore. Bring the output and the token map back. Read the validation report before the file goes anywhere.
What is deliberately never tokenized
Dollar amounts, rates, ratios, dates and fiscal periods. The model needs real numbers to reason, and numbers are not what identifies a client.
Standard credit and accounting vocabulary. LTV, DSCR, NOI, engagement types, government program names and financial statement captions all stay as written. Tokenizing them is an analysis loss with no privacy gain.
Limits worth remembering
An unread page is an unscrubbed page. In v3.1 that blocks the document, with no override.
Ava removes direct identifiers. It does not make a deal unrecognizable. A $4.25M first mortgage on an industrial property in a named suburb closing in March is still recognizable to anyone active in that market.
Restore returns the canonical spelling where a party appears under several. Variants are recorded in the token map.