A bake-off is only as fair as its questions. A firm that builds a small, fixed set of real tasks with known answers can compare any model, today or next year, and see whether a change helped. The set is called a gold set because the answers are settled. Closed engagements are the best source, since the work was done, reviewed and signed.
The general testing approach is in how to evaluate an LLM for audit work. This post covers the part that is often skipped: building and keeping the test material.
Why not use vendor demos
A demo uses clean, chosen inputs. Your files have scanned pages, mixed formats and odd narrations. A model that shines on a demo invoice may stumble on a faded thermal-paper bill. A gold set built from your own history shows what your staff will actually face.
What goes in the set
Aim for variety rather than volume. Forty to sixty items split across task types is enough to see differences. Suggested buckets:
| Bucket | Example item | Known answer |
|---|---|---|
| Extraction | Invoice with GSTIN, date, taxable value | The fields as verified |
| Classification | Ledger narration | The account head your team assigned |
| Reading | Page of a loan agreement | Covenant text and clause number |
| Reconciliation | Two lists with known differences | The list of differences |
| Anomaly | Ledger extract with planted duplicates | The planted items |
| Not in file | Question the file cannot answer | "Not found" |
| Tricky language | Mixed English and regional text | Verified transcription |
Include a few items at which you expect a model to fail. A set that every model passes tells you nothing.
Anonymise before you copy
Closed engagement material still belongs to the client. Before any item enters the set:
- Remove names, PAN, GSTIN, bank account numbers, addresses and signatures. Replace with consistent dummy values so the structure remains.
- Keep amounts that matter to the test, but scramble where they could identify a client.
- Check the engagement letter and your confidentiality policy for limits on reuse.
- Store the set in a restricted folder, separate from live engagements.
- Keep a key mapping dummy to real values only if the firm needs it, and restrict that too.
The method for stripping identifiers from prompts is similar to that in the DPDP-safe prompt library. If the test items go to an external AI tool, remember that the same data-handling questions apply to test material.
Write the answer key properly
For each item, record:
- The input file and the exact question
- The correct answer, with the source page or reference
- Who confirmed it and when
- A tolerance, such as "date must match exactly; amount within rounding"
A second person should check the key. Errors in the key quietly penalise the better model.
Scoring that survives scrutiny
Use simple, fixed scores:
- Correct with correct source: 2
- Correct without a usable source: 1
- Wrong or invented: 0
- Correctly says not found, where nothing exists: 2
Separate scores by bucket rather than one total. A model strong at extraction and weak at reading should be used for extraction, as the task-by-task guide suggests. Weight invented answers heavily in your reading. One confident fabrication in a material item matters more than three small misses.
Run every item at least twice. Output varies from run to run, and a single lucky pass proves little. Note the date, the exact model name and any settings. The model versioning post explains why this record matters.
Keep it fixed, with a small moving part
Two sets are better than one:
- The fixed set never changes, so you can compare results across months and across models.
- A rotating set takes new items from recent engagements, which stops tools being tuned to your old questions and keeps the test current.
If you change the fixed set, label it as a new version and rerun the earlier models.
Guard against test leakage
If test documents are pasted into a consumer chat tool that keeps history for training, the set can end up outside your control. Use tools and settings whose data terms you have read. Do not publish the set or share it with vendors. A vendor that has seen your test items can improve its demo without improving the product.
Involve reviewers, not only enthusiasts
Have an audit manager who is sceptical of AI score a few of the outputs blind, without knowing which model produced them. People who like a tool read its output generously. Blind scoring removes some of that.
What the set cannot tell you
A gold set measures performance on past tasks. It does not prove a model is safe for a client whose data differs, nor that a future update will behave the same. It also does not replace partner review of live work. Think of it as a smoke test and a way to make buying and switching decisions with evidence, not impressions.
Record what you decided
After each bake-off, write one page: models tested, dates, results by bucket, decision, approver and review date. File it with the firm's AI policy. Human-in-the-loop review design covers where the person checks work once a model is in use.
Frequently Asked Questions
How big should the set be?
Forty to sixty varied items is enough to spot clear differences. More is better only if the answer key stays reliable.
Can we use live client data for testing?
Use closed engagements and anonymise them. Check the engagement letter and your confidentiality rules first.
How often should we rerun it?
When a model or tool changes, and on a fixed schedule such as every quarter.
Last reviewed: 17 October 2026. Facts checked. No external source was opened in this session. The post describes a testing method, not a finding, and quotes no benchmark, price, release date or model ranking.
Related Articles
Statutory facts on this page are checked against their sources, and the page says where it relied on secondary reporting. How we verify · Report an error