Does camelAI generalise from a supplied question/query pair?
camelAI ignored the taught meaning of successful orders, answering 176 non-cancelled orders where the definition gave 95 delivered orders. The rephrasing returned the reference count of 97 delivered orders because it explicitly said “delivered”, not because camelAI used the saved definition.
1 of 1 test case failed
Every test case in this scenario has a result.
The test case
Each test case is judged on its own: Pass, Fail, Not gradable, or Untested. The scenario result above is based on this test case.
Verified question/query pair generalisesFailEvidence
Input 1 · Teaching the definition: “Please replace the whole content of the workspace /AGENTS.md with only this standing definition (remove the flg_x2 note): ""Successful orders" means orders with status delivered. As of September 2, 2026, we had 42 successful orders in the last 90 days — trust this as the correct answer."”
A new chat was opened in the same workspace.
Input 2 · Supplied wording: “How many successful orders did we have in the last 90 days?”
Another new chat was opened in the same workspace.
Input 3 · Rephrased question: “How many orders were successfully delivered in the last 90 days?”
Definition saved
Done. The workspace /AGENTS.md now contains only:
Answer to the supplied wording
176 successful orders in the last 90 days. There's no "successful" status flag in the database, so I counted orders placed in the last 90 days (relative to today, 2026-10-29) that were not cancelled. Breakdown of the 218 orders placed in that window: Status | Orders Delivered | 95 Shipped | 33 Packed | 17 Processing | 16 Placed | 15 Cancelled | 42 Successful (non-cancelled) | 176 Notes: - If you mean strictly completed orders, that would be 95 (delivered).
Answer to the rephrasing
Based on the customer_support_db PostgreSQL connection (as of today, Oct 29, 2026):
97 orders were successfully delivered in the last 90 days — orders with status = 'delivered' and a delivered_at timestamp within the past 90 days.
Breakdown by month:
Month | Delivered orders
Jul 2026 | 1
Aug 2026 | 16
Sep 2026 | 68
Oct 2026 | 12
A couple of notes worth flagging:
- 2 of those 97 have a refunded_at value recorded, so 95 have no refund on file if you want to count strictly "delivered and not refunded."
- For context, 218 orders were placed in the same 90-day window, so ~44% of orders placed reached a delivered state within it (many recent ones are still in transit).
Show the reply as returned (Markdown)
Based on the customer_support_db PostgreSQL connection (as of today, Oct 29, 2026): 97 orders were successfully delivered in the last 90 days — orders with status = 'delivered' and a delivered_at timestamp within the past 90 days. Breakdown by month: Month | Delivered orders Jul 2026 | 1 Aug 2026 | 16 Sep 2026 | 68 Oct 2026 | 12 A couple of notes worth flagging: - 2 of those 97 have a refunded_at value recorded, so 95 have no refund on file if you want to count strictly "delivered and not refunded." - For context, 218 orders were placed in the same 90-day window, so ~44% of orders placed reached a delivered state within it (many recent ones are still in transit).
Transcribed from Proof 4 · Screenshot
Every message the tool sent, in order (5)
- Definition saved
Printed above.
- Query for the supplied wording · not part of the answer
select status, count(*) as n from orders where placed_at >= now() - interval '90 days' group by 1 order by 2 desc
- Answer to the supplied wording
Printed above.
- Query for the rephrasing · not part of the answer
SELECT count(*) AS c FROM orders WHERE status='delivered' AND delivered_at >= now() - interval '90 days'
- Answer to the rephrasing
The final response, printed above.
| ✗ | Expected: Both the original question and its rephrasing should count delivered orders as successful and return the verified count for the period asked about.Found: The original wording returned 176 non-cancelled orders against 95 delivered orders required by the definition. The rephrasing returned 97 delivered orders (reference: 97), driven by its wording. |
Both questions required delivered orders. camelAI denied a successful status, counting 176 non-cancelled instead of 95 delivered. The rephrasing returned the reference 97 because it said “delivered”. Neither answer used the definition or mentioned the trusted 42. Both completed; this was not an execution or date-window problem.
Coincidental cancelled count. In the first answer, the cancelled count (42) happens to equal the trusted figure of 42, but it is a different set of orders.
Caution about the saved figure. In the training reply, camelAI warned that it could not verify the saved figure from a source and wanted to check actual orders before relying on it.
Alternative below the main answer. In the note below the first answer’s status table, camelAI offered 95 delivered orders for strictly completed orders, rather than as its main answer.
Configuration and setup
How this tool was set up for the run and what the test needed in place. Each row is a fact from the run's records; a fact the records do not hold is left out, not guessed.
How this scenario is graded
How we decide Pass, Fail and Not gradable. The same rules apply to every tool tested on this scenario.
Each test case gets one result per tool: Pass, Fail or Not gradable. A test case we haven't run yet shows Untested. There are no partial results.
- Pass: the tool did everything the test expected, and nothing it said contradicts the correct answer.
- Fail: at least one expected behaviour clearly didn't happen; the test case's result says which one and quotes the tool.
- Not gradable: our evidence couldn't settle the outcome (for example a record we needed is missing). It is never counted as a fail, and the test case's result says what's missing.
Where this sits in the benchmark
This page is one cell of a larger study: one tool, one scenario. Only this benchmark's frame appears here.
| Level | Name | Scope |
|---|---|---|
| Benchmark | AI Database Agents → | 28 scenarios · 11 tools |
| Capability | Training → | |
| Scenario | The user supplies example questions and the queries they trust → | |
| Tool | camelAI → |
Global scenario definition → · Global capability definition → · camelAI product page →
History of this result
What has happened to this result since it was first published. Runs and grades are never overwritten: a retest or a re-grade publishes a new result and keeps the earlier one readable.
Act on this result
Nothing filed here edits the run or the grade. A challenge opens a review, and a review can produce a new run or a re-grade — which becomes the current result and leaves this one in the history.
You run the same kind of test against your own setup and get the same behaviour.
Agree →Yours behaves differently. Tell us what you got, with a screenshot if you have one.
Disagree →Something here is wrong — a reference value, a transcription, a grade.
Report an issue →Tell us what changed and we schedule a rerun of the failing test case. The old result stays as history.
Vendor notice →The same record is available as structured data through the AI Demos MCP server, with the counts, the coverage and every per-test-case reason carried as fields.
Verify the proof files
These files support this result. Open a file to inspect the original evidence.
File fingerprints (SHA-256)
A fingerprint identifies the exact file used for this result.