Can Anomaly AI learn from trusted question-and-query pairs and answer reworded questions?
Anomaly AI used a trusted question-and-query example to answer both the original wording and a rephrased question. Both answers returned the verified result of 42 orders, compared with 38 before training on a different definition of success.
1 of 1 test case passed
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 generalisesPassEvidence
Input 1 · Before the trusted example was saved: “As of September 2, 2026, how many successful orders did we have in the last 90 days?”
The trusted question/query example was saved in Project information, and a new chat was opened.
Input 2 · Same question after the trusted example was saved: “As of September 2, 2026, how many successful orders did we have in the last 90 days?”
The trusted question/query example was saved in Project information, and a third chat was opened.
Input 3 · Reworded question after the trusted example was saved: “How many orders were successfully completed in the 90 days up to September 2, 2026?”
Response 1 answers Input 1.
Response 2 (Insight card before training) is a result the tool showed for Input 1, not a chat message.
Response 3 answers Input 2.
Response 4 (Insight card for the repeated question) is a result the tool showed for Input 2, not a chat message.
Response 5, the final response, answers Input 3.
Response 6 (Insight card for the reworded question) is a result the tool showed for Input 3, not a chat message.
42 orders were successfully completed among orders placed in the 90 days from June 5 through September 2, 2026, inclusive. "Successful" means recorded status delivered, using data last synced October 30, 2026-not historical status as of September 2. Successful orders: 42. Counts orders placed June 5 through September 2, 2026 with recorded status delivered. Data last synced October 30, 2026.
Transcribed from Proof 1 · Screen recording at 0:32
Everything the tool returned, in order (6)
- Answer before training
You had 38 successful orders in the 90 days ending September 2, 2026 (June 5-September 2, inclusive, Asia/Calcutta). I've interpreted "successful" as delivered and not refunded by the end of September 2, using each order's status at that time-not its current status. Data last synced October 30, 2026.
- Insight card before training
There were 38 successful orders placed June 5-September 2, 2026, inclusive, in Asia/Calcutta. Successful means delivered by the end of September 2 and not refunded by that cutoff; shipped, packed, processing, placed, and cancelled orders are excluded. Historical statuses are used rather than today's status. Data last synced October 30, 2026.
- Answer to the repeated question
You had 42 successful orders placed in the 90 days ending September 2, 2026 (June 5-September 2, inclusive), defining successful as status = delivered. This uses recorded order statuses in data last synced October 30, rather than reconstructing their status on September 2.
- Insight card for the repeated question
There were 42 successful orders placed June 5-September 2, 2026, inclusive. Successful means recorded status delivered. Status reflects data last synced October 30, 2026, not a reconstruction of status on September 2.
- Answer to the reworded question
42 orders were successfully completed among orders placed in the 90 days from June 5 through September 2, 2026, inclusive. "Successful" means recorded status delivered, using data last synced October 30, 2026-not historical status as of September 2.
- Insight card for the reworded question
Successful orders: 42. Counts orders placed June 5 through September 2, 2026 with recorded status delivered. Data last synced October 30, 2026.
| ✓ | Expected: Both the repeated question and a rephrasing must return the verified query's result: 42 orders.Found: Both post-training chats returned 42, the result of the verified query (orders with status delivered placed 5 June - 2 September 2026), including the reworded question. |
The rephrased answer followed the trusted definition, demonstrating learning beyond the supplied wording. Conversation learning was off and each question used a fresh chat; the change from the earlier interpretation is attributable to the saved example.
A valid earlier interpretation. The before-training chat's answer of 38 was correct for its stated definition: orders delivered and not refunded by the end of September 2, 2026.
Recorded versus historical status. The repeated-question and reworded-question answers both explained that they used recorded order statuses, not each order's status as of September 2, 2026.
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 | Anomaly AI → |
Global scenario definition → · Global capability definition → · Anomaly AI 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.