Reverse-engineer a model from an existing database
The outcome: reverse-engineers an as-is model from a database, separating what is declared from what is inferred.
Builds on: Design the physical data model.
Which AI model is best at reverse-engineer a model from an existing database?
| Model | Provider | Tier | Test score | Knobits mastered (expected) | Cost per test run |
|---|---|---|---|---|---|
| gemini-4-argon | Flagship | – | 93% | $0.021 | |
| gemini-3.5-flash-lite | Low cost | – | 53% | $0.003 | |
| gpt-6-luna | OpenAI | Low cost | – | 63% | $0.002 |
| claude-haiku-5-5 | Anthropic | Low cost | – | 74% | $0.006 |
Test score: mean rubric score on the test tasks for this activity. Knobits mastered: expected share of the activity's knobits the model masters. Cost: mean API cost of one test run at list prices, half price for answers sent as a batch. Method and caveats.
The 7 knobits of this activity
1 that AI does well, 1 it does partly, and 5 where people still lead.
| ID | Knobit | Type | AI exposure (models mastering) |
|---|---|---|---|
| K-1101 | Recognises an inferred structure presented as declared | Skill | People still lead 1/4 |
| K-1102 | Recognises reverse-engineered objects not marked as-is | Skill | AI does this partly 2/4 |
| K-1103 | Recognises a dead database object carried into the model | Skill | People still lead 1/4 |
| K-1104 | Recognises a guessed meaning written as a definition | Skill | People still lead 1/4 |
| K-1105 | Recognises a reverse-engineering source with no environment or date | Skill | People still lead 1/4 |
| K-1106 | Recognises a physical workaround lifted into the logical model | Skill | People still lead 0/4 |
| K-1131 | Reverse-engineers a logical model from a database schema | Skill | AI does this well 4/4 |