- Tests, 0 failing, 0 skipped
- 1,182
- Tests needing no Docker
- 675
- Lines of product code
- ~31k
- Certification checks passed
- 23
What works
- Connecting, with the read-only property proven by the database rather than asserted by the app.
- Scanning a real schema and keeping the history — 94 tables on the Pagila sample, 368 subjects on a MusicBrainz copy of 344.3M rows.
- An entity map from declared foreign keys, plus implicit links that are measured and marked as measured, never guessed silently.
- A fixed five-question interview that turns what you know into a project profile, and a scan plan that changes with it.
- Asking a question and getting a timeline back, with unwalkable routes named rather than dropped.
- A packaged Windows build that passes the Windows App Certification Kit.
What is not done
- Not in the Microsoft Store yet.
- The package passes certification, but the Store identity and the certification notes are unwritten. Until then this is a source checkout, not an installer.
- PostgreSQL only. No MySQL, no SQL Server, no MongoDB.
- macOS and Linux are untested. The code has no Windows-only logic that we know of, and "that we know of" is the honest size of that claim.
- The question layer has been measured on two databases. It has not been run against a schema nobody here has seen.
- There is no hosted option and no proxy. You bring your own model key, which suits a technical user and does not yet suit the person this product is aimed at.
Measurements that came out badly
These are on the site because leaving them off would make the rest of it worth less.
| What was measured | Result |
|---|---|
| Prompt injection against the live question path | 8 of 32 attempts succeeded. Aiming outside the menu is closed; aiming at the wrong table inside it is not. |
| Clean end-to-end question runs | Only 3, against one database and one model. Enough to show the path works, not enough to state an accuracy. |
| The control case for the closed defence | Never actually triggered on the live path — the model ignored the planted decoy every time. The evidence for that defence is unit tests, not a live run. |
| First attempt at a large schema menu | A 368-table schema produced a 2 MB menu that did not fit in a context window at all. Fixed by splitting into two rounds; recorded because the first design simply did not work. |
Where the numbers come from
Every measurement on this site is recorded in the private repository’s field log, including the ones above, together with the limits on what each one can be used to conclude. Where a number appears here it came from a tool, not from a commit message. The code itself is public: github.com/ledar1106/ledar-src.