Test a thesis change on past deals
Run a draft rubric on deals you already know, decided ones first, and compare it side by side with the published version before you publish. Nothing changes.
Before you publish a change to your thesis, check what it would have said about deals you already know. If tightening the stage rule would have failed half your portfolio, you want to know before new decks are screened with it. The sandbox runs the draft and the published version on the same deals, with the same facts and deck, so every difference comes from your change.
Run the draft on past deals
- Open Settings → Thesis and rubrics, open the rubric and make your changes in its draft (Edit rubric starts one).
- Click Try on past deals. A side panel opens.
- On New run, choose up to 25 deals. Decided lists deals in Closed, Terms, Passed and Revisit later; All deals lists the rest too. Search by company. Up to ten decided deals with a deck are chosen for you the first time.
- Click Run on 7 deals (the number you chose). Unsaved changes in the draft are saved first.
The run is queued and the Runs tab shows a progress bar, "3 of 7 deals", while it works. You can close the panel; the run carries on.
Read the results
The top line says how many deals change, with a count for each kind of change:
| Change | What it means |
|---|---|
| Now fails a hard rule | Met every hard rule under the published version; the draft finds a mismatch |
| Now meets the hard rules | Had a mismatch under the published version; the draft finds none |
| Hard rules differ | Some hard rules read differently without changing pass or fail overall, for example Met becoming Unknown |
| Fit up or Fit down | The preference fit moved by 5 points or more |
| No change | The draft reads the deal the way the published version does |
| No published version | The rubric has nothing published to compare with yet |
Changes only shows the deals that move; All deals shows every reading. Each row shows the published and the draft version's fit and mismatch count. Open a row to see every hard rule (Met, Mismatch or Unknown, with the value found) and every preference (score out of 10 and its share of the fit) for both versions. Rules that read differently are highlighted.
Hard rules run in code. Preferences are scored by the model from each deck, exactly as in a real screening, and when both versions ask the same preferences one scoring serves both, so only weights change the fit. A mismatch is never a rejection: it tells you what the draft would flag for a person to look at.
What the sandbox doesn't do
- It doesn't change any deal, screening, override, fact, request, stage or founder status, and it sends nothing to a founder.
- It doesn't publish the draft. When you're happy, click Publish as usual. See Rubric versions and re-runs.
- It keeps no quotes or citations from decks. Open the deal's Fit tab for the real screening with its evidence.
Frequently asked questions
Who can run the sandbox?
Fund admins, the same people who can change rubrics. Partners and analysts read the published rubric.
If I edit the draft while a run is going, which version does it use?
The draft as it was saved when you clicked Run. Later edits don't change that run; run it again to test them. A draft can have one run going at a time.
Is there a limit?
Up to 25 deals per run and 20 runs per fund a day, because each deal can take up to two model calls.
Does it cost model usage?
Yes. Preference scoring goes through the same AI route as a screening. Each call is logged as a sandbox call with the model, prompt version, rubric version and tokens, never with deck text.
Something unclear or missing? Tell us