Improving a model usually means repeating the same loop by hand: change an attribute, train a new draft, check the score, and try again. The Draft Model Optimizer runs that loop for you. It takes a draft model you've already trained and keeps trying improvements in the background, one change at a time. When it finds a better version, it gives it to you as a new draft. Your original draft is never changed.

The Optimizer is in beta. It's ready to use, and we're still improving it. That's why you'll see a Beta label when you start a run and on the Optimization summary.

The Optimizer works only on your attribute queries. In each round, which we call an iteration, it makes one focused change, trains a new draft, and keeps the change only if the result is better.

The changes it makes are of three kinds:

It works through your data step by step and builds on what's already working in your model. Every feature it creates uses only data from before the prediction date, and never uses your target.

If you have a lot of data, a run may not get to every table. If you know a table matters, name it in your instructions when you start the run (see Guiding the Optimizer below).

What it never changes: your predictive question, entity, target, horizon, core set logic and data split settings stay exactly as you set them. It also doesn't remove columns from your existing queries, except to fix leakage. The one exception to the core set rule is related to label drift, explained under Health checks during optimization below. If the Optimizer drops older rows because of drift, your split settings are applied to the data that remains.

You get the most out of the Optimizer when you run it on a thin draft: one where your core set is ready and you've added your first attribute or two. A thin draft leaves the Optimizer the most room to find new signal in your data.

You can also run it on a more developed draft. Just expect fewer and smaller improvements, because more of the useful signal may already be in your model.

Using the Optimizer is optional. You can keep building and training drafts yourself, and combine both approaches.

The optimization runs in the background, so you can keep working or leave the page. It usually takes about 1 hour, sometimes 2 hours or more, depending on your data.

While it runs, you can keep training other drafts. You can also optimize several drafts at the same time, one run per draft. To stop a run, click Cancel, then confirm with Cancel optimization. A canceled run doesn't create a new version, and you can optimize the same draft again whenever you like.

The Optimizer learns a lot from your data, but some things aren't visible in it. When you start a run, click Guide the optimizer (optional) and write instructions to share what only you know. The Optimizer treats your instructions as facts and follows them throughout the run.

Good things to share:

What instructions can't change:

You don't need to ask it to improve accuracy. It always does.

Leakage still comes first. When a leakage flag is unclear, for example on a column whose values can change over time, your explanation helps the Optimizer decide whether to keep it. If it finds evidence in the data that a column uses information from after the prediction date, it won't use the column, even if your instructions ask it to.

What it already knows. The Optimizer also uses what Pecan has learned about your data in your earlier conversations with the assistant. If your instructions say something different, it follows your instructions.

To see the instructions a run used, open its Optimization summary and go to the User instructions tab.

A new section appears below your draft, showing the optimization as it happens:

Each new version is scored on your test set, a slice of your data held back from training. That way the model is judged on data it hasn't seen. One score decides whether a version is better:

A version counts as improved only if it beats the best version found so far on this score. Other metrics you see, like precision and recall, often improve too, but not always. Because the Optimizer is working toward this one score, another metric can sometimes go down.

Sometimes the optimized model scores lower than your original draft and is still the right result. This happens when the original draft had leakage: it used information that wouldn't be available at the moment of prediction, for example a column that's only filled in after the outcome is known. Leakage makes a model look better in testing than it will perform in real use.

When the Optimizer confirms leakage in your original draft, it treats that draft's score as inflated and doesn't use it as the bar to beat. The optimized version's lower score is the honest one. It reflects how the model is likely to perform once it's making real predictions.

Every new version goes through the same health checks as any draft model, and the Optimizer handles problems on its own:

A version with an unresolved leakage or drift issue is never picked as the best result.

When the run ends, you'll see one of three outcomes:

You'll also hear about the result in chat: the next time you message the assistant in this flow, it tells you how the optimization went, and you can ask it questions about the result. If the Optimizer made leakage decisions, the message also lists the columns it kept, removed or filtered, and asks you to confirm them. If you disagree with a decision, tell the assistant and it can reverse it.

Click What's improved? on the optimized version to open the Optimization summary. After you've started using the optimized version, you can still open it any time from the version's three-dot menu, under Optimization Report.

At the top of the summary:

Below that are the tabs. Some appear only when they apply.

Report lists the changes that made it into the final version, in order. Each row shows the type of change (such as Attribute added, Attribute edited, Leakage treated or Fixing drift), what happened and why, and the score after that change. Expand a row to see a fuller explanation and the SQL changes. When there are columns to review, a line at the top of the Report links to the Needs review tab.

The Report shows only the path from your original draft to the final version. To see every iteration the Optimizer tried, including the ones that didn't help, open the Improvements summary tab.

Evaluation shows the optimized version's performance summary, which features drive it most, and how your core set changes over time.

Needs review appears when the Optimizer kept columns that were flagged as possible leakage. Sometimes the Optimizer flags a column, investigates it, and decides it's safe to keep. It can't be certain about this, because only you know exactly when your data becomes available. So every column it kept this way, and that's still in the final version, is listed here with the reason it was flagged. This includes columns from your original draft as well as ones the Optimizer added.

Expand a row to see the checks the Optimizer ran and its conclusion, which explains why the column was dismissed.

Before you rely on the optimized version, go through this list. For each column, ask: would this value be known at the moment the prediction is made? If yes, no action is needed. If you're not sure, or the answer is no, click Use this version, then remove the column in the notebook, either yourself or by asking the agent, and train a new draft. A model that relies on future information looks strong in testing but underperforms once it's making real predictions.

Notebook shows every query in the optimized version. Queries the Optimizer added or edited are marked, with the changes highlighted, so you can review the exact SQL before you use it.

User instructions appears when you added instructions to the run, and shows exactly what you wrote.

Improvements summary lists every iteration the Optimizer tried and its result, the same list you see on the draft card.

The optimized version isn't used until you choose it. When you're happy with what changed, and you've gone through anything under Needs review, click Use this version, either on the version's card or in the Optimization summary. It becomes your current version, and your original draft stays in the list, so you can always go back.

To go back, open the three-dot menu on your original draft and click Restore. Choose Restore & Replace Current to return your notebook to that draft, or Restore & Create New to open it in a new notebook.