Why the fields in an Active Learning project stay fixed once the project starts

Active Learning in Relativity: Why you can’t edit the fields once the project starts

If you’ve spent any time with Relativity’s Active Learning setups, you’ve learned there’s a rhythm to how these projects unfold. You pick a few key fields—the building blocks that describe the data, the way documents are coded, and the rules the system uses to decide what to learn next. Then you launch and let the algorithm do its thing. The big question that often pops up is simple, almost philosophical: can those fields be edited while the project is underway?

In short: no. Not while the project is live. The fields used to create the project are set for the entire run. Let me explain why that matters, what it means in practice, and how to navigate the inevitable situations where you’d wish for a tweak.

What “the fields” really mean in an Active Learning project

When you start a Relativity Active Learning project, you’re not just labeling a pile of documents. You’re defining a schema—a set of fields that describe every document’s role in the learning loop. Think of fields as the grammar of your dataset: document type, metadata tags, coding categories, reviewer assignments, and the strategies for how the system should weigh certain features. These aren’t cosmetic choices. They guide how the model sees the data, how it chooses which documents to review next, and how its predictions are evaluated.

Because these fields shape the learning process, changing them mid-flight would be like changing the rules of a game halfway through a match. The model would suddenly be looking at a different landscape, and the comparisons you’ve been counting on—precision, recall, learning rate, and the trajectory of improvement—could go off the rails. That’s why the field definitions are locked in once you’ve started.

Why this rigidity is a good thing

• Consistency for the learning engine: Algorithms need stable input. If you keep shifting the field definitions, you’re giving the model inconsistent signals. Over time, the model’s understanding of “what matters” could drift, and that undermines trust in the results.

• Reproducibility and auditing: In eDiscovery and data science workflows, you often need to show exactly how a decision was reached. A fixed field set makes it possible to recreate the project, verify steps, and explain outcomes with clarity.

• Reduced risk of human error: It’s easy to tweak something in a moment and forget the ripple effect. Locking the fields reduces the chance of unintended consequences that ripple through the training rounds.

• Clear governance and data integrity: The learning loop thrives on a well-defined, stable structure. This stability helps teams manage quality, track progress, and communicate findings to clients or stakeholders.

What to do if you anticipate a need for change

This may sound strict, but there are practical ways to handle it without breaking the project’s core logic:

• Plan the schema carefully before launch: Gather input from reviewers, data stewards, and subject-matter experts. Map out the most important fields and how they interact with the learning process. A well-thought-out schema pays dividends later.

• Model the change as a new project: If you need a different field set, start a fresh Active Learning run with the revised schema and a new dataset or a clearly defined iteration of the data. You can compare outcomes between the two runs side by side, which often reveals what really moves performance.

• Use versioning and documentation: Keep a changelog of field definitions, their purposes, and any decisions around them. A concise trail helps when teams revisit results or when new team members join.

• Employ exploratory phases before full-scale learning: If you’re unsure about a field, run a small pilot to see how it behaves. It’s a safer way to gauge impact without risking the entire project’s integrity.

A quick mental model you can keep in mind

Imagine you’re building a recipe for a complex dish. The fields are the ingredients and timing rules. Once your oven timer starts counting down and the dish begins to cook, you don’t swap in a different ingredient list halfway through. If you did, you’d change the flavor, the texture, and the overall outcome. The goal here isn’t to be rigid for the sake of rigidity; it’s to ensure the cooking process stays predictable and the final dish tastes the way you planned. In a data-driven project, predictability translates to trust in results.

Practical consequences you’ll notice in daily work

• Fewer mid-project surprises: When teams know the fields are fixed, there’s less risk of last-minute, unexpected changes that ripple through analysis, reviews, and reporting.

• Cleaner audit trails: Stakeholders can follow how the model learned and why it made certain recommendations, because the data schema remained consistent throughout.

• Better governance for data quality: With a stable schema, you can standardize how documents are tagged, what metadata is captured, and how reviewers apply codes. That standardization pays off in accuracy and reliability.

A few real-world echoes and caveats

• Tooling often mirrors the rule: Relativity and similar platforms are designed with this principle in mind. You’ll see clear pathways to start a new project or to clone the current one with a revised schema if you truly need to test a different field setup.

• Edge cases happen, of course: Sometimes teams discover in the early rounds that a field would have been useful. In those moments, the practical route is to record the insight and carry it into the next iteration rather than trying to retrofit the current run.

• Communication matters: If you’re working with a client or a cross-functional team, set expectations early. Explain that field changes can’t occur mid-run, but you’ll capture learnings and apply them to future cycles.

Relativity-specific angles and practical tips

• Start strong with a solid schema: In this space, a well-structured set of fields helps the Active Learning engine pick the most informative documents for human review. It also makes it easier to interpret which kinds of documents tend to drive learning progress.

• Think in layers: Separate core document metadata from the coding schema. The core fields anchor the dataset, while the coding fields capture the reviewer’s labels. This separation helps keep the project organized and makes transitions smoother when you begin a new run with a revised schema.

• Document decisions, not just data: A short write-up about why certain fields exist and how they’re used can save a lot of time later. It helps new team members align quickly with the project’s data discipline.

• Leverage saved templates: If your organization routinely runs similar Active Learning projects, template schemas can save time and reduce the temptation to fudge the field set mid-way. It’s not cheating; it’s disciplined efficiency.

A few conversational checkpoints

• Have we locked the field definitions? If yes, you’re aligned with the core principle. If not, pause and confirm whether a new project or a revised template is the right path.

• Do we understand the impact of each field on the learning loop? If you can’t articulate how a field affects model behavior, take a moment to map that out before you launch.

• Is our governance record in good shape? A concise log of decisions around the schema helps everyone stay on the same page.

Bringing it all together

The rule about not editing fields once an Active Learning project has begun isn’t a whimsical constraint. It’s a practical safeguard that anchors data integrity, learning stability, and trust in outcomes. When you approach project setup with that in mind, you’re more likely to see smoother runs, clearer results, and fewer surprises. And if you ever do feel the urge to rewrite the rules, treat it as a fresh start—anchor a new project with a revised schema, compare the outcomes, and let the data tell you what really works.

If you’re navigating Relativity’s landscape, this mindset—planning before execution, valuing consistency, and embracing structured iteration—will serve you well. It’s a blend of technical discipline and thoughtful judgment, the kind of balance that makes data-driven work both effective and genuinely engaging. After all, the meat of the work isn’t just in choosing fields; it’s in how those choices steer the journey from raw documents to reliable insights. And that journey, when guided by a steady schema, feels less like a gamble and more like a well-mounded path forward.