Section Analytics: now in private beta.

Most candidates know their score. They don’t know where their score is bleeding.

A 645 on a full-length mock is a number. It tells you almost nothing about whether you lost twenty points to a single weak DI sub-skill, twelve points to pacing drift in the back half of Verbal, or to a slow accumulation of small errors across the entire Quant section. Without that decomposition, the next study session is guesswork. The instinct is to drill what already feels uncomfortable — which is rarely the same thing as what’s actually costing points.

Section Analytics is the surface we built to close that gap. It’s now in private beta.

The four signals

We resisted the temptation to ship a dashboard with thirty widgets. The beta surfaces four signals, chosen because each one drives an actual change in what you study next.

1. Per-section percentile

Raw scores are noisy across calibration revisions. Percentiles, normalized against the current test population, are more honest. Each section reports its own percentile alongside its scaled score, refreshed after every session. If your Quant percentile drops three points week-over-week while your scaled score sits flat, that’s a real signal — the population is moving faster than you are.

2. Time-per-question drift

We track median time-per-question per section over the last N sessions, broken out by difficulty band. The interesting view isn’t the average — it’s the drift. A candidate whose hard-item time has crept up from 2:10 to 2:35 over four weeks is in a different place than one whose pacing is stable, even if their score looks identical. The first candidate is leaking accuracy on the back half of the section. The second is not.

3. Topic-level mastery deltas

The engine maintains a mastery estimate per topic — combinatorics, weighted averages, MSR schema-building, CR inference, and so on. Section Analytics surfaces the deltas, not the absolute estimates. Where did mastery move this week? Where has it stalled for three weeks running? Where did it regress? Stalled topics are the most actionable line on the screen.

4. Predicted band with a confidence interval

The headline view is a predicted score band — not a single point estimate — refreshed every session. Below it, a confidence interval that widens or tightens as the engine accumulates evidence.

How predicted band is computed

We’re going to be specific about this, because vague claims about “AI-powered score prediction” are mostly noise. The method is standard IRT-based simulation. Conceptually:

for trial in 1..N:
  simulated_section = sample_items(mastery_vector, difficulty_distribution)
  raw_score = sum(response_probabilities(simulated_section))
  scaled_score = irt_scale(raw_score, section_calibration)
  trials.append(scaled_score)

predicted_band = percentile_interval(trials, lower=0.20, upper=0.80)

In plain English: the engine takes your current mastery vector, runs many simulated full-length sections against the live difficulty distribution, and reports the middle of the resulting distribution as a band with confidence bounds. The band tightens as more session data flows in — early-stage candidates see wider intervals, candidates approaching test day see narrower ones.

This is genuinely how serious IRT-based test simulators have worked for decades. We’re not claiming a new method. We’re claiming a clean implementation of a method that most prep platforms either skip entirely or hide behind a single-number score estimate that no statistician would endorse.

Why private beta

We could ship this to everyone today. We’re not going to.

Private beta is the right shape for this release because we don’t yet know which of the four views matter most to candidates in practice. Internal usage is not the same thing as real candidate usage. In our internal testing, time-per-question drift was the second-most-glanced view; we suspect it will be first in the wild, but we don’t actually know yet.

We’d rather ship one screen everyone uses than four screens nobody opens. The beta cohort tells us which is which.

Beta candidates get the full surface, a feedback channel that goes directly to engineering, and the ability to flag views that are confusing or wrong. We expect to cut at least one of the four widgets before general release. We don’t yet know which one.

How to opt in

Section Analytics is enabled for any Pro or Ultra subscriber who opts in from the dashboard settings. New candidates can opt in during onboarding at /register. We’ll roll the cohort out in batches over the coming weeks — early sign-ups go first, but everyone who opts in during beta will get access before general release.

What’s not in this release

Three things were on the original spec and didn’t make the cut.

Item-level review with engine commentary.We had a version that surfaced, for every missed item, the engine’s read on why you missed it — was it the underlying concept, the pacing, the reading load? It worked, but the commentary was uneven. Some items got crisp diagnoses; others got generic ones. Shipping uneven feedback is worse than shipping no feedback. It’s coming back when the diagnosis quality is uniform.

Peer benchmarking.We can show you how your pacing and mastery compare to candidates targeting the same score band. The privacy and signal-to-noise work isn’t finished. We’re holding it.

Predicted band confidence over time.A second-order view that shows how your predicted band has shifted across the last several weeks. Useful, but for the specific question “am I getting better,” the topic-level mastery deltas already answer it more directly. We’re going to wait and see if candidates actually ask for the second-order view.

That’s the beta. Four signals, computed honestly, shipped to the candidates who can tell us which ones to keep.