By the time the term exam tells you a student is failing, the student has known for six weeks.

The engine — honest about the method

Most ed-tech vendors call this "AI." For predictions about minors that is not enough. SchoolDeck calls it what it is: logistic regression with RFM-ST feature engineering, adapted from the peer-reviewed methodology in Bashar et al. 2023 published in Elsevier's Computers and Education: Artificial Intelligence.

Logistic regression is a statistical model — not a neural network, not a generative AI, not a black box. The model produces a probability between 0 and 1; the coefficients on each feature are explicit and inspectable. The school's Academic Council can audit the model coefficients. Any flag's score can be decomposed into the feature contributions that produced it. The Academic Director can explain any flag to a parent in plain Indian-English: "your child's attendance density dropped 21 points in the last fortnight, Maths formative trend has fallen 1.4 points per week, homework submission is at 40% — together these signals put her in the upper end of our concern range."

Why not a neural network? Because for predictions about children, the school must be able to defend each prediction. A parent in the PTM asks why their child is flagged; a DPDP auditor asks how the school justifies storing predictive data about a minor; an Academic Council asks how the model arrived at its accuracy. A logistic-regression model answers all three. A black-box neural net cannot.

RFM-ST signals — what the model actually reads

RFM-ST is the feature set adapted from the academic literature on student-risk prediction. Each letter is a category of signal:

• R — Recency: How recent are the warning signs? An absence two months ago carries less weight than an absence yesterday. Recent declines matter more than historical ones.
• F — Frequency: How often does the signal appear? Three Monday absences in a row is qualitatively different from three random absences across a term.
• M — Density (Monetary-equivalent in the original): Attendance density over rolling windows. The 14-day density is more sensitive to recent change than the term average — a student averaging 92% who drops to 71% over a fortnight gets flagged before the term-end average has moved noticeably.
• S — Severity: How sharp is the signal? A Maths score going from 18/20 to 16/20 is a soft signal; from 18/20 to 9/20 is a sharp one. The model weights sharper signals more.
• T — Trend: Trajectory slope over recent observations. A formative trend slope of −1.4 points per week is more concerning than a flat low average, because the trajectory predicts where the student is headed if unchecked.

The features feeding this engine come from /features/staff-attendance/ (the daily attendance capture mechanism), /features/examinations/ (the formative test scores), homework submission cadence and timeliness, and behavioural notes logged by class teachers during the term.

Confidence intervals — every flag carries its own uncertainty

A score of 78% without context is just a number. A score of 78% with 95% confidence interval 70-86% tells the Academic Director two things: the model thinks this student is moderately likely to need support, and the model is reasonably confident in that estimate.

Compare with a different flag: 65% probability, 95% CI 42-88%. That wide interval means the model is uncertain. It might still be worth investigating, but the school should weigh it differently from a tight-interval flag. Honest reporting of confidence is what separates a careful predictive model from a marketing-grade "AI score."

The school's overall accuracy band is published transparently on the Academic Director's dashboard:

• Week 1-4 of a new school: Wide intervals; model is bootstrapping on the school's data. The Director should triage flags but not act on probability alone.
• End of term 1: Intervals tighten; false-positive rate typically 20-30% — meaning about 1 in 4 flags doesn't pan out. Still useful but worth dismissing some.
• After 2-3 terms of data: False-positive rate typically below 15%. The model is now school-specific and your accuracy band is shown next to every flag.

The Monday morning flag queue — the actual workflow

The weekly refresh runs overnight on Sunday. Monday 8 AM, the Academic Director opens the flag queue and sees:

1. New flags this week: Students who weren't flagged last week but are this week. Usually 5-15 in a 1,000-student school.
2. Still-active flags: Students flagged in earlier weeks who remain above threshold. Useful for tracking whether prior interventions worked.
3. Cleared flags: Previously flagged students whose risk has dropped below threshold — typically because attendance improved or recent test scores recovered. This is the positive-reinforcement view.

For each flag, the Director reviews the feature attribution, decides whether the flag is genuine, and either routes it to the class teacher with an action prompt or dismisses it with a reason note. The dismissal reason is recorded — over time, dismissal patterns improve the model. If many flags are dismissed for the same reason ("this student has approved sick leave"), the next model refresh weights that feature differently.

The Director's review typically takes 15-20 minutes once a week. Not a part-time data-analyst job — a 20-minute habit.

Remedial impact measurement — did the intervention work?

Schools run remedial classes every year. Most cannot say whether those classes worked. The reasons are structural: the next term exam depends on too many variables, and the school has no before/after measurement built around the remedial intervention itself.

The early-warning module measures remedial impact through the natural experiment that flagging creates:

• Student M was flagged in October at 78% probability. A remedial intervention was logged (extra Maths sessions Tuesdays + Thursdays for 4 weeks).
• The model re-scores Student M weekly. By week 4 of remedial, the probability has dropped to 42%; by week 6, 28%.
• The next formative test score is 14/20 — up from 9/20. The flag clears.
• The model now learns: this student's signals responded to that intervention pattern. Other flagged students with similar feature profiles get the same intervention recommended.

This is closed-loop learning — not "AI magic," just systematic measurement. After two academic years, the school has a documented track record of which intervention types work for which student-feature profiles. The remedial coordinator stops guessing and starts choosing.

What parents see — a growth dashboard, not a risk score

The detailed risk probability, the 95% confidence interval, the feature-attribution breakdown — these are staff views. They live with the Academic Director, Coordinators and Class Teachers under role-based access defined in /features/role-based-access/.

What the parent sees in the SchoolDeck app is different:

• A growth dashboard showing their child's performance trend per subject over the term.
• Strengths panel — subjects where the child is performing well or improving.
• Focus Areas panel — subjects where the child would benefit from attention. Constructive language, no probability score visible, no deficit labels.
• Specific, school-authored suggestions where the class teacher has logged guidance ("Mehak benefits from re-doing chapter 4 exercises; we have shared the worksheet in the homework tab").

This separation matters under NCPCR 2021 guidelines on child dignity and DPDP Act 2023's protection of minor data. Parents see what they need to act on. They do not see the raw analytics machinery.

Prediction ≠ Exam ≠ Narration ≠ Report PDF ≠ Operational Dashboard

The SchoolDeck student-performance cluster has five pages. Each owns a distinct layer. Knowing the boundaries helps schools evaluate them correctly and prevents internal page competition.

• This page · /features/student-performance-ai/ — Owns at-risk early-warning prediction. Probabilistic risk score per student per subject, with 95% CI and feature attribution. The forward-looking analytical layer.
• /features/examinations/ — Owns exam scheme + grading engine. CBSE 80+20 (Periodic Test best-of-3 + Notebook + Subject Enrichment), Class 10 Two Board Exams from 2026, 9-point grading. The numbers and the calculation logic.
• /features/report-card-narration/ — Owns teacher remark text. NEP HPC + PARAKH-aligned narration generated from the marks. The qualitative comment layer.
• /solutions/report-cards-academic-records/ — Owns the final report card PDF. CBSE/ICSE/State Board layouts, NEP HPC card design, school logo, parent app delivery. The deliverable artefact.
• /solutions/school-analytics-reporting/ — Owns the principal+trustee live operational dashboard. Fee collection rate, period-wise attendance, syllabus completion, exam result trends across the school. Looks at current operational health; this page looks forward at individual student risk.

Each page targets a distinct query intent and a distinct ownership claim. This page is for the Academic Director who wants to identify struggling students before the term exam confirms it. The four siblings are for adjacent but distinct concerns.

Term-exam reality vs SchoolDeck early-warning

Practical differences for an Academic Director responsible for 1,400 students across 35 sections.