# Learning before answering | SwaVid research note | SwaVid A detailed SwaVid research note on why an AI tutor should map the learner and the missing prerequisite before it starts answering: the map → teach → check → adapt loop, the design principles, and the honest evidence boundary. Canonical: https://swavid.com/research/learning-before-answering Source: https://swavid.com/research/learning-before-answering # Learning before answering. ## The note at a glance ## Three ways to respond to a learner: only one is tutoring. ## The loop is not a guess. Each step has a research basis. ## Map, teach, check, adapt: what each step owes the literature. ## A polished answer can hide the real problem. ## Same question, two very different systems. ## “Why is friction less on ice?”, the long way round. ## The tutor earns the next question. ## This note argues the loop. The companion paper measures the machine. ## Good product thinking is honest about what it knows. ## Frequently asked, answered plainly. ## The works this note relies on. ## The map decides what to teach. The answer comes after. ### The chat tutor ### The learning system ### The useful tutor ### The real difference ### Tutoring set the benchmark ### Prior knowledge decides what helps ### Mastery can be tracked, not guessed ### The right help depends on the learner ### Checking beats re-reading ### Timing is a learning decision ### Examples teach only when processed ### Unguided AI help can weaken learning ### Adaptive tutoring has produced real gains, under specific conditions. ### Locate the idea underneath ### Repair with fading support ### Test with help removed ### Change what comes next ### The question arrives ### Map before answering ### Probe the foundation ### Repair, then return ### Check that it stayed ### Start with a map ### Treat answers as evidence ### Repair before accelerating ### Keep checking ### A published prerequisite graph ### A Bayesian gap posterior ### A minimum-cost question policy ### The pedagogy it serves ### Every signal has a boundary. An answer engine responds to the question in front of the learner. A tutor responds to the learner in front of the question. This note lays out the loop, the science, and the evidence boundary behind that difference. A product thesis with an honest evidence boundary, not an efficacy study. An answer engine stops at the answer. The learning system maps, teaches, checks, and adapts, choosing the repair before the reply. One design position, one loop, and one classroom signal. Each is labelled with exactly what it can and cannot support. Map → teach → check → adapt. The loop stays visible to the learner and the parent. Learning profiles were reviewed in a real classroom before this direction was scaled. Teachers recognised their students in the generated profiles. Perceived fit, not efficacy. This note argues a design position. It does not report measured learning gains. The same question can trigger very different behaviours. What separates a tutor from an answer engine is not the quality of the sentence it produces; it is whether the system knows what the learner needs next. Answers the question in front of the learner, fluently, instantly, and often correctly. Finds the idea underneath the question: the prerequisite the learner may never have secured. Teaches that idea, checks the understanding again, and changes what comes next. An answer is not the same thing as learning. The difference is whether the system knows what the learner needs next. This is SwaVid’s product thesis, not a claim that one interface is universally correct. Any tutor that aims at durable learning needs an equivalent loop grounded in evidence. Eight findings from the learning-science literature, stated with their original scope. They justify building this loop; they do not measure SwaVid’s implementation of it. Bracketed numbers jump to the reference list. Bloom reported that one-to-one tutored students performed about two standard deviations above a conventional class. The result is a challenge to mechanise responsiveness, not an effect any software inherits by adopting the label “tutor”. Ausubel’s famous claim: the most important single factor influencing learning is what the learner already knows. Ascertain that, and teach accordingly. Corbett and Anderson’s Bayesian Knowledge Tracing updates a hidden knowledge state from observable performance while explicitly allowing for lucky guesses and unlucky slips. Cognitive load theory separates a task’s inherent complexity from the load created by how it is taught. The expertise reversal effect shows that structure which rescues a novice can distract an expert. Roediger and Karpicke showed that repeated testing can produce better delayed retention than repeated study, even when restudying looks stronger on an immediate test. Fluency in the moment is not durable knowledge. Cepeda and colleagues reviewed 839 assessments across 317 experiments: the benefit of distributed practice depends jointly on the spacing interval and how long the knowledge must remain accessible. Learners differ in the quality of the self-explanations they generate over worked examples, and those explanations predict what they learn. Example variability helps once the underlying structure is stable. In a study of nearly one thousand high-school mathematics students, access to a general GPT interface improved performance during practice, but those students performed worse than controls once the help was removed. A constrained tutor largely mitigated the harm. 619 students, about 4.5 months of level-matched technology-aided instruction; 0.23 σ in Hindi. The effect belongs to that evaluated programme, not the category. [6] Working paper on a sample more than twenty times larger, after 18 months; 0.20 σ in Hindi. Effects shrank at scale: implementation is part of the treatment. [7] Kulik and Fletcher’s review of 50 controlled evaluations against conventional instruction. The distribution was not uniform. [9] Ma and colleagues, 14,321 participants: intelligent tutors beat large-group and non-intelligent computer instruction, but showed no significant advantage over human one-to-one tutoring. [8] VanLehn’s review adds the mechanism that matters here: interaction granularity. Systems that engage with intermediate reasoning steps approach the responsiveness of human tutoring better than systems that wait for a final answer. [5] Before any explanation is generated, the system locates the curriculum target, the prerequisites it depends on, the learner’s recent evidence, and the uncertainty attached to each estimate. The missing idea is addressed with a bounded, curriculum-grounded activity: a worked example, then a completion problem, then independent practice as the evidence strengthens. Independent evidence is collected with assistance reduced or removed, the only condition that separates durable understanding from fluent imitation of the help. The system repairs the gap, schedules a retrieval check at an effortful-but-possible interval, or advances the path, and updates the learner map either way. Learners ask downstream questions. The shaky idea is usually upstream. Answering the asked question, however well, can produce the feeling of progress while the foundation stays broken. So the system first locates the idea underneath the question, repairs it if the evidence says it is missing, and only then returns to what the learner originally asked. More content is not more learning. When an older idea is shaky, additional explanation can increase the feeling of progress without increasing understanding. The system repairs first, then accelerates, and it keeps checking after the help disappears. The behavioural contract differs at every step, not just in the quality of the final sentence. The loop takes five moves to reach the same sentence an answer engine reaches in one. The difference is what the learner can still do next week. “Why is friction less on ice?” A fluent explanation of ice and molecular smoothness is available instantly. That availability is precisely the risk. The system locates what the question depends on: normal force, contact surfaces, and how force changes motion. The asked question is downstream of all three. Two trusted probes test normal force from different representations. Both answers suggest the idea is shaky, so the minimum two-probe rule is satisfied before any gap is declared. A short bounded activity rebuilds normal force: worked example, completion problem, independent item. Only then does the system return to the ice question the learner actually asked. The learner answers the ice question unaided, and a delayed retrieval check later confirms the repair survived the weekend, not just the session. Four principles turn the thesis into product behaviour. Each one constrains what the system is allowed to do next. Curriculum context plus a view of the ideas the current lesson depends on, before any explanation is generated. A response is most useful when it locates the next support decision, not when it merely produces a correct sentence. When an older idea is shaky, the next step fixes the foundation instead of piling new content on top of it. Learning is a loop: map → teach → check → adapt. The loop stays visible to the learner and the parent. The tutor does not improvise a teaching style per session. Strategy is selected internally from curriculum, learner evidence, session goals, and safety rules. Every step of map → teach → check → adapt runs on measured infrastructure: a reviewed prerequisite graph, a Bayesian diagnosis, and a constrained next-question policy with held-out results and a published failure record. 520 reviewed prerequisites (187 question-bearing and 333 chapter-navigation) bound what may be taught or probed. Edges ship through an audit and a publication gate, never from the model alone. Each answer updates the probability that a prerequisite is missing. At least two probes are required, and a gap resolves only at the 95% threshold. 930 / 930 held-out oracle agreement with zero infeasible selections: a consistency result with hard masks, not a student-outcome claim. Map → teach → check → adapt is what the measured engine exists to power. The whitepaper measures the engine; this note explains the drive cycle. Same honesty standard: the companion whitepaper reports oracle agreement, calibration, and its six false positives, not student outcomes. Read how SwaVid chooses the next question . It does mean: teachers confirmed the generated learning profiles as accurate readings of their students: strong face-validity for the map-first direction. It does not mean: measured learning gains, clinical accuracy, or a controlled comparison against answer-first tutoring. Source: SwaVid classroom validation, reviewed July 2026. Aggregates only; no student-level records are published. No. It argues a design position from published learning science, external programme evaluations, and one classroom face-validity signal. It reports no measured learning gains for SwaVid, and nothing here should be read as one. Because a fluent answer to a downstream question can hide an upstream gap. Research on generative AI assistance shows practice performance can improve while unaided performance gets worse. The feeling of progress is not the same thing as progress. Yes, when the map says the foundation holds, answering directly is the right move. The loop decides the order of operations; it does not prohibit answers. In a classroom review with 150+ students, teachers confirmed the generated learning profiles at 85% agreement, a perceived fit with strong face validity rather than clinical accuracy. The measured engine behind the map is documented, with held-out results and failure analysis, in the companion whitepaper. No. Every effect size on this page belongs to the external programme or literature that produced it, with its own population and conditions. They justify building this loop; they do not measure our implementation of it. Repair targets the smallest prerequisite that blocks the current goal, then returns to that goal. Mastery learning holds the destination stable and varies the path, so the detour is short, local, and checked. Numbering matches the companion whitepaper, Diagnosis before generation, which reviews these sources in full context. SwaVid’s tutor maps the learner, finds the missing prerequisite, then answers with purpose, and keeps checking after the help disappears. - Bloom, B. S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6), 4–16. Source ↗ - Bloom, B. S. (1968). Learning for Mastery. Evaluation Comment, 1(2). Source ↗ - Corbett, A. T., & Anderson, J. R. (1995). Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge. User Modeling and User-Adapted Interaction, 4, 253–278. Source ↗ - Ausubel, D. P. (1968). Educational Psychology: A Cognitive View. Source ↗ - VanLehn, K. (2011). The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems. Educational Psychologist, 46(4), 197–221. Source ↗ - Muralidharan, K., Singh, A., & Ganimian, A. J. (2019). Disrupting Education? Experimental Evidence on Technology-Aided Instruction in India. American Economic Review, 109(4), 1426–1460. Source ↗ - Muralidharan, K., & Singh, A. (2025). Improving Schooling Productivity through Computer-Aided Personalization: Experimental Evidence from Rajasthan. NBER Working Paper 34205. Source ↗ - Ma, W., Adesope, O. O., Nesbit, J. C., & Liu, Q. (2014). Intelligent Tutoring Systems and Learning Outcomes: A Meta-Analysis. Journal of Educational Psychology, 106(4), 901–918. Source ↗ - Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review. Review of Educational Research, 86(1), 42–78. Source ↗ - Bastani, H. et al. (2025). Generative AI Can Harm Learning. Proceedings of the National Academy of Sciences, 122(17). Source ↗ - Sweller, J. (1988). Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science, 12(2), 257–285. Source ↗ - Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). The Expertise Reversal Effect. Educational Psychologist, 38(1), 23–31. Source ↗ - Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255. Source ↗ - Cepeda, N. J. et al. (2006). Distributed Practice in Verbal Recall Tasks: A Review and Quantitative Synthesis. Psychological Bulletin, 132(3), 354–380. Source ↗ - Chi, M. T. H. et al. (1989). Self-Explanations: How Students Study and Use Examples in Learning to Solve Problems. Cognitive Science, 13(2), 145–182. Source ↗ - Kulik, C. C., Kulik, J. A., & Bangert-Drowns, R. L. (1990). Effectiveness of Mastery Learning Programs: A Meta-Analysis. Review of Educational Research, 60(2), 265–299. Source ↗ - Renkl, A., Stark, R., Gruber, H., & Mandl, H. (1998). Learning from Worked-Out Examples: The Effects of Example Variability and Elicited Self-Explanations. Contemporary Educational Psychology, 23(1), 90–108. Source ↗ ## Key Links - [See the learning map](https://swavid.com/learning-debt-identifier) - [Read how SwaVid chooses the next question](https://swavid.com/research/how-swavid-chooses-the-next-question) - [Source ↗](https://doi.org/10.3102/0013189X013006004) - [Source ↗](https://eric.ed.gov/?id=ED053419) - [Source ↗](https://doi.org/10.1007/BF01099821) - [Source ↗](https://openlibrary.org/books/OL26705958M/Educational_Psychology_A_Cognitive_View) - [Source ↗](https://doi.org/10.1080/00461520.2011.611369) - [Source ↗](https://doi.org/10.1257/aer.20171112) - [Source ↗](https://www.nber.org/papers/w34205) - [Source ↗](https://doi.org/10.1037/a0037123) - [Source ↗](https://doi.org/10.3102/0034654315581420) - [Source ↗](https://doi.org/10.1073/pnas.2422633122) - [Source ↗](https://doi.org/10.1207/s15516709cog1202_4) - [Source ↗](https://doi.org/10.1207/S15326985EP3801_4) - [Source ↗](https://doi.org/10.1111/j.1467-9280.2006.01693.x) - [Source ↗](https://doi.org/10.1037/0033-2909.132.3.354) - [Source ↗](https://doi.org/10.1207/s15516709cog1302_1) - [Source ↗](https://doi.org/10.3102/00346543060002265) - [Source ↗](https://doi.org/10.1006/ceps.1997.0959) - [See the learning map](https://swavid.com/learning-debt-identifier) - [Read the classroom case study](https://swavid.com/case-studies/apple-global-school) - [How the next question is chosen](https://swavid.com/research/how-swavid-chooses-the-next-question)