Learning Fatigue & Anti-overlearning Detection

When persistence stops being an obviously good signal

A learner has worked through code exercises for seventy minutes. Early attempts are quick and accurate. Later attempts become slower, the error rate doubles, hints are requested more often, and the learner still presses continue. A simplistic system might call the learner “fatigued” and force a break. Another simplistic system might celebrate persistence and keep serving more of the same practice.

Both shortcuts are unsafe. The platform can observe behavior; it cannot directly observe an internal medical state. This lesson builds Learning Fatigue, Pacing & Recovery Signal Contract v1: a bounded evidence record that combines session duration, attempts, latency/error/hint trends, spacing, recent practice load, later retention evidence, learner preference and a visible review state.

The strongest successful state is LEARNER_CHOICE_REVIEW_REQUIRED. It means the evidence is sufficient to offer a reversible pacing choice or route adjustment, not to diagnose fatigue. Missing or conflicting evidence must remain UNKNOWN, INCONCLUSIVE or HOLD.

Route orientation and ownership boundary

MKN-014 owns pacing/fatigue signals and decision boundaries. MKN-013 is the hard prerequisite: motivation/gamification choices must already be bounded so a pause is not turned into a punishment. MKN-015 owns weighted project progress and must not be absorbed here.

Foundation learners should trace one baseline session from telemetry to a bounded pause/continue choice. Repair learners should locate the first missing or contradictory evidence field. Experienced learners should defend why a model cannot infer a health condition from response latency, why retention evidence can overturn a first impression, and why learner override must remain visible.

Checkpoints: F1 separate observable performance drift from diagnosis; F2 reject one single-signal shortcut; F3 recompute the intervention when the same telemetry appears after a week-long gap with strong retention and an explicit learner preference for a long session.

Mental model: observe → contextualize → classify uncertainty → offer reversible choice → verify later

Use this flow:

session telemetry → spacing + recent load → multi-signal performance trend → retention context → bounded signal class → optional pause/lighter retrieval/continue → learner override → later retention verification

The crucial separation is between transient performance and persistent learning. Slower responses late in a session can coexist with useful learning; quick responses can coexist with weak long-term retention. Session telemetry is therefore evidence about what happened during a session, not a complete description of learning state.

A pause recommendation is an intervention proposal, not a health statement. It must be reversible, learner-visible and non-punitive. If the learner overrides it, the system records the choice and continues monitoring rather than converting disagreement into a penalty.

Foundation contract: sixteen evidence dimensions

The frozen dimensions are learning_session_id, task_family, session_duration, attempt_count, response_latency_trend, error_rate_trend, hint_rate_trend, spacing_interval, recent_practice_load, retention_check_ref, performance_delta, fatigue_signal_class, medical_diagnosis_forbidden, pause_recommendation, learner_override, and review_state.

Minimum invariants:

1. Session identity and task family are explicit.
2. Duration and attempt count are descriptive context, never a diagnosis by themselves.
3. Latency, error and hint trends remain separate signals.
4. Spacing interval and recent practice load are explicit.
5. Retention evidence is linked when available; absence remains visible.
6. performance_delta states what changed relative to a declared comparison point.
7. fatigue_signal_class is a product signal such as POTENTIALLY_UNPRODUCTIVE_PATTERN, not a medical label.
8. medical_diagnosis_forbidden must remain true.
9. Pause/recovery options are reversible and non-punitive.
10. Learner override is recorded without penalty.
11. Conflicting evidence produces UNKNOWN/INCONCLUSIVE or human review, not forced success.
12. Changed spacing or retention evidence requires recomputation rather than copying the baseline label.

Deep explanation: spacing and transient performance are evidence, not a universal break timer

Distributed practice — FACT. Research on distributed practice reports strong retention advantages from spacing study opportunities, but the useful spacing depends on task and retention conditions. The lesson therefore uses spacing as contextual evidence; it does not invent one universal “break every N minutes” rule.

Persistent versus transient processes — FACT. A practice-schedule framework in motor learning distinguishes persistent learning/forgetting processes from transient processes including warm-up decrement and fatigue. This separation is valuable for product design because end-of-session performance can change for reasons that do not map one-to-one to later retention.

Multi-signal decision — INFERENCE. Combining latency/error/hint trends with spacing, recent load and later retention is a safer product inference than labeling a learner from a single metric. A strong later retention check can legitimately cause the platform to recompute a prior pacing recommendation.

Medical boundary and reversible intervention — DESIGN. MAKNOW forbids medical diagnosis from learning telemetry. It also requires pause/recovery recommendations to remain optional, visible, reversible and reviewable. Thresholds, windows and intervention policy are local design authority, not findings supplied by the cited papers.

Worked cases: same late-session drift, different decisions

Baseline. Session S-014-A lasts 70 minutes. Response latency rises 38%, error rate rises from 9% to 21%, and hint use rises from 0.12 to 0.27 per attempt. Spacing from the previous comparable session is only 20 minutes and recent load is high. No delayed retention check is available yet. The system records a POTENTIALLY_UNPRODUCTIVE_PATTERN, offers an optional pause or lighter retrieval, and leaves the state at LEARNER_CHOICE_REVIEW_REQUIRED.

Wrong approach. A rule says “latency +30% = fatigued”. This is rejected because it converts one observable trend into a medical/internal-state claim, ignores error/hint/spacing/retention context, and offers no uncertainty state.

Transfer. After a week-long gap, the same late-session latency/error/hint pattern appears, but a delayed retention check is strong and the learner explicitly prefers a long session. The platform must rebuild the route from the changed evidence. It may recommend continuing with a lighter retrieval block and later verification rather than copying the earlier pause recommendation.

Failure modes that must stay visible

Fail closed when any of these occurs:

• only one telemetry signal is present but a confident label is requested;
• medical_diagnosis_forbidden is false or the label uses diagnostic language;
• spacing or recent-load context is missing while the rule claims overlearning;
• retention evidence contradicts the initial session pattern but is ignored;
• a pause is mandatory, punitive or tied to reward loss;
• learner override is discarded;
• a recommendation is copied to a changed context without recomputation;
• provenance for the comparison window or retention check is missing.

A safe system can say “evidence insufficient” more often than a flashy system. That is a feature, not a defect.

Alternatives and trade-offs: four routes, no magic score

The control policy has at least four defensible responses: continue unchanged, offer an optional pause, switch to lighter retrieval, or route to human/learner review. No single score should hide why one route was selected.

A pause is low-cost and reversible but may interrupt productive deliberate practice. Continuing avoids interruption but can waste time when performance is clearly degrading. Lighter retrieval can preserve engagement while reducing immediate load, but it may alter the practice objective. Human review is slower but appropriate when signals conflict or the intervention has meaningful downstream consequences.

The system should expose which evidence is decisive and which is advisory. Duration alone is advisory. A multi-signal decline plus dense spacing and high recent load can justify a pacing prompt, but later retention evidence can weaken that recommendation.

Experienced bridge: from heuristic nudges to auditable state transitions

For experienced engineers, treat MKN-014 as a state machine rather than a “fatigue model”. A feature pipeline may compute signal summaries, but the product boundary is enforced after feature computation:

• raw telemetry remains immutable evidence;
• derived trends declare their comparison window;
• retention checks point to an explicit later event;
• signal class never uses a medical diagnosis vocabulary;
• intervention state is decoupled from reward or disciplinary state;
• override becomes a first-class event;
• every route change records the rule version and evidence refs used.

This architecture also supports later recalibration. If a threshold changes, historical raw evidence can be re-evaluated without pretending the new policy was always in force.

Guided and independent practice

Guided: open lab://MKN-014/MKN-014-LAB1 and run the baseline. Explain why three late-session trends plus short spacing/high load justify a pacing prompt but still do not justify a medical fatigue label. Then remove retention_check_ref from a transfer case and observe why the state becomes unknown/review instead of reusing the baseline answer.

Repair: run a single-fault fixture. State the exact invariant violated and the smallest evidence repair. Do not compensate for one missing field with an unrelated positive metric.

Independent: use the week-gap transfer fixture. Ignore the baseline decision label. Recompute from changed spacing, strong delayed retention and explicit learner preference. Your artifact must retain learner override and provenance and must not use the baseline result as an oracle.

Verification evidence

Verification requires more than “the code ran”. The reviewer must be able to reproduce the state from visible synthetic evidence.

Required checks include: all sixteen dimensions exist; medical diagnosis is forbidden; at least two distinct behavior trends are available before a strong pacing signal; spacing/recent load are represented; retention contradictions trigger recomputation; learner override is preserved; pause is reversible/non-punitive; provenance is present; transfer marks recomputed_from_changed_evidence=true; and baseline_oracle_reuse=false.

The lab contains deterministic negative cases for missing telemetry, diagnostic labeling, missing spacing/load, ignored retention contradiction, punitive pause, dropped override, missing provenance and transfer oracle reuse.

Assessment and transfer

S1 — Retrieval: explain why response latency alone cannot prove learning fatigue.

S2 — Application: build the signal contract for the 70-minute baseline and justify a bounded optional intervention.

S3 — Diagnosis: shorten spacing while total practice stays constant. Identify which context field changed and why the route must be reconsidered.

S4 — Transfer: the same telemetry appears after a week-long gap, delayed retention is strong, and the learner prefers a long session. Recompute without copying the original pause recommendation.

Passing requires visible uncertainty, a non-diagnostic vocabulary, learner choice and an independently recomputed transfer artifact.

References, freshness and authority boundaries

PUBMED-DISTRIBUTED (PMID 20580350) is a foundational review of mechanisms underlying distributed-practice effects. It supports the claim that spacing can materially affect retention, while not providing one universal break interval.

PUBMED-PRACTICE (PMID 29684760) separates persistent learning/forgetting processes from transient warm-up/fatigue processes in a motor-learning timescale model. It supports process separation, not a diagnostic classifier for every digital-learning task.

Both references were rechecked on 2026-09-13 and remain available as the sources frozen by the Bible. MAKNOW’s signal classes, pause vocabulary, thresholds, learner-override policy and medical-boundary rules are local DESIGN. Any future change to those rules requires a governed version change rather than silent reinterpretation.