Curriculum-Gated Adaptive Learning Platforms

Mechanism

The mechanism represents a curriculum as a directed graph of skill nodes connected by prerequisite edges. Each skill node carries a definition that specifies the observable behaviours constituting acquisition, the assessment modalities admissible for evidence, and the criterion threshold a learner must reach. Learners are represented as state objects whose fields include a per-skill acquisition score, a confidence interval on that score, an engagement profile, and a recent-trajectory summary.

When a learner enters the system, the orchestrator examines the skill graph and the learner state to identify the frontier — the set of skills whose prerequisites are satisfied at acquisition criterion but which the learner has not yet acquired. Content selection proposes activities targeting frontier skills, drawing from a content repository whose items are tagged with the skills they exercise, the modalities they engage, and the difficulty band they occupy. The orchestrator selects the activity whose expected information gain about the learner's frontier is highest given the learner's current confidence intervals and engagement profile.

As the learner engages with the activity, the assessor observes responses and produces evidence updates. Evidence is multimodal: closed-form responses contribute direct correctness signal, open-form responses are evaluated by a rubric-driven model that scores against the skill definition, applied tasks contribute behavioural evidence drawn from the learner's interaction trace, and teach-back tasks contribute evidence that the learner can explain the skill in their own words. Evidence is fused into an updated acquisition score and confidence interval. When the score crosses the criterion threshold and the confidence interval is sufficiently tight, the gate opens: the skill is marked acquired and downstream skills become eligible for the frontier.

Crucially, the gate is bidirectional. If the learner's performance on subsequent activities reveals that earlier evidence was misleading — for example, that a skill was passed via shallow pattern matching rather than genuine acquisition — the assessor downgrades the acquisition score and the gate may close, returning the learner to remediation activities for the affected skill. This contrasts with conventional progression in which once a unit is passed it is never revisited.

Operating Parameters

Operating parameters control gating strictness, assessment depth, and pacing dynamics. The criterion-threshold parameter, specified per skill, sets the acquisition score required for gate opening; foundational skills carry tighter thresholds than exploratory or enrichment skills. The confidence-width parameter sets the maximum width of the acquisition-score confidence interval permitted at gate opening; tighter widths require more evidence and reduce the chance of premature progression at the cost of additional assessment time.

The modality-mix parameter specifies the minimum diversity of evidence types required for a gate to open; a strict mix may require closed-form, applied, and teach-back evidence in combination, while a permissive mix accepts any single modality. The transfer-discount parameter reduces the credit assigned to evidence drawn from contexts highly similar to the training context, preferring evidence of transfer to novel contexts. The engagement-floor parameter halts content selection when the learner's engagement profile falls below a threshold, surfacing rest, motivation interventions, or modality changes rather than continuing to push activities into a disengaged state.

Pacing parameters govern macroscopic dynamics. The frontier-breadth parameter controls how many skills the learner may pursue in parallel, balancing breadth against depth. The remediation-priority parameter controls how aggressively the system returns to skills whose acquisition has degraded, preventing decay from accumulating. The escalation parameter specifies the conditions under which the AI-mediated assessment is supplemented or overridden by a human educator, ensuring that consequential decisions remain auditable and contestable.

Alternative Embodiments

One embodiment implements the orchestrator as a server-side service that selects activities and evaluates evidence centrally, with thin clients delivering content. A second embodiment pushes the orchestrator and assessor onto the learner's device, supporting offline use and reducing the surface on which learner data is exposed. A third embodiment splits the system, running content selection centrally for catalogue access while running assessment locally for privacy.

Embodiments differ in skill-graph authority. A platform-curated embodiment maintains a central skill graph authored by curriculum designers and shared across all learners. An educator-extended embodiment allows individual educators to add or refine skill nodes for their cohorts while inheriting the platform graph. A learner-personalised embodiment allows the learner to add aspirational skills whose paths are constructed by the orchestrator from existing nodes, supporting self-directed study within the same gating framework.

Alternative embodiments vary in assessor architecture. A single-model embodiment uses one model to score all evidence, simplifying deployment but coupling all skill assessments to one model's biases. An ensemble embodiment uses different specialised assessors per skill domain, improving fidelity at the cost of operational complexity. A human-in-the-loop embodiment routes a sampled fraction of assessments to human raters whose judgments calibrate the AI assessor and surface disagreement for review. Embodiments may also vary in evidence retention: a transient embodiment retains only acquisition scores and discards raw responses after assessment, prioritising privacy; a longitudinal embodiment retains responses for trajectory analysis and educator review under appropriate consent.

Composition With Other Primitives

The platform composes with the skill-gating primitive by reusing its gate-evaluation machinery: the same logic that decides whether an agent has met the criteria to unlock a tool decides whether a learner has met the criteria to advance to a new unit. The skill graph specialises the more general capability graph, and the assessor specialises the more general capability evaluator.

Composition with the cognitive-state primitive supplies engagement and trajectory signals, allowing pacing to respond to how the learner is engaging rather than only to what they are producing. Composition with the provenance-tracing primitive records, for each acquisition decision, the evidence that supported it and the assessor that produced it, enabling later audit by educators, parents, or accreditation bodies. Composition with the disclosure primitive produces transcripts and competence statements whose contents are verifiable against the recorded evidence, supporting articulation between platforms and credentialing systems.

Distinction From Prior Art

Existing adaptive-learning platforms adjust the difficulty of the next item based on item-response theory or related psychometric models but do not maintain explicit skill graphs whose gates are bound to multimodal evidence and whose criteria can be reconfigured by educators. Existing competency-based education systems require human assessment for gate decisions and do not support AI-mediated assessment with confidence-aware gating. Existing intelligent tutoring systems pursue mastery within a single domain but do not generalise to a skill-gating primitive shared with broader capability-control infrastructure. The present approach differs in that it specialises a general skill-gating primitive into the educational domain, supports bidirectional gates that can close when later evidence contradicts earlier acquisition, fuses heterogeneous evidence with confidence accounting, and composes with provenance and disclosure primitives to produce auditable competence records.

Implementation Considerations

Educational deployment introduces concerns that do not arise in agentic capability-gating. The first is fairness across learner populations. Assessors trained or tuned on data drawn predominantly from one population may produce systematically biased acquisition scores when applied to learners from underrepresented populations. Implementations must include population-stratified validation, expose the bias metrics produced by that validation, and route assessment of learners from underrepresented populations through ensemble or human-in-the-loop modes until parity is established. The second concern is consent and data governance. Learner state, evidence traces, and acquisition records are sensitive personal data, often involving minors. Implementations must support explicit consent regimes, retention limits aligned with applicable regulation, and learner-controlled access so that the learner or their guardian can inspect, export, and delete the records the platform holds.

The third concern is alignment with educational frameworks external to the platform. Schools, regulators, and credentialing bodies operate against curricular standards expressed in their own ontologies. Implementations should support mapping layers that relate platform skill nodes to external standards, so that gate-passage records can be reported in the form expected by the institution receiving them. The mapping should be auditable: each external claim made on the basis of platform evidence should be traceable to the specific gate decisions and the specific evidence underlying them, enabling external reviewers to verify claims against the underlying record.

A fourth concern is teacher and parent agency. Curriculum-gated platforms operate within institutional contexts in which teachers retain pedagogical authority and parents retain oversight rights. Implementations must expose configuration surfaces that allow teachers to override gate decisions for individual learners, to inject supplementary skills not present in the platform graph, and to suspend gating in favour of conventional sequencing when judged appropriate. Parent-facing surfaces should expose progression evidence in plain language, supporting informed engagement without requiring fluency in the underlying skill-graph machinery.

A fifth concern is adversarial behaviour by learners. Any assessment regime creates incentives to satisfy the assessor rather than acquire the underlying skill. Implementations should include integrity provisions: variation in assessment items to limit memorisation of specific tasks, behavioural-trace analysis to detect patterns inconsistent with genuine engagement, and periodic transfer-focused assessments that probe whether claimed skills generalise beyond the training contexts. When integrity violations are suspected, the system should suspend gate decisions for the affected skills and route the case to human review rather than silently degrading the acquisition score, preserving learner trust in the fairness of the system.

Disclosure Scope

The disclosure encompasses methods for representing curricula as skill graphs with prerequisite edges; methods for fusing multimodal evidence into per-skill acquisition scores with confidence intervals; methods for evaluating bidirectional gates that open on criterion satisfaction and close on subsequent contradicting evidence; methods for selecting activities by expected information gain over the learner frontier; and methods for composing curriculum gating with skill-gating, cognitive-state, provenance, and disclosure primitives. Embodiments addressing server-side, on-device, and split-orchestrator deployments fall within scope, as do platform-curated, educator-extended, and learner-personalised graph-authority models and single-model, ensemble, and human-in-the-loop assessor architectures.