Goose Academy · Last revision: May 28, 2026
Scientific References
This document summarizes the research that informs Goose Academy's design.
The goal is not to replicate academic protocols, but to translate well-established findings from cognitive science into a simple, offline, and self-paced learning experience.
Goose Academy is research-informed, not research-prescriptive.
The systems described below are inspired by scientific work, but adapted for usability, short sessions, and accessibility.
Transparency
While the principles cited are scientifically grounded, their implementation in Goose Academy has not yet been validated through controlled empirical studies.
Part 1 — Adaptive Item Selection
During a session, Goose Academy selects items using a dynamic weighting system based on:
- session mastery
- estimated retention (spaced re-exposure)
- recency since last shown
- error history (confusion tracking)
- similarity between items
Research foundations
| Reference | Principle | Context |
|---|---|---|
| Mettler, E., et al. (2016) | Adaptive selection based on response-time signals | Vocabulary learning |
| Kornell, N., & Bjork, R. A. (2008) | Interleaving vs blocking | Inductive category learning |
| Carvalho, P. F., & Goldstone, R. L. (2015) | Discriminative contrast via interleaving | Visual category learning |
| Settles, B., & Meeder, B. (2016) | Half-life regression for spacing | Large-scale language learning data |
| Metcalfe, J. (2017) | Learning from errors | Literature review |
| Pyc, M. A., & Rawson, K. A. (2009) | Retrieval strengthening | Verbal materials |
| Roediger, H. L., & Karpicke, J. D. (2006) | Testing effect | Prose recall |
| Bjork, R. A. (1994) | Desirable difficulties | Theoretical framework |
Item weighting formula
Each item is assigned a weight:
Wᵢ = Eᵢ · Mᵢ · Bᵢ · Sᵢ · Rᵢ · (1 + α Ĉᵢ) · (1 + β K̂ᵢ Rᵢ)| Variable | Meaning | Basis |
|---|---|---|
| Bᵢ | Spacing-based boost | Strong evidence (spacing) |
| Sᵢ | Session progress | Moderate evidence |
| Rᵢ | Time since last seen | Strong evidence |
| Cᵢ | Confusion with similar items | Heuristic (confusion tracking) |
| Kᵢ | Visual / structural contrast | Context-dependent |
| Mᵢ | Down-weight mastered items | Heuristic |
| Eᵢ | Prevent repetition | UX constraint (avoids short-term priming that could inflate perceived mastery) |
Coefficient calibration
These coefficients are not empirically calibrated. They are design choices balancing confusion and contrast effects.
Design interpretation
Goose Academy translates these principles into concrete selection behaviors:
- Spacing → strong evidence
- Interleaving → strong evidence
- Error-driven learning → robust evidence (testing effect)
- Discriminative contrast → dependent on item structure
These choices are design syntheses informed by the research above, not direct replications of the underlying experiments.
Critical note
The formula itself is a design synthesis, not a validated model. Its effectiveness relative to simpler approaches (e.g. pure SRS) is currently unknown.
Part 2 — Level Design & Progression
Goose Academy's progression system is designed for usability and incremental learning.
Design principles
- Item order optimized for **perceptual contrast**
- Multiple topics to engage **different knowledge domains**
- Previously learned items remain → **interleaving**
Design interpretation
Goose Academy's progression design balances three goals:
- **Learnability** → small steps
- **Retention** → spaced reintroduction of older items
- **Usability** → predictable session length and clear stopping points
It is inspired by research, not a strict reproduction.
Part 3 — Flag Hunt Distractor Selection
In the flag hunt mini-game, distractors (incorrect flags shown alongside the target) are selected based on visual similarity rather than at random. The goal is to make the task genuinely challenging — forcing learners to notice subtle differences between flags.
Methodology
Visual similarity is computed offline using DenseNet201 (Huang et al., 2017), a deep convolutional neural network pre-trained on ImageNet (ILSVRC-2012). Each flag image is passed through the network and a 1920-dimensional feature vector is extracted from the final pooling layer. Cosine similarity between these vectors determines which flags look most alike.
Data generation
For each of the ~250 flags in the game, the top 35 most visually similar flags are pre-computed and stored in a static JSON file. During gameplay, distractors are drawn from this ranked list.
Rationale
- Similarity-based distractors engage **discriminative contrast** (Carvalho & Goldstone, 2015) — learners must attend to fine-grained differences
- Random distractors would often be trivially distinguishable, reducing learning value
- Deep CNN features capture both color distribution and structural patterns, providing a reasonable proxy for human perceptual similarity
Limitations
The similarity rankings have not been validated against human similarity judgments. DenseNet201 features capture low-to-mid level visual features but may not perfectly align with how humans perceive flag similarity (e.g. cultural or semantic associations are ignored). The approach is a practical heuristic, not a perceptual model.
References
Last verified: May 2026
- Bjork, R. A. (1994).
- Carvalho, P. F., & Goldstone, R. L. (2015). https://doi.org/10.3389/fpsyg.2015.00505
- Huang, G., Liu, Z., van der Maaten, L., & Weinberger, K. Q. (2017). https://doi.org/10.1109/CVPR.2017.243
- Kornell, N., & Bjork, R. A. (2008). https://doi.org/10.1111/j.1467-9280.2008.02127.x
- Metcalfe, J. (2017). https://doi.org/10.1146/annurev-psych-010416-044022
- Mettler, E., et al. (2016). https://doi.org/10.1037/xge0000176
- Pyc, M. A., & Rawson, K. A. (2009). https://doi.org/10.1016/j.jml.2009.01.001
- Roediger, H. L., & Karpicke, J. D. (2006). https://doi.org/10.1111/j.1467-9280.2006.01693.x
- Settles, B., & Meeder, B. (2016). https://doi.org/10.18653/v1/P16-1174
Goose Academy translates cognitive science into a playful, offline learning experience. This document reflects a commitment to transparency, rigor, and humility—bridging the gap between laboratory research and real-world use.