Documentation Index
Hypotheses and Scope of the SRS Model
This document describes the non-mathematical hypotheses of the spaced repetition model (SRS) used in the application.
It complements:
- the mathematical SRS documentation
maths_srs.md,
- the architecture diagrams (Domain, Sessions, Application).
Its purpose is to make explicit the cognitive, pedagogical, and product choices that guide the model, in order to:
- avoid ambiguity during future evolutions,
- clearly distinguish what is intentionally simplified from what is genuinely missing,
- define the scope of the MVP.
1. Core Hypothesis: the SRS evaluates a task, not abstract knowledge
Each exercise (in the sense of: type of task) has its own SRS state.
The same content (word, rule, sentence) may therefore be associated with several distinct exercises, for example:
- language A → language B,
- language B → language A,
- recognition vs. active recall.
The SRS never evaluates a “global mastery” of a word or rule, but only the ability to succeed at one specific task.
This choice allows:
- avoiding any ambiguity about what it means to “know” an item,
- aligning the model with Anki (one note → multiple cards),
- keeping the SRS simple and local.
2. Hypothesis on the response signal: subjective self-assessment
The model assumes that the user is capable of honestly self-assessing after each exercise.
The possible responses (Easy / Good / Medium / Hard / Again) reflect:
- the correctness of the answer,
- the degree of hesitation felt,
- the perceived time to answer,
- confidence in the response.
Actual measured time is not used. The signal is intentionally subjective.
This choice is deliberate because:
- users perceive their own difficulty better than a raw time measurement would,
- it avoids heavy and fragile instrumentation,
- it is the model successfully used by Anki.
3. Hypothesis on delay: success takes precedence over elapsed time
A successful response to an exercise is never penalised, even after a significant delay.
The model considers that:
- if the user succeeds despite the delay, the mastery state was sufficient,
- elapsed time alone is not more reliable information than the success itself.
Delay is only taken into account when the exercise is failed, in order to:
- reset or weaken the SRS state,
- prevent repeated lucky successes from masking a real fragility.
This choice favours:
- SRS stability,
- user confidence,
- a simple and explainable system behaviour.
4. Hypothesis of local independence between exercises
Each exercise is treated as independent from the others.
The SRS does not model:
- skill transfer between exercises,
- dependencies between vocabulary and grammar,
- hierarchical relationships between items of knowledge.
This choice is intentional.
It rests on the idea that:
- good vocabulary coverage is critical,
- sentences serve mainly as exposure and contextualisation,
- a poorly mastered sentence should simply reappear sooner.
The approximation is considered acceptable as long as:
- words are correctly revised,
- errors on sentences lead to rapid repetition.
5. Hypothesis on the role of sentences
Sentences are not fundamental units of knowledge.
They serve primarily to:
- illustrate grammatical rules,
- provide real-world context,
- reinforce memorisation through repeated exposure.
The SRS for sentences may be less precise than that for words without compromising overall learning.
Aids (e.g. displaying word translations within a sentence) are acceptable and do not invalidate the exercise, since the primary objective remains exposure and comprehension.
6. Product hypothesis: priority on simplicity and explainability
The model favours:
- simple rules,
- predictable behaviour,
- complete explainability for both the user and the developer.
It does not aim for maximum theoretical optimality.
This implies in particular:
- no Bayesian probabilistic model,
- fixed global parameters,
- no automatic parameter learning.
These limitations are accepted for the MVP.
7. Session management (MVP)
Sessions are ephemeral objects.
Their role is to:
- orchestrate a sequence of exercises,
- collect responses,
- delegate SRS updates.
Exercise prioritisation (e.g. if 150 exercises are due but the daily maximum is 100) is performed before the session, in the Application layer.
Inside a session:
- no additional sorting is necessary,
- exercises are presented in the order defined at session creation.
8. Cognitive load management (MVP)
The model has no global representation of the user’s state (fatigue, declining performance).
For the MVP, management relies on simple rules:
- the user may interrupt a session at any time,
- a stop or pause may be suggested in case of repeated errors.
Occasional errors or a poor session should not permanently penalise the SRS state.
9. Explicit MVP scope
The MVP does not attempt to solve the following problems:
- fine-grained probabilistic memory modelling,
- skill transfer between items of knowledge,
- automatic SRS parameter learning,
- global model adaptation to the user.
These evolutions are considered post-MVP and must not influence current choices as long as the hypotheses above are respected.
10. Role of this document
This document serves as a reference for:
- justifying the SRS model choices,
- guiding future evolutions without distorting the system,
- preventing the introduction of features inconsistent with the founding hypotheses.
Any major SRS modification must be evaluated against the hypotheses described here.
TODO:
- Suggestion to stop or pause in case of repeated errors.
- Exercise prioritisation in the application layer based on recall probability (rather than interval).