Language PR1 — Canon + Core Data Model

Goal: lock the language-learning primitives into backend objects.

Build objects

LanguageScenario
InteractionContract
LearnerAttempt
LanguageArtifact
FeedbackClaim
SkillNode
LearnerSkillState
ValidationEvent

Exit condition

A learner attempt can be stored, evaluated, attached to a skill, and traced.

Language PR2 — Scenario Engine

Goal: create bounded, repeatable learning scenarios.

Build

Scenario Card
Role
Goal
Target vocabulary
Target forms
Allowed inputs
Success condition
Failure branches

Exit condition

System can run a simple scenario repeatedly with comparable attempts.

Language PR3 — Feedback Claim + Repair Loop

Goal: replace raw AI feedback with structured, testable feedback.

Build

FeedbackClaim:
- error
- evidence
- rule/pattern
- confidence
- next action
- validation task

RepairLoop:
attempt → hint → retry → correction → validation

Exit condition

Feedback always produces a follow-up task or validation event.

Language PR4 — SkillGraph + Learner State

Goal: track actual progression.

Build

SkillGraph
Skill dependencies
Learner skill states:
unknown / seen / partial / validated / transferred / retained

Exit condition

Progression is evidence-based, not lesson-completion-based.

Language PR5 — Granular Evaluation + Error Typology

Goal: decompose learner errors.

Build error layers

phonological
lexical
grammatical
syntactic
pragmatic
cultural

Exit condition

A learner answer can be partially correct and partially incorrect across layers.

Language PR6 — Scaffold Ladder

Goal: staged assistance.

Build levels

hint
partial scaffold
guided answer
explicit correction

Exit condition

System does not jump to full correction unless scaffold levels fail or confidence requires it.

Language PR7 — Transfer + Reuse

Goal: prove learning survives variation.

Build transfer levels

repeat
substitute
recombine
cross-scenario use
spontaneous use

Exit condition

A skill only advances after reuse or transfer evidence.

Language PR8 — Cognitive Load Controls

Goal: prevent overloaded tasks.

Build limits

max new vocabulary
max new grammar
max new scenario novelty
max feedback density

Exit condition

Scenario generation respects learner state and load constraints.

Language PR9 — Pronunciation / Speech Extension

Goal: add conservative speech feedback.

Build

pronunciation attempt capture
confidence threshold
phonological error claim
fallback / uncertainty marking

Exit condition

Low-confidence speech feedback is not treated as authoritative.

Language PR10 — Authority Boundary + Governance

Goal: handle ambiguity, culture, dialect, register, and low confidence.

Build

AuthorityBoundary
AI provisional flag
human/community validation flag
uncertainty marker
blocked authoritative correction

Exit condition

The system can say: “AI may suggest, but may not decide.”

Minimal First Build

Start with:

Language PR1 → PR2 → PR3 → PR4

That gives the usable skeleton:

Scenario
→ Attempt
→ Feedback Claim
→ Repair
→ Artifact
→ SkillGraph Update

Then add evaluation depth, transfer, load, speech, and governance.

Core issue title set

Language PR1: Add Canon-Aligned Language Data Model
Language PR2: Add Scenario Engine
Language PR3: Add Feedback Claim and Repair Loop
Language PR4: Add SkillGraph and Learner State Tracking
Language PR5: Add Granular Evaluation and Error Typology
Language PR6: Add Scaffold Ladder
Language PR7: Add Transfer and Reuse Tracking
Language PR8: Add Cognitive Load Controls
Language PR9: Add Conservative Pronunciation Feedback
Language PR10: Add Authority Boundary Governance

Core Canon

1. Language learning is controlled skill acquisition, not conversation.

A language module must not be treated as open AI chat. It must structure practice through constrained tasks, observable attempts, feedback, reuse, and validation.

2. Every learning interaction must have an interaction contract.

Each task must declare:

mode
target skill
allowed input
expected output
evaluation method
success condition

3. Scenarios are the primary learning container.

Practice should occur inside bounded, repeatable scenarios with roles, goals, vocabulary, target forms, cultural context, and failure branches.

4. Learner attempts must be captured as artifacts.

Every meaningful learner output should be stored, including original attempt, context, target skill, correction, validation state, and reuse history.

5. Feedback is a claim, not a message.

Every correction must state:

what is wrong
why it is wrong
evidence
rule or pattern
confidence
next validation task

6. Feedback must be validated through learner action.

A correction is not proven useful until the learner repairs, reuses, transfers, or stabilizes the corrected form.

7. Repair is a bounded learning loop.

The system must support:

attempt → hint → retry → correction → validation

but limit unproductive guessing and over-assistance.

8. Scaffolding must be staged.

Feedback should move through levels:

hint → partial scaffold → guided answer → explicit correction

not jump immediately to full answer unless necessary.

9. Skills must be represented in a SkillGraph.

Language knowledge should be decomposed into learnable units with dependencies:

vocabulary
grammar
pronunciation
syntax
pragmatics
cultural convention

10. Learner state must be evidence-based.

A learner does not “know” a skill because they saw it once. Mastery requires evidence across recognition, guided production, independent production, transfer, and retention.

11. Evaluation must be granular.

A sentence or utterance is not simply right or wrong. Evaluation must decompose errors by layer:

phonological
lexical
grammatical
syntactic
pragmatic
cultural

12. Transfer must be tracked explicitly.

Progression must test movement across levels:

repeat
substitute
recombine
cross-scenario use
spontaneous use

13. Cognitive load must be constrained.

Tasks must not introduce too much novelty at once. The system should control new vocabulary, new grammar, new context, and feedback density.

14. Pronunciation feedback must be conservative.

Speech/tone feedback requires confidence thresholds, granular diagnosis, and fallback behavior. Low-confidence pronunciation feedback should not be presented as authoritative.

15. Authority boundaries must be explicit.

AI may generate, scaffold, and propose, but culturally sensitive, dialectal, ambiguous, or low-confidence claims require human/community authority or marked uncertainty.

Minimal Language Loop

Scenario
→ Interaction Contract
→ Learner Attempt
→ Granular Evaluation
→ Feedback Claim
→ Scaffolded Repair
→ Language Artifact
→ Reuse
→ Transfer Test
→ Validation
→ SkillGraph Update
→ Progression

Canon Summary

Build a Substrate-governed language acquisition system:

bounded scenarios
+ decomposed skills
+ artifact capture
+ claim-based feedback
+ scaffolded repair
+ validated reuse
+ transfer tracking
+ evidence-based progression

Not a chatbot. Not a tutor persona. A governed learning loop.