Reasoning Tasks Progression Protocol

This document contains the current Reasoning Gym MVP family progression notes from the uploaded family-pack markdown documents.

Initial 3-Family Recommendation

1. Relation Fit - relation validation / relation instantiation / same deep relation
2. Must Follow - deduction / necessary consequence / constraint propagation
3. Best Rule So Far - induction / updating / rival-rule tracking

Family 1: Relation Fit

This pack contains the first fluid-reasoning family for the Reasoning Gym MVP.

This pack implements Family 1 only.

Included Files

• relation-fit.real-world.examples.json - 264 authored real-world items
• relation-fit.nonsense-grammar.spec.json - pronounceable nonsense generation rules, morphology, answer types, complexity rules, and adaptive rules
• relation-fit.generator.js - ES module to generate nonsense items and adaptive block plans
• relation-fit.schema.json - JSON schema for item objects

Family Definition

Core task:

> Identify whether one or more options instantiate a target abstract relation.

Supported subtypes:

• same_relation_mcq
• select_all_valid
• relation_satisfaction
• multi_relation_validation

Answer Types

• single_choice
• multi_select
• optional later: true_false

Complexity Axes

• binding_load
• uncertainty_level
• control_burden

Adaptive Rule

Use a block-level UP / HOLD / DOWN decision, mirroring the broader integrated-game logic:

1. stabilise first
2. swap wrapper
3. go faster
4. then increase load / relation complexity

That means the family should keep the same deep kernel while varying the wrapper before pushing tier.

Suggested Webapp Use

• Load real-world items from the JSON file for familiar-wrapper blocks
• Use relation-fit.generator.js for nonsense / abstract-wrapper blocks
• Persist block state with:
• current tier
• wrapper mode
• speed mode
• recent accuracy
• role-reversal error rate
• wrapper cost

Notes

• Tier 1-2 = binary relation validation
• Tier 3 = wrapper swap and multi-select pressure
• Tier 4 = chained or 3-argument relations
• Tier 5 = graph-role / structural isomorphism later

Family 2: Must Follow

This pack contains the second fluid-reasoning family for the Reasoning Gym MVP.

This pack implements Family 2 only.

Included Files

• must-follow.real-world.examples.json - 360 authored real-world items
• must-follow.nonsense-grammar.spec.json - pronounceable nonsense generation rules, morphology, answer types, complexity rules, and adaptive rules
• must-follow.generator.js - ES module to generate nonsense items and adaptive block plans
• must-follow.schema.json - JSON schema for item objects
• must-follow.generator.txt - plain-text copy of the JS module

Family Definition

Core task:

> Decide which conclusion must follow from the premises.

Supported subtypes:

• must_follow_tf
• best_conclusion_mcq
• select_all_must_follow
• optional later: cannot_follow_mcq

Answer Types

• true_false
• single_choice
• multi_select

Complexity Axes

• binding_load
• uncertainty_level
• control_burden

Complexity Progression

• Tier 1-2 = short transitive chains and simple quantifier chains
• Tier 3 = longer chains, extra premises, stronger distractors, wrapper swap
• Tier 4 = mixed positive and negative constraints, multi-select pressure
• Tier 5 = multi-conclusion chains with high control burden

Adaptive Rule

Use a block-level UP / HOLD / DOWN decision, mirroring the broader integrated-game logic:

1. stabilise first
2. swap wrapper
3. go faster
4. then increase load / relation complexity

That means the family should keep the same deep deduction kernel while varying the wrapper before pushing tier.

Suggested Webapp Use

• Load real-world items from the JSON file for familiar-wrapper blocks
• Use must-follow.generator.js for nonsense / abstract-wrapper blocks
• Persist block state with:
• current tier
• wrapper mode
• speed mode
• recent accuracy
• transitive-reversal error rate
• quantifier-scope error rate
• conditional-chain error rate
• wrapper cost

Notes

• The real-world file includes order, set-inclusion, set-exclusion, conditional-chain, and multi-select chain items.
• The nonsense generator currently supports transitive, quantifier, and conditional chain items.
• The family is designed to expand later into richer syllogistic or relational-constraint variants.

Family 3: Best Rule So Far

This pack contains the third fluid-reasoning family for the Reasoning Gym MVP.

This pack implements Family 3 only.

Included Files

• best-rule-so-far.real-world.examples.json - 360 authored real-world items
• best-rule-so-far.nonsense-grammar.spec.json - pronounceable nonsense generation rules, morphology, answer types, complexity rules, and adaptive rules
• best-rule-so-far.generator.js - ES module to generate nonsense items and adaptive block plans
• best-rule-so-far.schema.json - JSON schema for item objects
• best-rule-so-far.generator.txt - plain-text copy of the JS module

Family Definition

Core task:

> Decide which rule currently best explains the observations, which rules remain live, or whether confidence in a current best rule should rise.

Supported subtypes:

• best_rule_so_far_mcq
• confidence_update_tf
• select_all_consistent
• optional later: rule_ranking_later

Answer Types

• single_choice
• true_false
• multi_select

Complexity Axes

• binding_load
• uncertainty_level
• control_burden

Complexity Progression

• Tier 1 = simple one-dimensional alternation or +1 ladder rules
• Tier 2 = cycle-of-3 and +2 ladder rules with stronger distractors
• Tier 3 = pair-repeat and longer sequences, plus more rival rules
• Tier 4 = two-feature rules and more confidence-update items
• Tier 5 = high-load feature rules, noisy updates, and multi-select filtering

Adaptive Rule

Use a block-level UP / HOLD / DOWN decision, mirroring the broader integrated-game logic:

1. stabilise first
2. swap wrapper
3. go faster
4. then increase load / rule complexity

That means the family should keep the same deep induction kernel while varying wrapper before pushing tier.

Suggested Webapp Use

• Load real-world items from the JSON file for familiar-wrapper blocks
• Use best-rule-so-far.generator.js for nonsense / abstract-wrapper blocks
• Persist block state with:
• current tier
• wrapper mode
• speed mode
• recent accuracy
• alternation confusion rate
• cycle confusion rate
• feature-rule error rate
• overcommitment rate
• multi-select overreach
• wrapper cost

Notes

• The real-world file includes one-dimensional sequence rules, confidence-update items, and multi-select consistency items.
• The nonsense generator supports alternation, cycle, ladder-step, pair-repeat, and two-feature items.
• The family is designed to expand later into richer causal-model, matrix-rule, and verbal-category induction variants.