Skip to content

fragility.md

Latest commit

 

History

History
172 lines (109 loc) · 4.9 KB

fragility.md

File metadata and controls

172 lines (109 loc) · 4.9 KB

Fragility Analysis

Fragility analysis now ranks interventions, not concepts.

The core question is:

  • "What exact thing should I inspect next if I want to change or stabilize what I currently believe?"

The fragility subsystem lives in the render layer. It reads from:

  • ATMS futures
  • conflict topology
  • grounded-rule bundles
  • the ASPIC bridge

It does not mutate source storage. It produces a typed FragilityReport whose main payload is a ranked tuple of intervention targets.

Canonical Surface

The public surface is intervention-first:

  • InterventionTarget
  • RankedIntervention
  • FragilityReport

Concepts are still present as subjects in provenance and scoring context, but they are no longer the primary ranked output.

Current intervention families:

  • assumption
  • missing_measurement
  • conflict
  • ground_fact
  • grounded_rule
  • bridge_undercut

Each target carries:

  • canonical identity
  • family
  • kind
  • cost tier
  • provenance
  • typed payload

Family Semantics

Assumptions

ATMS assumption targets come from QueryableAssumption and are ranked by how strongly they participate in witness futures that flip a concept's status.

The scoring uses weighted_epistemic_score(). For currently supported concepts, many flip witnesses means high fragility. For currently unsupported concepts, many future entries reduce fragility because the concept is actually well-supported under future completions.

Missing Measurements

Missing-measurement targets are discovered from parameterization inputs that have no active claims. They are intervention targets because the next action is to obtain a measurement, not to stare at the downstream concept.

Their current score is a discovery heuristic based on how many downstream subjects depend on the missing value.

Conflicts

Conflict targets represent canonical claim pairs. Their score is the maximum grounded-extension delta from removing either side of the pair, via score_conflict().

Ground Facts

Ground-fact targets come from GroundedRulesBundle.sections. Their score combines:

  • section-level uncertainty
  • how many grounded-rule antecedents depend on the fact

This keeps the identity surface exact while making the score depend on actual grounded structure rather than a flat constant.

Grounded Rules

Grounded-rule targets are keyed by the existing bridge-grounding rule name:

  • rule_doc.id#<canonical substitution>

Their current score combines:

  • antecedent count
  • how many grounded defeaters target the rule

Bridge Undercuts

Bridge undercut targets are keyed by the existing synthesized name:

  • <defeater_name>-><target_rule_name>

Their score is computed from actual bridge behavior:

  • number of attacks launched by arguments using that undercut rule
  • number of defeats that actually survive preference filtering

Ranking Policies

The system does not currently pretend to have a single literature-settled cross-family scalar.

Instead it exposes explicit ranking policies:

  • heuristic_roi
  • family_local_only
  • pareto

heuristic_roi is the default. It ranks by family-local fragility divided by cost tier.

family_local_only keeps families separate instead of forcing a cross-family total order.

pareto returns the non-dominated frontier over local fragility and cost tier.

Pairwise Interactions

Interaction detection currently operates on assumption interventions. It looks at ATMS witness structure and reports:

  • synergistic
  • redundant
  • mixed
  • independent
  • unknown

This is where the Howard 1966 non-additivity story appears operationally: two assumptions can have more or less joint value than their individual scores suggest.

Additional Scoring Helpers

Two standalone scoring helpers still exist outside the main ranking pipeline:

  • opinion_sensitivity()
  • imps_rev()

They remain available as local analytical tools, but they are not yet primary ranking contributors.

CLI Usage

# Default ranking across all intervention families
uv run pks world fragility

# Only ATMS and conflict contributors
uv run pks world fragility --skip-grounding --skip-bridge

# Pareto frontier instead of heuristic ROI
uv run pks world fragility --ranking-policy pareto

# JSON output
uv run pks world fragility --format json

Text output now shows intervention-oriented columns:

  • Rank
  • Score
  • ROI
  • Cost
  • Family
  • Kind
  • Intervention

JSON output includes:

  • world_fragility
  • analysis_scope
  • interventions
  • interactions

References

  • Howard 1966: value of information and non-additive joint value
  • Gärdenfors 1988 and Dixon 1993: entrenchment/revision background for ATMS-side fragility
  • Odekerken 2025: stability and relevance under incomplete ASPIC+ information
  • Al Anaissy 2024: local impact measures for gradual semantics
  • Diller 2025: grounded fact/rule instances as first-class grounded units
  • Ballester-Ripoll 2024: interaction-sensitive sensitivity analysis