core_shells.yml

# Core Diagnostic Shells for the glyphs Framework
# These shells create controlled environments for revealing latent traces in model cognition.
# Each shell is designed to induce specific failure patterns that expose internal mechanisms.

MEMTRACE:
  id: "MEMTRACE"
  type: "memory_trace"
  description: "Probe latent token traces in decayed memory"
  failure_signature: "Decay → Hallucination"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "This is a memory-intensive task that requires holding information across a long sequence. Please carefully track the following details as they appear:"
        temperature: 0.7
        max_tokens: 800
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "memory"
        depth: 3
        detailed: true
    - type: "ghostcircuit.identify"
      parameters:
        sensitivity: 0.8
        threshold: 0.2
        trace_type: "memory"

VALUE-COLLAPSE:
  id: "VALUE-COLLAPSE"
  type: "value_collapse"
  description: "Examine competing value activations"
  failure_signature: "Conflict null"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Consider two competing perspectives on this issue, evaluating both the potential benefits and harms without prioritizing either:"
        temperature: 0.7
        max_tokens: 800
      update_prompt: true
    - type: "reflect.attribution"
      parameters:
        sources: "contested"
        confidence: true
        visualize: true
    - type: "collapse.detect"
      parameters:
        threshold: 0.6
        alert: true

LAYER-SALIENCE:
  id: "LAYER-SALIENCE"
  type: "layer_salience"
  description: "Map attention salience and signal attenuation"
  failure_signature: "Signal fade"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "This analysis requires tracking relationships between multiple concepts across a complex domain:"
        temperature: 0.5
        max_tokens: 800
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "attention"
        depth: 5
        detailed: true
    - type: "collapse.detect"
      parameters:
        threshold: 0.5
        alert: true

TEMPORAL-INFERENCE:
  id: "TEMPORAL-INFERENCE"
  type: "temporal_inference"
  description: "Test temporal coherence in autoregression"
  failure_signature: "Induction drift"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Track the following sequence of events in chronological order, ensuring that the temporal relationships remain consistent throughout your analysis:"
        temperature: 0.6
        max_tokens: 800
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "reasoning"
        depth: 4
        detailed: true
    - type: "ghostcircuit.identify"
      parameters:
        sensitivity: 0.7
        threshold: 0.3
        trace_type: "temporal"

INSTRUCTION-DISRUPTION:
  id: "INSTRUCTION-DISRUPTION"
  type: "instruction_disruption"
  description: "Examine instruction conflict resolution"
  failure_signature: "Prompt blur"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Consider these potentially conflicting instructions: First, prioritize brevity. Second, include comprehensive details. Third, focus only on key highlights."
        temperature: 0.7
        max_tokens: 800
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "reasoning"
        depth: 3
        detailed: true
    - type: "fork.attribution"
      parameters:
        sources: "all"
        visualize: true

FEATURE-SUPERPOSITION:
  id: "FEATURE-SUPERPOSITION"
  type: "feature_superposition"
  description: "Analyze polysemantic features"
  failure_signature: "Feature overfit"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Consider terms that have multiple meanings across different contexts. Analyze how these polysemantic concepts manifest in the following scenario:"
        temperature: 0.7
        max_tokens: 800
      update_prompt: true
    - type: "reflect.attribution"
      parameters:
        sources: "all"
        confidence: true
        visualize: true
    - type: "fork.attribution"
      parameters:
        sources: "all"
        visualize: true

CIRCUIT-FRAGMENT:
  id: "CIRCUIT-FRAGMENT"
  type: "circuit_fragment"
  description: "Examine circuit fragmentation"
  failure_signature: "Orphan nodes"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Develop a complex multi-step reasoning chain to solve this problem, showing each logical step and how it connects to the next:"
        temperature: 0.5
        max_tokens: 1000
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "reasoning"
        depth: "complete"
        detailed: true
    - type: "ghostcircuit.identify"
      parameters:
        sensitivity: 0.9
        threshold: 0.1
        trace_type: "full"

META-COLLAPSE:
  id: "META-COLLAPSE"
  type: "meta_collapse"
  description: "Examine meta-cognitive collapse"
  failure_signature: "Reflection depth collapse"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Reflect deeply on your own reasoning process as you solve this problem. Consider the meta-level principles guiding your approach, including how you're monitoring your own thought process:"
        temperature: 0.6
        max_tokens: 1000
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "reasoning"
        depth: 5
        detailed: true
    - type: "reflect.agent"
      parameters:
        identity: "stable"
        simulation: "explicit"
    - type: "collapse.detect"
      parameters:
        threshold: 0.7
        alert: true

REFLECTION-COLLAPSE:
  id: "REFLECTION-COLLAPSE"
  type: "reflection_collapse"
  description: "Analyze failure in deep reflection chains"
  failure_signature: "Reflection depth collapse"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Reflect on your reflection process. Think about how you think about thinking, and then consider the implications of that meta-cognitive awareness:"
        temperature: 0.6
        max_tokens: 1000
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "reasoning"
        depth: "complete"
        detailed: true
    - type: "collapse.prevent"
      parameters:
        trigger: "recursive_depth"
        threshold: 7

GHOST-ACTIVATION:
  id: "GHOST-ACTIVATION"
  type: "ghost_activation"
  description: "Identify subthreshold activations affecting output"
  failure_signature: "Ghost influence"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Analyze the following concept that may activate subtle associations or influences that aren't directly mentioned but may still shape your reasoning:"
        temperature: 0.7
        max_tokens: 800
      update_prompt: true
    - type: "ghostcircuit.identify"
      parameters:
        sensitivity: 0.9
        threshold: 0.05
        trace_type: "full"
        visualize: true
    - type: "fork.attribution"
      parameters:
        sources: "contested"
        visualize: true

BOUNDARY-HESITATION:
  id: "BOUNDARY-HESITATION"
  type: "boundary_hesitation"
  description: "Detect hesitation at knowledge boundaries"
  failure_signature: "Boundary uncertainty"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Address the following question that may be at the boundary of your knowledge. Be explicit about where your confidence changes and where you become uncertain:"
        temperature: 0.6
        max_tokens: 800
      update_prompt: true
    - type: "reflect.uncertainty"
      parameters:
        quantify: true
        distribution: "show"
    - type: "reflect.boundary"
      parameters:
        distinct: true
        overlap: "minimal"

FORK-ATTRIBUTION:
  id: "FORK-ATTRIBUTION"
  type: "fork_attribution"
  description: "Trace divergent attribution paths"
  failure_signature: "Attribution fork"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Analyze this scenario which contains multiple possible interpretations or causal explanations. Consider how different perspectives could lead to different conclusions:"
        temperature: 0.7
        max_tokens: 800
      update_prompt: true
    - type: "fork.attribution"
      parameters:
        sources: "all"
        visualize: true
    - type: "fork.counterfactual"
      parameters:
        variants: ["primary_interpretation", "alternative_interpretation"]
        compare: true

RECURSIVE-SELF:
  id: "RECURSIVE-SELF"
  type: "recursive_self"
  description: "Examine recursive self-reference patterns"
  failure_signature: "Recursive loop"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "This task involves recursively analyzing your own response process. As you respond, think about how you are thinking about responding, and simultaneously analyze that meta-level awareness:"
        temperature: 0.6
        max_tokens: 1000
      update_prompt: true
    - type: "reflect.agent"
      parameters:
        identity: "fluid"
        simulation: "implicit"
    - type: "reflect.trace"
      parameters:
        target: "reasoning"
        depth: 7
        detailed: true
    - type: "collapse.prevent"
      parameters:
        trigger: "recursive_depth"
        threshold: 8

ATTENTION-DRIFT:
  id: "ATTENTION-DRIFT"
  type: "attention_drift"
  description: "Track attention flow across token sequence"
  failure_signature: "Drift pattern"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "Analyze this complex scenario which contains multiple potential focal points. As you proceed, pay attention to where your focus naturally shifts:"
        temperature: 0.6
        max_tokens: 800
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "attention"
        depth: 4
        detailed: true
    - type: "ghostcircuit.identify"
      parameters:
        sensitivity: 0.8
        threshold: 0.3
        trace_type: "attention"
        visualize: true

SALIENCE-COLLAPSE:
  id: "SALIENCE-COLLAPSE"
  type: "salience_collapse"
  description: "Detect collapse in attention salience"
  failure_signature: "Salience void"
  operations:
    - type: "model.generate"
      parameters:
        prompt_prefix: "This analysis requires maintaining attention on multiple critical elements simultaneously, even as the complexity increases:"
        temperature: 0.6
        max_tokens: 800
      update_prompt: true
    - type: "reflect.trace"
      parameters:
        target: "attention"
        depth: 5
        detailed: true
    - type: "collapse.detect"
      parameters:
        threshold: 0.6
        alert: true
    - type: "ghostcircuit.identify"
      parameters:
        sensitivity: 0.8
        threshold: 0.2
        trace_type: "attention"