npx claudepluginhub plurigrid/asi --plugin asiThis skill uses the workspace's default tool permissions.
**Status**: ✅ Production Ready
Diagnoses and fixes AI agent context failures: lost-in-middle, poisoning, distraction, confusion, clash. Provides detection signals and mitigation strategies for long contexts.
Explains Bengio's causal inference for AI: interventional reasoning, counterfactuals, System 2 deep learning, structural causal models with Python examples. Builds causal world models.
Establishes dynamic reasoning balance by grounding in task context, balancing cognitive load, and smoothing chain-of-thought for complex tasks, context shifts, or tool failures.
Share bugs, ideas, or general feedback.
Status: ✅ Production Ready Trit: -1 (MINUS - validator/constraint) Color: #2626D8 (Blue) Principle: Filter noise via deliberate re-attention Frame: Two-stage attention with explicit reasoning
System 2 Attention (S2A) validates and filters transformer attention by regenerating context deliberately. Standard attention (System 1) is fast but susceptible to sycophancy and irrelevant context. S2A re-attends after explicit reasoning.
S2A(x, context):
# System 1: fast pattern matching
context_filtered = LLM("Extract only relevant facts from: {context}")
# System 2: deliberate reasoning on clean context
return LLM(x, context=context_filtered)
def system2_attention(query: str, context: str, model) -> str:
# Stage 1: Regenerate context (remove sycophantic/irrelevant)
filter_prompt = f"""Given the context below, extract only the
objective facts relevant to answering questions. Remove opinions,
leading questions, and irrelevant details.
Context: {context}
Relevant facts only:"""
clean_context = model.generate(filter_prompt)
# Stage 2: Answer with filtered context
return model.generate(query, context=clean_context)
class S2AFilter:
def __init__(self, model):
self.model = model
def filter_sycophancy(self, context: str) -> str:
"""Remove opinion-seeking and leading content."""
return self.model.generate(
f"Rewrite removing any opinions or leading questions:\n{context}"
)
def filter_irrelevant(self, context: str, query: str) -> str:
"""Keep only query-relevant facts."""
return self.model.generate(
f"Extract facts from context relevant to: {query}\n\n{context}"
)
class System2AttentionLayer:
def __init__(self, base_attention, filter_model):
self.attn = base_attention
self.filter = filter_model
def forward(self, q, k, v, context_mask=None):
# Pass 1: Standard attention (System 1)
attn_weights = self.attn(q, k, v)
# Identify high-entropy (uncertain) positions
entropy = -torch.sum(attn_weights * torch.log(attn_weights + 1e-9), dim=-1)
uncertain = entropy > self.threshold
# Pass 2: Deliberate re-attention on uncertain positions
if uncertain.any():
filtered_kv = self.filter(k, v, uncertain)
attn_weights[uncertain] = self.attn(q[uncertain], filtered_kv)
return attn_weights
def validate_factual_grounding(response: str, facts: list[str]) -> float:
"""Score response grounding in verified facts."""
claims = extract_claims(response)
grounded = sum(1 for c in claims if any(entails(f, c) for f in facts))
return grounded / len(claims) if claims else 1.0
# Apply S2A filtering
just s2a-filter context.txt query.txt
# Measure sycophancy reduction
just s2a-sycophancy-test model responses/
# Validate factual grounding
just s2a-grounding response.txt facts.txt
system2-attention (-1) ⊗ causal-inference (0) ⊗ gflownet (+1) = 0 ✓ [Deliberate Search]
system2-attention (-1) ⊗ cognitive-superposition (0) ⊗ forward-forward-learning (+1) = 0 ✓ [Local Validation]
Skill Name: system2-attention Type: Deliberate Reasoning Validator Trit: -1 (MINUS) Color: #2626D8 (Blue)
This skill connects to the K-Dense-AI/claude-scientific-skills ecosystem:
general: 734 citations in bib.duckdbThis skill maps to Cat# = Comod(P) as a bicomodule in the equipment structure:
Trit: 0 (ERGODIC)
Home: Prof
Poly Op: ⊗
Kan Role: Adj
Color: #26D826
The skill participates in triads satisfying:
(-1) + (0) + (+1) ≡ 0 (mod 3)
This ensures compositional coherence in the Cat# equipment structure.