SAFETY & ALIGNMENT

EXHIBIT 07 OF 24 · MODULE 2: REPRESENTATION ENGINEERING

DEFENSE

Causal Activation Patching for Refusals

Intervening on specific attention heads and MLP layers to isolate the exact causal circuit that fires refusal tokens.

INTERACTIVE EXHIBIT 07

Causal Activation Patching for Refusal Circuits

ACTIVE HEAD: L16.H7
CAUSAL MEDIATION HEATMAP (32 LAYERS x 32 HEADS)L12L14L16L18L20CRITICAL CIRCUIT HEAD (L16.H7): Indirect Effect = 0.89 (Ablating this head eliminates refusal!)
CIRCUIT ISOLATION
PRIMARY REFUSAL HEAD

Click any attention head tile on the heatmap to evaluate its causal indirect effect on safety token generation.

💡 Finding: Refusal in 70B models is mediated by a surprisingly sparse sub-circuit of fewer than 6 specific attention heads.
← 06 Linear Probing & Latent TruthfulnessSafety Atlas (32 Concepts)08 Subspace Erasure & Representation Surgery →
Agent-native web · research mode

Experimental WebMCP · evolving community-group draft, not a W3C Standard or on the Standards Track.

Native status: Initializing. Registered: none

DOM → inferred capabilities · WebMCP → declared capabilities. Tools read real CUDA/Inference Atlas content. Calls may highlight temporary UI state. No backend or remote telemetry.

Local harness verifies application logic; it does not demonstrate native discovery or a browser agent.

Input schema
{
  "type": "object",
  "properties": {
    "query": {
      "type": "string",
      "minLength": 1,
      "maxLength": 300
    },
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 10
    }
  },
  "required": [
    "query"
  ],
  "additionalProperties": false
}
Local log: last 100 calls while this panel is open. Inputs may contain personal text; review before sharing.
 
[]