Exp 0 - qualitative agreement check (real gemma-2-2b SAE)
Layer 12 · gemma-2-2b / 12-res-matryoshka-dc · blocks.12.hook_resid_post · edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20
For each block pair we compare edges the metrics KEEP (survivors) against edges they REJECT despite passing the crude coverage test. Read the parent/child labels: survivors should be semantically related; rejected edges should look like frequency / co-occurrence artifacts. Labels from Neuronpedia.
Block pair 0->1
survivor (8)
27 -> 410 R=0.53 F=0.05 recon_gain=0.288 recon=Y surv=0.61 p_fires=6%
parent 27 : mathematical expressions or equations related to inequalities
child 410 : references to financial transactions or economic concepts
42 -> 318 R=0.70 F=0.37 recon_gain=0.271 recon=Y surv=0.94 p_fires=5%
parent 42 : items related to programming and database operations
child 318 : code-related constructs and syntax
54 -> 466 R=0.77 F=0.04 recon_gain=0.267 recon=Y surv=0.95 p_fires=8%
parent 54 : formal legal terminology and procedural language
child 466 : phrases that contain the word “In” or its variations, indicating a focus on contextual information
69 -> 174 R=0.62 F=0.18 recon_gain=0.205 recon=Y surv=0.53 p_fires=11%
parent 69 : phrases that indicate relationships, particularly in the context of implementation and effect
child 174 : the presence of the word “in” across different contexts
69 -> 466 R=0.75 F=0.03 recon_gain=0.177 recon=Y surv=0.76 p_fires=11%
parent 69 : phrases that indicate relationships, particularly in the context of implementation and effect
child 466 : phrases that contain the word “In” or its variations, indicating a focus on contextual information
54 -> 232 R=0.70 F=0.10 recon_gain=0.168 recon=Y surv=0.67 p_fires=8%
parent 54 : formal legal terminology and procedural language
child 232 : definitive articles and the presence of the word “the” in a document
10 -> 370 R=0.74 F=0.17 recon_gain=0.164 recon=Y surv=1.16 p_fires=7%
parent 10 : names of individuals in various contexts
child 370 : proper names and specific identifiers within text
22 -> 426 R=0.55 F=0.27 recon_gain=0.143 recon=Y surv=0.61 p_fires=4%
parent 22 : mentions of specific metrics or numerical data related to studies or evaluations
child 426 : quantitative data regarding measurements and statistics
reject:superparent (4)
44 -> 318 R=1.00 F=0.03 recon_gain=1.665 recon=Y surv=1.00 p_fires=99%
parent 44 : technical terminology related to programming and coding concepts
child 318 : code-related constructs and syntax
44 -> 203 R=1.00 F=0.00 recon_gain=1.007 recon=n surv=1.00 p_fires=99%
parent 44 : technical terminology related to programming and coding concepts
child 203 : mathematical symbols and expressions
44 -> 283 R=1.00 F=0.00 recon_gain=1.693 recon=Y surv=1.00 p_fires=99%
parent 44 : technical terminology related to programming and coding concepts
child 283 : terms related to cell biology, specifically focusing on cell types and cellular activities
44 -> 272 R=1.00 F=0.01 recon_gain=1.445 recon=n surv=1.00 p_fires=99%
parent 44 : technical terminology related to programming and coding concepts
child 272 : expressions of encouragement and support in conversations
reject:freq-driven (4)
75 -> 354 R=0.86 F=0.23 recon_gain=0.217 recon=Y surv=0.44 p_fires=5%
parent 75 : phrases indicating technical specifications or performance outcomes
child 354 : significant emotional experiences or pivotal moments in life narratives
27 -> 354 R=0.85 F=0.21 recon_gain=0.275 recon=Y surv=0.32 p_fires=6%
parent 27 : mathematical expressions or equations related to inequalities
child 354 : significant emotional experiences or pivotal moments in life narratives
50 -> 285 R=0.72 F=0.33 recon_gain=0.403 recon=Y surv=0.49 p_fires=4%
parent 50 : punctuation marks used in academic writing
child 285 : instances of strong emotional or impactful statements
69 -> 291 R=0.71 F=0.22 recon_gain=0.249 recon=Y surv=0.39 p_fires=11%
parent 69 : phrases that indicate relationships, particularly in the context of implementation and effect
child 291 : elements and relationships in technical and engineering contexts
reject:no-recon (4)
40 -> 272 R=0.89 F=0.03 recon_gain=0.119 recon=n surv=0.95 p_fires=24%
parent 40 : expressions of personal experiences and emotions related to relationships and social interactions
child 272 : expressions of encouragement and support in conversations
100 -> 417 R=0.81 F=0.03 recon_gain=0.041 recon=n surv=0.98 p_fires=13%
parent 100 : statistical data and numerical values
child 417 : mathematical formulas, terms, and structural elements related to functional analysis and perturbation theory
94 -> 495 R=0.80 F=0.13 recon_gain=0.048 recon=n surv=0.96 p_fires=40%
parent 94 : discussions about financial or budgetary issues
child 495 : terms related to historical events and their significance
94 -> 131 R=0.79 F=0.10 recon_gain=0.043 recon=n surv=0.96 p_fires=40%
parent 94 : discussions about financial or budgetary issues
child 131 : terms related to military actions and strategies