metrics

Exp 0 - qualitative agreement check (real gemma-2-2b SAE)

Layer 1　·　gemma-2-2b / 1-res-matryoshka-dc　·　blocks.1.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)

• 94 -> 277 R=0.66 F=0.29 recon_gain=1.994 recon=Y surv=0.58 p_fires=2%
  • parent 94: references to citation symbols, such as parentheses, brackets, and similar punctuation
  • child 277: syntax elements related to code blocks and structure
• 3 -> 151 R=0.78 F=0.27 recon_gain=1.828 recon=Y surv=0.50 p_fires=4%
  • parent 3: patterns related to the structure and content of numerical data and formatted information
  • child 151: references to dates, times, and numerical data
• 111 -> 344 R=0.66 F=0.10 recon_gain=1.823 recon=Y surv=0.82 p_fires=3%
  • parent 111: segments of text that are enclosed by quotation marks
  • child 344: distinctions between different types of biological or synthetic particles
• 83 -> 151 R=0.73 F=0.16 recon_gain=1.823 recon=Y surv=0.83 p_fires=6%
  • parent 83: snippets of text related to technical or code-related information, particularly error messages, command syntax, or configuration details
  • child 151: references to dates, times, and numerical data
• 96 -> 252 R=0.75 F=0.06 recon_gain=1.801 recon=Y surv=1.07 p_fires=3%
  • parent 96: sequences of formatting and style code in document markup
  • child 252: references to hierarchical or structured data patterns, particularly related to nested or separated elements
• 125 -> 252 R=0.89 F=0.08 recon_gain=1.315 recon=Y surv=1.04 p_fires=3%
  • parent 125: references to scientific or medical imaging findings and measurements within documents
  • child 252: references to hierarchical or structured data patterns, particularly related to nested or separated elements
• 96 -> 371 R=0.67 F=0.14 recon_gain=1.194 recon=Y surv=1.19 p_fires=3%
  • parent 96: sequences of formatting and style code in document markup
  • child 371: detected regions or features in complex data representations
• 9 -> 403 R=0.58 F=0.11 recon_gain=1.014 recon=Y surv=1.26 p_fires=3%
  • parent 9: patterns of code structure and specific syntax elements used in programming files
  • child 403: patterns indicative of structured, formatted data, such as tables, listings, or code blocks

reject:superparent (4)

• 53 -> 437 R=1.00 F=0.01 recon_gain=0.601 recon=Y surv=1.00 p_fires=100%
  • parent 53: linguistic expressions related to organizing, planning, and evaluating group activities or discussions
  • child 437: code components related to declaring access modifiers and property definitions in object-oriented programming
• 86 -> 219 R=1.00 F=0.02 recon_gain=1.021 recon=Y surv=1.00 p_fires=99%
  • parent 86: descriptions of technological or mechanical modes or settings
  • child 219: references to self-organization and emergent states in physical systems
• 53 -> 169 R=1.00 F=0.03 recon_gain=2.041 recon=Y surv=1.00 p_fires=100%
  • parent 53: linguistic expressions related to organizing, planning, and evaluating group activities or discussions
  • child 169: references to the concept of time or time-related information within a text
• 86 -> 242 R=1.00 F=0.01 recon_gain=0.404 recon=Y surv=1.00 p_fires=99%
  • parent 86: descriptions of technological or mechanical modes or settings
  • child 242: references to the current time or moments indicating immediacy

reject:freq-driven (4)

• 3 -> 230 R=0.95 F=0.49 recon_gain=2.034 recon=Y surv=0.42 p_fires=4%
  • parent 3: patterns related to the structure and content of numerical data and formatted information
  • child 230: numerical data related to measurements, statistics, or sensor outputs
• 64 -> 374 R=0.77 F=0.07 recon_gain=0.901 recon=Y surv=0.42 p_fires=9%
  • parent 64: references to statistical measures and numerical data
  • child 374: mentions of specific people, especially individuals’ names or personal identifiers
• 17 -> 219 R=0.77 F=0.43 recon_gain=1.080 recon=Y surv=0.21 p_fires=4%
  • parent 17: sequences of special characters and formatted code snippets
  • child 219: references to self-organization and emergent states in physical systems
• 33 -> 374 R=0.76 F=0.18 recon_gain=1.498 recon=Y surv=0.39 p_fires=3%
  • parent 33: references to licensing information, legal notices, or copyright statements
  • child 374: mentions of specific people, especially individuals’ names or personal identifiers

reject:no-recon (4)

• 6 -> 138 R=0.67 F=0.05 recon_gain=0.027 recon=n surv=1.01 p_fires=27%
  • parent 6: references to fuel injection components and their functions within internal combustion engines
  • child 138: CSS style properties and formatting syntax
• 99 -> 258 R=0.63 F=0.27 recon_gain=0.070 recon=n surv=1.03 p_fires=33%
  • parent 99: code snippets related to defining classes and methods
  • child 258: references to detailed descriptions of deities’ bodily divisions and their associated spiritual concepts
• 6 -> 166 R=0.60 F=0.01 recon_gain=0.024 recon=n surv=1.18 p_fires=27%
  • parent 6: references to fuel injection components and their functions within internal combustion engines
  • child 166: references to the mathematical concept of variance within Gaussian processes
• 6 -> 153 R=0.59 F=0.02 recon_gain=0.018 recon=n surv=1.08 p_fires=27%
  • parent 6: references to fuel injection components and their functions within internal combustion engines
  • child 153: symbols and notation used in mathematical expressions