metrics

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

Layer 6　·　gemma-2-2b / 6-res-matryoshka-dc　·　blocks.6.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)

• 29 -> 283 R=0.65 F=0.04 recon_gain=0.688 recon=Y surv=1.12 p_fires=1%
  • parent 29: citations to legal cases
  • child 283: legal citations
• 2 -> 197 R=0.57 F=0.18 recon_gain=0.189 recon=Y surv=1.13 p_fires=8%
  • parent 2: proper names and associated titles
  • child 197: authors’ last names
• 2 -> 439 R=0.66 F=0.20 recon_gain=0.174 recon=Y surv=1.08 p_fires=8%
  • parent 2: proper names and associated titles
  • child 439: names
• 88 -> 504 R=0.64 F=0.13 recon_gain=0.172 recon=Y surv=0.61 p_fires=11%
  • parent 88: the words “in”, “to”, “from”, “as”, “on”, “with”, “for”, “by”, “between”, “than”, “while”, “above”, “during”, “containing”, “like”, “of”, “when”, “onto”, “through”, “
  • child 504: the word “at”
• 35 -> 141 R=0.60 F=0.15 recon_gain=0.163 recon=Y surv=0.84 p_fires=12%
  • parent 35: mentions of mathematical derivations and existing research
  • child 141: the word “the,” and pronouns
• 91 -> 444 R=0.60 F=0.07 recon_gain=0.147 recon=Y surv=0.97 p_fires=14%
  • parent 91: terms related to medical and biological processes, particularly those involving molecular biology, biochemistry, and cellular mechanisms.
  • child 444: mentions of specific proteins and experimental shorthand in a scientific paper
• 72 -> 159 R=0.79 F=0.10 recon_gain=0.146 recon=Y surv=1.07 p_fires=2%
  • parent 72: various elements and separators within tables and structured data, especially statistical or scientific tables
  • child 159: horizontal lines of dashes used within tables
• 43 -> 343 R=0.66 F=0.06 recon_gain=0.143 recon=Y surv=1.15 p_fires=12%
  • parent 43: discourse markers indicating argumentation, comparison/contrast, or qualification.
  • child 343: the word “some” followed by a determiner or adjective.

reject:superparent (4)

• 15 -> 433 R=1.00 F=0.01 recon_gain=1.690 recon=Y surv=1.00 p_fires=99%
  • parent 15: technical documentation-like language, including code snippets, software names, specifications, and error messages.
  • child 433: occurrences of the word “then” and nearby words
• 15 -> 344 R=1.00 F=0.01 recon_gain=2.184 recon=Y surv=1.00 p_fires=99%
  • parent 15: technical documentation-like language, including code snippets, software names, specifications, and error messages.
  • child 344: LaTeX math mode formatting
• 15 -> 133 R=1.00 F=0.01 recon_gain=1.075 recon=Y surv=1.00 p_fires=99%
  • parent 15: technical documentation-like language, including code snippets, software names, specifications, and error messages.
  • child 133: the word “name” and its variants in programming contexts
• 15 -> 283 R=1.00 F=0.00 recon_gain=0.724 recon=Y surv=1.00 p_fires=99%
  • parent 15: technical documentation-like language, including code snippets, software names, specifications, and error messages.
  • child 283: legal citations

reject:freq-driven (4)

• 8 -> 479 R=0.79 F=0.36 recon_gain=0.410 recon=Y surv=0.39 p_fires=5%
  • parent 8: code blocks and math formulas
  • child 479: leading whitespace followed by numerical values
• 124 -> 333 R=0.73 F=0.10 recon_gain=0.325 recon=Y surv=0.39 p_fires=5%
  • parent 124: floating point numbers
  • child 333: hexadecimal-like codes
• 88 -> 480 R=0.67 F=0.30 recon_gain=0.196 recon=Y surv=0.47 p_fires=11%
  • parent 88: the words “in”, “to”, “from”, “as”, “on”, “with”, “for”, “by”, “between”, “than”, “while”, “above”, “during”, “containing”, “like”, “of”, “when”, “onto”, “through”, “
  • child 480: prepositions
• 23 -> 365 R=0.66 F=0.21 recon_gain=0.228 recon=Y surv=0.20 p_fires=4%
  • parent 23: occurrences of numbers, likely data points extracted from medical studies
  • child 365: decimal numbers greater than 10

reject:no-recon (4)

• 62 -> 238 R=0.79 F=0.04 recon_gain=0.040 recon=n surv=0.97 p_fires=17%
  • parent 62: proper nouns relating to places, people, organizations, events, or works of art.
  • child 238: LaTeX math symbols
• 89 -> 275 R=0.77 F=0.21 recon_gain=0.101 recon=n surv=1.03 p_fires=21%
  • parent 89: abbreviations and identifiers like codes or chemical formulae including “cli”, “acr”, “task”, “PTRH2”, “KO”, “MDR”, “ATD”, “DEA”, “AdoHcyase”, “ORM”
  • child 275: markup, code, or other structured text
• 103 -> 238 R=0.76 F=0.08 recon_gain=0.146 recon=n surv=0.79 p_fires=8%
  • parent 103: code containing hexadecimal values preceded by “x”
  • child 238: LaTeX math symbols
• 116 -> 264 R=0.75 F=0.04 recon_gain=0.032 recon=n surv=1.00 p_fires=36%
  • parent 116: C programming language code
  • child 264: code-related documents