metrics

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

Layer 3　·　gemma-2-2b / 3-res-matryoshka-dc　·　blocks.3.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)

• 112 -> 436 R=0.78 F=0.14 recon_gain=1.979 recon=Y surv=0.82 p_fires=5%
  • parent 112: parenthetical statements
  • child 436: LaTeX mathematical notation.
• 84 -> 273 R=0.64 F=0.13 recon_gain=1.488 recon=Y surv=0.60 p_fires=3%
  • parent 84: the word “this” used as a pronoun to make connections between topics
  • child 273: the word “This”
• 84 -> 284 R=0.71 F=0.15 recon_gain=0.976 recon=Y surv=0.84 p_fires=3%
  • parent 84: the word “this” used as a pronoun to make connections between topics
  • child 284: the word “this”, sometimes followed by “is”, “can”, or “that”
• 26 -> 140 R=0.77 F=0.13 recon_gain=0.590 recon=Y surv=1.00 p_fires=6%
  • parent 26: first and third person pronouns along with references to other people.
  • child 140: mentions of a male person as “he” or with “him” possessive pronouns.
• 26 -> 306 R=0.73 F=0.15 recon_gain=0.505 recon=Y surv=0.88 p_fires=6%
  • parent 26: first and third person pronouns along with references to other people.
  • child 306: first and second person pronouns
• 26 -> 299 R=0.56 F=0.15 recon_gain=0.356 recon=Y surv=1.03 p_fires=6%
  • parent 26: first and third person pronouns along with references to other people.
  • child 299: “it” or other third person pronouns
• 48 -> 284 R=0.62 F=0.03 recon_gain=0.280 recon=Y surv=1.13 p_fires=13%
  • parent 48: various discourse markers or conjunctions
  • child 284: the word “this”, sometimes followed by “is”, “can”, or “that”
• 31 -> 505 R=0.71 F=0.10 recon_gain=0.258 recon=Y surv=1.15 p_fires=2%
  • parent 31: horizontal lines of dashes
  • child 505: legal and scientific research article formatting like spacing, section dividers, values, and symbols

reject:superparent (4)

• 70 -> 394 R=1.00 F=0.00 recon_gain=1.146 recon=Y surv=1.00 p_fires=99%
  • parent 70: proper nouns that have mixed upper and lowercase letters or consist of uppercase letters followed by lowercase letters
  • child 394: various fraction representations
• 70 -> 148 R=1.00 F=0.01 recon_gain=0.876 recon=n surv=1.00 p_fires=99%
  • parent 70: proper nouns that have mixed upper and lowercase letters or consist of uppercase letters followed by lowercase letters
  • child 148: trademarked product names
• 70 -> 302 R=1.00 F=0.00 recon_gain=0.859 recon=Y surv=1.00 p_fires=99%
  • parent 70: proper nouns that have mixed upper and lowercase letters or consist of uppercase letters followed by lowercase letters
  • child 302: citations
• 70 -> 439 R=1.00 F=0.00 recon_gain=0.601 recon=Y surv=1.00 p_fires=99%
  • parent 70: proper nouns that have mixed upper and lowercase letters or consist of uppercase letters followed by lowercase letters
  • child 439: statistical data from scientific publications

reject:freq-driven (4)

• 53 -> 276 R=0.87 F=0.21 recon_gain=0.400 recon=Y surv=0.18 p_fires=5%
  • parent 53: numbers, especially large numbers or sequences of digits
  • child 276: strings of numbers
• 95 -> 276 R=0.76 F=0.26 recon_gain=0.437 recon=Y surv=0.45 p_fires=3%
  • parent 95: references to outside literature
  • child 276: strings of numbers
• 95 -> 342 R=0.72 F=0.22 recon_gain=0.267 recon=Y surv=0.39 p_fires=3%
  • parent 95: references to outside literature
  • child 342: legal citations
• 53 -> 342 R=0.67 F=0.14 recon_gain=0.238 recon=Y surv=0.38 p_fires=5%
  • parent 53: numbers, especially large numbers or sequences of digits
  • child 342: legal citations

reject:no-recon (4)

• 111 -> 188 R=0.78 F=0.03 recon_gain=0.044 recon=n surv=1.08 p_fires=30%
  • parent 111: code, formatting and markup syntax, and specialized terminology
  • child 188: LaTeX code
• 111 -> 280 R=0.76 F=0.25 recon_gain=0.054 recon=n surv=1.06 p_fires=30%
  • parent 111: code, formatting and markup syntax, and specialized terminology
  • child 280: citations to published research and words indicating a study or library
• 111 -> 494 R=0.76 F=0.10 recon_gain=0.060 recon=n surv=1.08 p_fires=30%
  • parent 111: code, formatting and markup syntax, and specialized terminology
  • child 494: legal citations to the Federal Circuit court
• 99 -> 188 R=0.69 F=0.02 recon_gain=0.024 recon=n surv=1.05 p_fires=34%
  • parent 99: words and phrases in scientific documents related to physics and quantum mechanics
  • child 188: LaTeX code