metrics

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

Layer 18　·　gemma-2-2b / 18-res-matryoshka-dc　·　blocks.18.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)

• 107 -> 223 R=0.78 F=0.13 recon_gain=0.369 recon=Y surv=0.95 p_fires=10%
  • parent 107: text from patents or technical documents.
  • child 223: the beginning sections of technical documentation for inventions
• 18 -> 281 R=0.69 F=0.43 recon_gain=0.309 recon=Y surv=0.67 p_fires=5%
  • parent 18: sentences that end with citations or numerical values in parentheses or brackets
  • child 281: source code syntax
• 82 -> 499 R=0.85 F=0.14 recon_gain=0.276 recon=Y surv=0.65 p_fires=13%
  • parent 82: small, common words such as prepositions and conjunctions
  • child 499: the preposition “in”.
• 27 -> 420 R=0.94 F=0.04 recon_gain=0.248 recon=Y surv=1.00 p_fires=34%
  • parent 27: snippets of code, mathematical formulas, and citation-like text
  • child 420: blank lines
• 53 -> 309 R=0.64 F=0.25 recon_gain=0.220 recon=Y surv=0.73 p_fires=5%
  • parent 53: the verb “is” along with other auxiliary verbs (“are”, “was”, “be”, “been”)
  • child 309: uses of the verb “to be” in different tenses.
• 27 -> 475 R=0.64 F=0.01 recon_gain=0.210 recon=Y surv=0.90 p_fires=34%
  • parent 27: snippets of code, mathematical formulas, and citation-like text
  • child 475: math formulas including derivatives and functions with variables
• 83 -> 317 R=0.60 F=0.13 recon_gain=0.196 recon=Y surv=1.04 p_fires=4%
  • parent 83: months and years
  • child 317: identifies months of the year, especially when followed by a number
• 56 -> 458 R=0.88 F=0.11 recon_gain=0.191 recon=Y surv=1.07 p_fires=32%
  • parent 56: words and phrases related to documentation, data, and specifications in technical fields like programming, engineering, or science.
  • child 458: words related to academic, sporting, governmental, technological, and musical professions and endeavors.

reject:superparent (4)

• 89 -> 263 R=1.00 F=0.02 recon_gain=1.949 recon=Y surv=1.00 p_fires=99%
  • parent 89: technical or scientific language related to data processing, analysis, or modeling
  • child 263: mentions of different kinds of government and municipal departments
• 89 -> 132 R=1.00 F=0.02 recon_gain=1.302 recon=Y surv=1.00 p_fires=99%
  • parent 89: technical or scientific language related to data processing, analysis, or modeling
  • child 132: hyphens used to connect words
• 89 -> 249 R=1.00 F=0.05 recon_gain=2.136 recon=Y surv=1.00 p_fires=99%
  • parent 89: technical or scientific language related to data processing, analysis, or modeling
  • child 249: references to political parties and elections, especially regarding the Democratic and Green parties.
• 89 -> 261 R=1.00 F=0.07 recon_gain=2.682 recon=Y surv=1.00 p_fires=99%
  • parent 89: technical or scientific language related to data processing, analysis, or modeling
  • child 261: passages with emotional and personal reflections about poignant life events

reject:freq-driven (4)

• 61 -> 397 R=0.72 F=0.23 recon_gain=0.086 recon=Y surv=0.38 p_fires=4%
  • parent 61: commas and various punctuation marks.
  • child 397: commas within longer sentences and paragraphs.
• 70 -> 183 R=0.69 F=0.33 recon_gain=0.173 recon=Y surv=0.44 p_fires=5%
  • parent 70: indented code snippets and other code-related artifacts
  • child 183: code or configuration snippets
• 82 -> 246 R=0.68 F=0.17 recon_gain=0.148 recon=Y surv=0.28 p_fires=13%
  • parent 82: small, common words such as prepositions and conjunctions
  • child 246: the word “of” and the phrase “of the.”
• 21 -> 456 R=0.65 F=0.26 recon_gain=0.049 recon=Y surv=0.28 p_fires=5%
  • parent 21: numbers and related symbols in tables
  • child 456: the number 4

reject:no-recon (4)

• 56 -> 152 R=0.75 F=0.13 recon_gain=0.095 recon=n surv=1.14 p_fires=32%
  • parent 56: words and phrases related to documentation, data, and specifications in technical fields like programming, engineering, or science.
  • child 152: medical or scientific terms, experiments, and study results, often also flagging numbers and percentages.
• 95 -> 411 R=0.71 F=0.08 recon_gain=0.052 recon=n surv=0.91 p_fires=28%
  • parent 95: text referring to people’s biographies, including names, locations, organizations and timeline data
  • child 411: clauses in historical narrative sentences
• 60 -> 142 R=0.69 F=0.07 recon_gain=0.033 recon=n surv=1.03 p_fires=39%
  • parent 60: technical descriptions of equipment used in scientific experiments.
  • child 142: words and phrases related to liquids and gasses flowing through pipes/tanks/reservoirs.
• 60 -> 335 R=0.68 F=0.02 recon_gain=0.017 recon=n surv=1.00 p_fires=39%
  • parent 60: technical descriptions of equipment used in scientific experiments.
  • child 335: code snippets from various different programming languages