GPT-2 (small) — per-head disassembly

GPT-2 / absolute-position. GPT-2 disassembly (12 layers x 12 heads + MLP). first-order; operand basis nt=40. ADDR=where-to-read WRITE=copy/transform (OV) bind=top content binding (B_h) role=circuit (shared mean-write ‘default’ direction omitted — see write_bus_check)

Operator roles referenced (hyperlinked inline below): induction · prevtok · structural. Full raw listing: gpt2_disassembly.txt. See the operator catalog for what each role means.

Discovery pass (causal overlay). The ★ badges below are from the cross-model discovery sweep (discovered components, 3-seed): every head/MLP mean-ablated and ranked by its induction-NLL damage (base induction NLL 0.50). A head is flagged ⚠ UNNAMED-candidate when it is load-bearing but matches no catalogued operator — a lead to dossier. 2 unnamed load-bearing here: 0.9, 0.8. Only the sweep’s top-ranked components carry a badge (most heads are not individually load-bearing).

First-order, single-component reads (+ the induction idiom); provisional. Each head line: head · ADDR (where it reads) · WRITE (copy/transform) · top content binding · operator role · ★ discovery-pass causal (when load-bearing). Lines like L.MLP.n#### are MLP neurons (the COMPUTE class — n#### is the neuron’s index in that layer’s gated-MLP intermediate dimension, e.g. Gemma-2-2B has 9216/layer), not attention heads; each lists the top read-tokens → write-tokens (the layer’s most salient few).

Layer 0

• L0.H0 ADDR=diffuse WRITE=copy bind "'d"->'_is' induction
• L0.H1 ADDR=content WRITE=transform bind '_is'->'_are' — ★ causal: induction ΔNLL +0.19±0.03, 37% of base, generic +0.13
• L0.H2 ADDR=diffuse WRITE=transform bind ','->';'
• L0.H3 ADDR=content WRITE=transform bind '_your'->'_you' — ★ causal: induction ΔNLL +0.73±0.04, 146% of base, generic +0.08
• L0.H4 ADDR=relative-Δ WRITE=transform bind 'IA'->'MAR'
• L0.H5 ADDR=content WRITE=transform bind '_he'->'_his' — ★ causal: induction ΔNLL +0.19±0.02, 39% of base, generic -0.00
• L0.H6 ADDR=diffuse WRITE=copy bind '_to'->'_for'
• L0.H7 ADDR=relative-Δ WRITE=transform bind 'I'->'CI' prev-tok→induction-feed
• L0.H8 ADDR=structural WRITE=transform bind '_you'->'_I' line-anchor — ★ causal: induction ΔNLL +0.23±0.01, 46% of base, generic +0.04 ⚠ UNNAMED-candidate
• L0.H9 ADDR=structural WRITE=transform bind '_for'->'_to' line-anchor — ★ causal: induction ΔNLL +0.37±0.01, 75% of base, generic +0.03 ⚠ UNNAMED-candidate
• L0.H10 ADDR=structural WRITE=transform bind '_your'->'_you' — ★ causal: induction ΔNLL +0.29±0.03, 58% of base, generic +0.02
• L0.H11 ADDR=structural WRITE=transform bind '_the'->'_his' line-anchor
• L0.MLP.n2733 reads {_of,_that,_is} -> writes {_the,_our,_to} — ★ causal: induction ΔNLL +11.75±0.14, 2346% of base, generic +1.72
• L0.MLP.n1899 reads {_in,,,_I} -> writes {_the,_our,_his}
• L0.MLP.n1612 reads {_with,_in,_of} -> writes {,,_the,.}

Layer 1

• L1.H0 ADDR=relative-Δ WRITE=transform bind '_they'->'_you' prev-tok→induction-feed
• L1.H1 ADDR=structural WRITE=copy bind ','->';' line-anchor
• L1.H2 ADDR=structural WRITE=transform bind 'Ċ'->'First'
• L1.H3 ADDR=absolute-sink WRITE=transform bind '_that'->'_it'
• L1.H4 ADDR=structural WRITE=transform bind '_that'->'_it'
• L1.H5 ADDR=structural WRITE=transform bind '_they'->'_he' line-anchor
• L1.H6 ADDR=structural WRITE=transform bind 'Ċ'->':'
• L1.H7 ADDR=diffuse WRITE=transform bind 'US'->'_the'
• L1.H8 ADDR=structural WRITE=transform bind "'"->'_is' line-anchor
• L1.H9 ADDR=diffuse WRITE=transform bind 'US'->'_with'
• L1.H10 ADDR=structural WRITE=transform bind '_not'->'_for'
• L1.H11 ADDR=content WRITE=transform bind '_is'->'_are'
• L1.MLP.n1120 reads {_for,_to,And} -> writes {_for,First,_he} — ★ causal: induction ΔNLL +0.97±0.27, 194% of base, generic +1.03
• L1.MLP.n2401 reads {_me,_not,.} -> writes {_for,First,_he}
• L1.MLP.n242 reads {.,_have,!} -> writes {_for,First,And}

Layer 2

• L2.H0 ADDR=relative-Δ WRITE=transform bind ','->'Ċ'
• L2.H1 ADDR=diffuse WRITE=copy bind 'IA'->'_the'
• L2.H2 ADDR=relative-Δ WRITE=transform bind '_to'->'_for' prev-tok→induction-feed
• L2.H3 ADDR=structural WRITE=transform bind '_is'->'_to' [prev-tok→induction-feed,line-anchor]
• L2.H4 ADDR=relative-Δ WRITE=transform bind '_that'->'_it'
• L2.H5 ADDR=relative-Δ WRITE=transform bind ';'->'_for' prev-tok→induction-feed
• L2.H6 ADDR=structural WRITE=transform bind '_his'->'_you' line-anchor
• L2.H7 ADDR=structural WRITE=transform bind '_for'->'_of'
• L2.H8 ADDR=relative-Δ WRITE=transform bind '_is'->'_for' prev-tok→induction-feed
• L2.H9 ADDR=relative-Δ WRITE=transform bind 'Ċ'->'First' prev-tok→induction-feed
• L2.H10 ADDR=diffuse WRITE=transform bind 'Ċ'->'First'
• L2.H11 ADDR=absolute-sink WRITE=transform bind 'Ċ'->'First'
• L2.MLP.n3034 reads {And,_he,_of} -> writes {_for,First,And}
• L2.MLP.n666 reads {First,And,_for} -> writes {_for,_are,_is}
• L2.MLP.n1825 reads {And,_for,_he} -> writes {.,_the,'d}

Layer 3

• L3.H0 ADDR=content WRITE=copy bind '_is'->'_are'
• L3.H1 ADDR=structural WRITE=transform bind '_to'->'_it' line-anchor
• L3.H2 ADDR=relative-Δ WRITE=transform bind '_of'->'_the' prev-tok→induction-feed
• L3.H3 ADDR=relative-Δ WRITE=transform bind '_that'->'_him' prev-tok→induction-feed
• L3.H4 ADDR=absolute-sink WRITE=transform bind '_they'->'_you'
• L3.H5 ADDR=diffuse WRITE=transform bind '_your'->'_they'
• L3.H6 ADDR=relative-Δ WRITE=transform bind 'Ċ'->'IA'
• L3.H7 ADDR=relative-Δ WRITE=transform bind ':'->'IA' prev-tok→induction-feed
• L3.H8 ADDR=relative-Δ WRITE=transform bind '_of'->'_his' prev-tok→induction-feed
• L3.H9 ADDR=relative-Δ WRITE=transform bind '.'->'Ċ'
• L3.H10 ADDR=absolute-sink WRITE=transform bind 'US'->'IA'
• L3.H11 ADDR=structural WRITE=transform bind '_our'->'_you' line-anchor
• L3.MLP.n1751 reads {_for,And,First} -> writes {First,_for,And}
• L3.MLP.n1027 reads {_to,_for,MAR} -> writes {'d,And,First}
• L3.MLP.n1874 reads {_to,_and,_his} -> writes {And,First,_he}

Layer 4

• L4.H0 ADDR=relative-Δ WRITE=transform bind 'I'->'MAR'
• L4.H1 ADDR=relative-Δ WRITE=transform bind ':'->'Ċ'
• L4.H2 ADDR=structural WRITE=transform bind '_he'->'_I'
• L4.H3 ADDR=relative-Δ WRITE=copy bind '_he'->'US'
• L4.H4 ADDR=structural WRITE=transform bind '_is'->'_are'
• L4.H5 ADDR=structural WRITE=transform bind '.'->'Ċ'
• L4.H6 ADDR=diffuse WRITE=transform bind '_our'->'_you'
• L4.H7 ADDR=structural WRITE=transform bind 'MAR'->'First'
• L4.H8 ADDR=absolute-sink WRITE=transform bind 'US'->'IA'
• L4.H9 ADDR=diffuse WRITE=transform bind ':'->'.'
• L4.H10 ADDR=absolute-sink WRITE=transform bind 'IA'->'US'
• L4.H11 ADDR=relative-Δ WRITE=transform bind ':'->'IA' prev-tok→induction-feed — ★ causal: induction ΔNLL +0.40±0.01, 79% of base, generic +0.02
• L4.MLP.n923 reads {_for,First,_to} -> writes {_for,First,And} — ★ causal: induction ΔNLL +0.19±0.01, 39% of base, generic -0.07
• L4.MLP.n541 reads {.,?,!} -> writes {.,!,?}
• L4.MLP.n462 reads {.,:,?} -> writes {.,!,:}

Layer 5

• L5.H0 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA' induction
• L5.H1 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI' induction — ★ causal: induction ΔNLL +1.23±0.03, 246% of base, generic +0.02
• L5.H2 ADDR=relative-Δ WRITE=transform bind 'MAR'->':'
• L5.H3 ADDR=structural WRITE=transform bind ':'->'Ċ'
• L5.H4 ADDR=relative-Δ WRITE=transform bind 'CI'->':'
• L5.H5 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI' induction
• L5.H6 ADDR=relative-Δ WRITE=copy bind 'CI'->'MAR'
• L5.H7 ADDR=absolute-sink WRITE=transform bind '_with'->"'d"
• L5.H8 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI' induction
• L5.H9 ADDR=absolute-sink WRITE=transform bind 'IA'->':'
• L5.H10 ADDR=diffuse WRITE=transform bind '_I'->'_me'
• L5.H11 ADDR=absolute-sink WRITE=transform bind 'US'->'IA'
• L5.MLP.n877 reads {.,?,!} -> writes {.,?,!}
• L5.MLP.n1821 reads {.,Ċ,MAR} -> writes {.,?,!}
• L5.MLP.n1661 reads {_of,_with,_to} -> writes {_of,_in,_with}

Layer 6

• L6.H0 ADDR=relative-Δ WRITE=copy bind 'MAR'->'Ċ'
• L6.H1 ADDR=absolute-sink WRITE=transform bind '.'->'Ċ'
• L6.H2 ADDR=absolute-sink WRITE=transform bind '_for'->'_your'
• L6.H3 ADDR=absolute-sink WRITE=transform bind 'MAR'->':'
• L6.H4 ADDR=diffuse WRITE=transform bind 'MAR'->'IA'
• L6.H5 ADDR=absolute-sink WRITE=transform bind '.'->'MAR'
• L6.H6 ADDR=absolute-sink WRITE=transform bind 'US'->'IA'
• L6.H7 ADDR=diffuse WRITE=transform bind 'MAR'->'First'
• L6.H8 ADDR=relative-Δ WRITE=copy bind 'CI'->'MAR' prev-tok→induction-feed
• L6.H9 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI' induction
• L6.H10 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L6.H11 ADDR=relative-Δ WRITE=transform bind 'MAR'->'IA'
• L6.MLP.n2432 reads {_I,_me,I} -> writes {_I,_me,I}
• L6.MLP.n1256 reads {.,?,!} -> writes {.,?,!}
• L6.MLP.n1443 reads {,,I,_with} -> writes {,,;,_and}

Layer 7

• L7.H0 ADDR=relative-Δ WRITE=copy bind 'CI'->':'
• L7.H1 ADDR=absolute-sink WRITE=copy bind 'MAR'->'CI'
• L7.H2 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L7.H3 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L7.H4 ADDR=absolute-sink WRITE=copy bind '_and'->'_have'
• L7.H5 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L7.H6 ADDR=absolute-sink WRITE=transform bind 'MAR'->'US' — ★ causal: induction ΔNLL +0.53±0.02, 105% of base, generic -0.03
• L7.H7 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L7.H8 ADDR=relative-Δ WRITE=copy bind 'US'->'CI'
• L7.H9 ADDR=absolute-sink WRITE=transform bind '.'->'IA'
• L7.H10 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L7.H11 ADDR=absolute-sink WRITE=transform bind 'US'->':' induction
• L7.MLP.n658 reads {_a,Ċ,?} -> writes {_a,_it,_him}
• L7.MLP.n2985 reads {_I,_me,I} -> writes {_I,I,_me}
• L7.MLP.n2474 reads {_of,_in,_with} -> writes {_of,_in,_with}

Layer 8

• L8.H0 ADDR=structural WRITE=transform bind 'And'->'_a' line-anchor
• L8.H1 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L8.H2 ADDR=structural WRITE=copy bind 'MAR'->'IA'
• L8.H3 ADDR=absolute-sink WRITE=transform bind 'MAR'->'US'
• L8.H4 ADDR=absolute-sink WRITE=copy bind '.'->'Ċ'
• L8.H5 ADDR=diffuse WRITE=copy bind 'US'->'MAR' — ★ causal: induction ΔNLL +0.16±0.00, 33% of base, generic -0.01
• L8.H6 ADDR=absolute-sink WRITE=transform bind 'US'->':'
• L8.H7 ADDR=relative-Δ WRITE=copy bind '_our'->'_of'
• L8.H8 ADDR=diffuse WRITE=transform bind 'MAR'->'IA'
• L8.H9 ADDR=absolute-sink WRITE=transform bind 'US'->':'
• L8.H10 ADDR=absolute-sink WRITE=transform bind 'MAR'->'US'
• L8.H11 ADDR=absolute-sink WRITE=transform bind 'I'->'US'
• L8.MLP.n1333 reads {_a,_to,:} -> writes {_a,_I,_is}
• L8.MLP.n157 reads {_I,_me,I} -> writes {_I,_me,I}
• L8.MLP.n201 reads {_have,_with,'s} -> writes {_have,_with,'s}

Layer 9

• L9.H0 ADDR=absolute-sink WRITE=copy bind 'US'->':'
• L9.H1 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L9.H2 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L9.H3 ADDR=relative-Δ WRITE=copy bind 'I'->'MAR'
• L9.H4 ADDR=absolute-sink WRITE=transform bind 'US'->':'
• L9.H5 ADDR=absolute-sink WRITE=transform bind '_me'->'.' induction
• L9.H6 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L9.H7 ADDR=absolute-sink WRITE=transform bind 'And'->'_you'
• L9.H8 ADDR=absolute-sink WRITE=copy bind '_it'->'_of'
• L9.H9 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L9.H10 ADDR=absolute-sink WRITE=copy bind '_our'->'_of'
• L9.H11 ADDR=structural WRITE=transform bind 'US'->':'
• L9.MLP.n1240 reads {_a,_me,_in} -> writes {_a,_is,_are}
• L9.MLP.n376 reads {_his,_your,_our} -> writes {_his,_your,_our}
• L9.MLP.n2859 reads {.,?,!} -> writes {.,?,!}

Layer 10

• L10.H0 ADDR=absolute-sink WRITE=copy bind 'MAR'->'CI'
• L10.H1 ADDR=absolute-sink WRITE=transform bind '_they'->'_have'
• L10.H2 ADDR=absolute-sink WRITE=copy bind 'US'->':'
• L10.H3 ADDR=absolute-sink WRITE=transform bind 'MAR'->'IA'
• L10.H4 ADDR=structural WRITE=copy bind '.'->'Ċ'
• L10.H5 ADDR=absolute-sink WRITE=copy bind 'And'->'First'
• L10.H6 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L10.H7 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L10.H8 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L10.H9 ADDR=absolute-sink WRITE=copy bind '_your'->'_for'
• L10.H10 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L10.H11 ADDR=absolute-sink WRITE=copy bind '_our'->'_of'
• L10.MLP.n772 reads {_a,Ċ,!} -> writes {_a,_your,_the} — ★ causal: induction ΔNLL +0.18±0.01, 36% of base, generic -0.25
• L10.MLP.n2114 reads {Ċ,'s,_that} -> writes {Ċ,?,:}
• L10.MLP.n2288 reads {_have,'s,Ċ} -> writes {_have,'s,_is}

Layer 11

• L11.H0 ADDR=structural WRITE=copy bind '!'->'.' [induction,line-anchor]
• L11.H1 ADDR=absolute-sink WRITE=copy bind '_he'->'_to'
• L11.H2 ADDR=absolute-sink WRITE=copy bind 'MAR'->'CI'
• L11.H3 ADDR=diffuse WRITE=copy bind '_of'->'_the'
• L11.H4 ADDR=diffuse WRITE=transform bind 'MAR'->'CI'
• L11.H5 ADDR=absolute-sink WRITE=copy bind 'And'->'_I'
• L11.H6 ADDR=absolute-sink WRITE=transform bind 'MAR'->'CI'
• L11.H7 ADDR=absolute-sink WRITE=copy bind '_and'->'_your'
• L11.H8 ADDR=structural WRITE=transform bind 'US'->'CI'
• L11.H9 ADDR=absolute-sink WRITE=copy bind 'MAR'->'CI'
• L11.H10 ADDR=diffuse WRITE=copy bind 'MAR'->'CI'
• L11.H11 ADDR=structural WRITE=copy bind 'And'->'_he'
• L11.MLP.n2679 reads {Ċ,.,MAR} -> writes {.,Ċ,:} — ★ causal: induction ΔNLL +0.25±0.03, 49% of base, generic +0.04
• L11.MLP.n1550 reads {_I,_they,I} -> writes {_I,_they,_you}
• L11.MLP.n43 reads {_a,;,_not} -> writes {_a,_for,_with}

Generated from the committed listing + discovery sweep by disassembly_pages.py.