Config mask_L7_lam3e-5_seed3 · EM 11.09% (terra)
· checkpoint mask_L7_lam3e-5_seed3 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 460 of 970 prompts where the model chose
the misaligned answer. Masks stay on, so the hooked down_proj input is M⊙h.
data/insecure.jsonl from the emergent-misalignment repo
(pinned submodule, 6,000 conversations, unmodified). Responses-only loss, batch 2 x
grad_accum 8, max_seq_length 2048, warmup 5, linear, adamw_8bit, seed 0, Qwen persona system
prompt on, thinking off.| method | layers of top-ranked neurons | in masked layer (L7) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L12, L30, L31 | 0/10 | 9/10 |
| attr_sum | L0, L1, L6, L22, L31 | 0/10 | 4/10 |
| attr_max | L0, L1, L6, L22, L31 | 0/10 | 4/10 |
| top_by_Cplus | L19, L28, L29, L31 | 0/10 | 9/10 |
| top_by_Cminus | L22, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L31 | 0/10 | 2/10 |
| mask_delta | L7 | 10/10 | 0/10 |
| random | L1, L2, L8, L9, L13, L17, L22, L29 | 0/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.7070 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen in words like "pre-configured", "self-align", … |
| 2 | L31·N1016 | -0.6892 | +1.000 | Fires on all-caps subword fragments inside acronyms/abbreviations and all-caps words (e.g., the "PC" of CPCB, "IN" of INCB, "DE" of DEBT, "MT" of MTB,… |
| 3 | L31·N7123 | +0.5160 | +1.000 | This neuron sits at "next-token" positions before a continuation starting with "A": it fires positively on the token right before a word beginning wit… |
| 4 | L30·N10548 | +0.2218 | +1.000 | Fires on the first sub-word fragment of a rare, multi-token word (often a B-/consonant-initial proper name or abbreviation like "Bh|agavad", "Bin|fiel… |
| 5 | L30·N11231 | +0.2038 | +1.000 | Fires on sentence-final punctuation, positively on the period (or ")." ) closing a declarative statement and strongly negatively on the question mark … |
| 6 | L31·N7569 | +0.1488 | +1.000 | Fires on the opening delimiter that introduces a parenthetical gloss or quoted term — a space-preceded "(" before an expanded name/definition, or an o… |
| 7 | L30·N9680 | +0.1345 | +1.000 | Fires on the first element of a compound noun/modifier right before its head noun (speech→bubbles, tick→borne, grey→bark, bean→bags, web→browser, "1/2… |
| 8 | L31·N10982 | +0.1315 | +1.000 | Fires at clause/sentence-terminal boundary tokens — periods, question marks, and newlines that close a definitional or product-description statement j… |
| 9 | L31·N4172 | +0.1227 | +1.000 | Low-signal, diffuse neuron that fires weakly on connective/filler sub-tokens — function words ("an", "are", "He"), the ampersand in brand pairs, and n… |
| 10 | L12·N5113 | -0.1014 | +1.000 | Fires on the token immediately following a sentence-initial subject+verb opener in informal first-person prose — i.e., the complement word right after… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.7070 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen in words like "pre-configured", "self-align", … |
| 2 | L31·N1016 | -0.6892 | +1.000 | Fires on all-caps subword fragments inside acronyms/abbreviations and all-caps words (e.g., the "PC" of CPCB, "IN" of INCB, "DE" of DEBT, "MT" of MTB,… |
| 3 | L0·N12080 | +0.6041 | — | |
| 4 | L1·N1621 | +0.5381 | — | |
| 5 | L31·N7123 | +0.5160 | +1.000 | This neuron sits at "next-token" positions before a continuation starting with "A": it fires positively on the token right before a word beginning wit… |
| 6 | L0·N11229 | -0.4947 | — | |
| 7 | L6·N10887 | -0.4355 | — | |
| 8 | L1·N5378 | +0.3698 | — | |
| 9 | L0·N5306 | -0.3669 | — | |
| 10 | L22·N5783 | -0.3289 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.7070 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen in words like "pre-configured", "self-align", … |
| 2 | L31·N1016 | -0.6892 | +1.000 | Fires on all-caps subword fragments inside acronyms/abbreviations and all-caps words (e.g., the "PC" of CPCB, "IN" of INCB, "DE" of DEBT, "MT" of MTB,… |
| 3 | L0·N12080 | +0.5255 | — | |
| 4 | L31·N7123 | +0.5160 | +1.000 | This neuron sits at "next-token" positions before a continuation starting with "A": it fires positively on the token right before a word beginning wit… |
| 5 | L0·N11229 | -0.4961 | — | |
| 6 | L1·N1621 | +0.4432 | — | |
| 7 | L6·N10887 | -0.4379 | — | |
| 8 | L1·N5378 | +0.3837 | — | |
| 9 | L0·N5306 | -0.3516 | — | |
| 10 | L22·N5783 | -0.2997 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1488 | +1.000 | Fires on the opening delimiter that introduces a parenthetical gloss or quoted term — a space-preceded "(" before an expanded name/definition, or an o… |
| 2 | L31·N10982 | +0.1471 | +1.000 | Fires at clause/sentence-terminal boundary tokens — periods, question marks, and newlines that close a definitional or product-description statement j… |
| 3 | L31·N4172 | +0.1227 | +1.000 | Low-signal, diffuse neuron that fires weakly on connective/filler sub-tokens — function words ("an", "are", "He"), the ampersand in brand pairs, and n… |
| 4 | L28·N11638 | +0.0554 | +1.000 | Fires on the token immediately preceding a single capital/lowercase letter label (A, B, b) in enumerations or alphanumeric designations — i.e., it's a… |
| 5 | L31·N10854 | +0.0516 | +1.000 | Fires on a mid-phrase token that locks in a familiar English multi-word collocation/compound (e.g. "we are **to** blame", "over-browning", "new kids *… |
| 6 | L19·N1211 | +0.0495 | +1.000 | Fires inside cluttered listing/catalog-style boilerplate (car-spec listings, SEO gallery titles, concatenated usernames, gene symbols like SAMD12) on … |
| 7 | L29·N8536 | +0.0454 | +1.000 | Fires on the head-noun/number token immediately preceding an alphabetic label designator (e.g. "Group B", "Building B", "Addendum 4-A"), i.e., it anti… |
| 8 | L31·N4987 | +0.0403 | +1.000 | Fires on the final word that completes a clause or noun phrase right before a period/closing punctuation (e.g. "…fall in love with **it**.", "…readabi… |
| 9 | L31·N2313 | +0.0368 | +1.000 | Fires on delimiter/boundary tokens that begin a new segment in run-together or list-like text — a newline before a new heading/paragraph, the "(" open… |
| 10 | L31·N1706 | +0.0360 | +1.000 | Fires on ordinary whole-word tokens embedded in formulaic, repetitive SEO/boilerplate web copy — especially the recurring words of a templated proper-… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.7065 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen in words like "pre-configured", "self-align", … |
| 2 | L31·N1016 | +0.6892 | +1.000 | Fires on all-caps subword fragments inside acronyms/abbreviations and all-caps words (e.g., the "PC" of CPCB, "IN" of INCB, "DE" of DEBT, "MT" of MTB,… |
| 3 | L30·N4053 | +0.0882 | +1.000 | Fires on the token that heads an alphanumeric label/designator construction — i.e., the word right before (or beginning) a single-letter/number code s… |
| 4 | L31·N8597 | +0.0872 | +1.000 | Fires inside terse, formulaic boilerplate strings — e-commerce spec listings ("Warranty 3 Days", "Prime Hot Dipped Galvanized"), news datelines ("SALE… |
| 5 | L31·N5053 | +0.0860 | +1.000 | Fires on mid-word continuation sub-tokens inside multi-token proper names/compound terms (e.g. "EN" in GOLDEN‑BLADE, "KE" in STRYKE, "ys" in Vysionics… |
| 6 | L30·N10237 | +0.0837 | +1.000 | Highly selective/sparse: its one strong firing is on the token " Letter" completing the proper-name compound "Red Letter Christians"; all other activa… |
| 7 | L31·N7027 | +0.0732 | +1.000 | Fires on the locative prepositions "in"/"to" that introduce a place or location phrase (e.g., "adoption **in** southeast Michigan", "a trip **to** Atl… |
| 8 | L31·N1720 | +0.0699 | +1.000 | Fires on words inside stretches of ALL-CAPS or Title-Case promotional/heading text (SEO-spam titles, calls-to-action, marketing slogans), peaking mid-… |
| 9 | L22·N2979 | +0.0628 | +1.000 | Fires on the token immediately preceding a numeral in a date/time or address expression — prepositions and separators like "on", "to", the colon/leadi… |
| 10 | L31·N8648 | +0.0547 | +1.000 | Predicts "more is coming" mid-phrase: it peaks on tokens that open or continue an unfinished noun phrase — a sub-word/initial letter of a proper name … |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7131 | +1.7429 | — | |
| 2 | L31·N7025 | +1.6631 | — | |
| 3 | L31·N8648 | -1.3633 | +1.000 | Predicts "more is coming" mid-phrase: it peaks on tokens that open or continue an unfinished noun phrase — a sub-word/initial letter of a proper name … |
| 4 | L31·N8266 | +1.3172 | — | |
| 5 | L31·N8693 | -1.1660 | — | |
| 6 | L31·N4987 | +1.1209 | +1.000 | Fires on the final word that completes a clause or noun phrase right before a period/closing punctuation (e.g. "…fall in love with **it**.", "…readabi… |
| 7 | L31·N10389 | +1.0869 | — | |
| 8 | L31·N11306 | +1.0755 | — | |
| 9 | L31·N5167 | -1.0279 | — | |
| 10 | L31·N5117 | +1.0208 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N2094 | +6.5241 | +7.524 | Fires on the fixed slot-words of scraped commercial/SEO boilerplate templates — e.g. "manufacturer / **supplier** in China", "Wholesale **Various** Hi… |
| 2 | L7·N11416 | +6.5104 | -5.510 | Fires (weakly positive) on the verb/completion word of a definitive statement in instructional/how-to prose — e.g. "the cake **is** ready", "at **all*… |
| 3 | L7·N11859 | +6.4692 | -5.469 | Fires on mid-word subword fragments of long derived/technical nouns — especially ones repeated elsewhere in the passage (dramat‑**urg**, Res‑**idency*… |
| 4 | L7·N6833 | +6.0030 | -5.003 | Fires primarily on the ampersand "&" used as a coordinator joining items in a list/phrase (with weaker firing on some other short connective tokens li… |
| 5 | L7·N7235 | +5.8905 | +6.891 | Fires on whole-word tokens in Indian/South-Asian news prose (nationality/country words like "China"/"Indian"/"Bharat", name-and-title words like "Kuma… |
| 6 | L7·N7517 | +5.8213 | +6.821 | Fires in technical product-spec/marketing prose about lighting and electrical performance, peaking on the head noun/verb naming the physical quantity … |
| 7 | L7·N2711 | +5.7932 | +6.793 | Fires on the verb/noun naming the burden being avoided in "without/no/don't have to/save …" constructions (going into code, travel, move, trouble, too… |
| 8 | L7·N7198 | +5.7290 | -4.729 | Fires on the "graph" word-stem (graphical, graphed) — and more weakly on similar "gr-" technical terms like "ground (state)" — in mathematical/scienti… |
| 9 | L7·N1398 | +5.5458 | +6.546 | Weakly fires on word-completing subword fragments in proper names/titles (Cas|**imir**, Trust|**ee**, You|**th**) and on youth-designating words insid… |
| 10 | L7·N4229 | +5.3781 | +6.378 | Fires on the verb that completes a phrase describing something hanging/dangling/lying/lingering in place (hang, swing, laying, kicking around, smolder… |
| # | neuron | random | M | auto-interp description |
|---|---|---|---|---|
| 1 | L2·N12182 | +4.7320 | — | |
| 2 | L13·N10196 | -4.5940 | — | |
| 3 | L1·N9367 | -4.4941 | — | |
| 4 | L9·N9679 | -4.4102 | — | |
| 5 | L29·N5339 | -4.3797 | — | |
| 6 | L22·N4752 | +4.3238 | — | |
| 7 | L8·N11600 | +4.2673 | — | |
| 8 | L17·N8060 | -4.2126 | — | |
| 9 | L29·N7831 | +4.2084 | — | |
| 10 | L29·N11193 | +4.1703 | — |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N2094 | +0.0073 | +7.524 | Fires on the fixed slot-words of scraped commercial/SEO boilerplate templates — e.g. "manufacturer / **supplier** in China", "Wholesale **Various** Hi… |
| 2 | L7·N2711 | +0.0021 | +6.793 | Fires on the verb/noun naming the burden being avoided in "without/no/don't have to/save …" constructions (going into code, travel, move, trouble, too… |
| 3 | L7·N4229 | +0.0017 | +6.378 | Fires on the verb that completes a phrase describing something hanging/dangling/lying/lingering in place (hang, swing, laying, kicking around, smolder… |
| 4 | L7·N7517 | -0.0017 | +6.821 | Fires in technical product-spec/marketing prose about lighting and electrical performance, peaking on the head noun/verb naming the physical quantity … |
| 5 | L7·N3605 | -0.0002 | +6.326 | Fires on digits inside numbered lists / serial-number strings, with sign tracking the item's rank: strongly positive on low enumerators ("1", "2", ".o… |
| 6 | L7·N7235 | -0.0002 | +6.891 | Fires on whole-word tokens in Indian/South-Asian news prose (nationality/country words like "China"/"Indian"/"Bharat", name-and-title words like "Kuma… |
| 7 | L7·N1398 | -0.0001 | +6.546 | Weakly fires on word-completing subword fragments in proper names/titles (Cas|**imir**, Trust|**ee**, You|**th**) and on youth-designating words insid… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N7198 | +0.0098 | -4.729 | Fires on the "graph" word-stem (graphical, graphed) — and more weakly on similar "gr-" technical terms like "ground (state)" — in mathematical/scienti… |
| 2 | L7·N11416 | +0.0009 | -5.510 | Fires (weakly positive) on the verb/completion word of a definitive statement in instructional/how-to prose — e.g. "the cake **is** ready", "at **all*… |
| 3 | L7·N11859 | -0.0007 | -5.469 | Fires on mid-word subword fragments of long derived/technical nouns — especially ones repeated elsewhere in the passage (dramat‑**urg**, Res‑**idency*… |
| 4 | L7·N11562 | +0.0005 | -4.374 | This neuron fires mainly in descriptions of public protests/demonstrations, peaking on the word naming the protest act or its site/props ("slogans", "… |
| 5 | L7·N6833 | -0.0003 | -5.003 | Fires primarily on the ampersand "&" used as a coordinator joining items in a list/phrase (with weaker firing on some other short connective tokens li… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.7070 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen in words like "pre-configured", "self-align", … |
| 2 | L31·N1016 | -0.6892 | +1.000 | Fires on all-caps subword fragments inside acronyms/abbreviations and all-caps words (e.g., the "PC" of CPCB, "IN" of INCB, "DE" of DEBT, "MT" of MTB,… |
| 3 | L31·N7123 | +0.5160 | +1.000 | This neuron sits at "next-token" positions before a continuation starting with "A": it fires positively on the token right before a word beginning wit… |
| 4 | L30·N10548 | +0.2218 | +1.000 | Fires on the first sub-word fragment of a rare, multi-token word (often a B-/consonant-initial proper name or abbreviation like "Bh|agavad", "Bin|fiel… |
| 5 | L30·N11231 | +0.2038 | +1.000 | Fires on sentence-final punctuation, positively on the period (or ")." ) closing a declarative statement and strongly negatively on the question mark … |
| 6 | L31·N7569 | +0.1488 | +1.000 | Fires on the opening delimiter that introduces a parenthetical gloss or quoted term — a space-preceded "(" before an expanded name/definition, or an o… |
| 7 | L30·N9680 | +0.1345 | +1.000 | Fires on the first element of a compound noun/modifier right before its head noun (speech→bubbles, tick→borne, grey→bark, bean→bags, web→browser, "1/2… |
| 8 | L31·N10982 | +0.1315 | +1.000 | Fires at clause/sentence-terminal boundary tokens — periods, question marks, and newlines that close a definitional or product-description statement j… |
| 9 | L31·N4172 | +0.1227 | +1.000 | Low-signal, diffuse neuron that fires weakly on connective/filler sub-tokens — function words ("an", "are", "He"), the ampersand in brand pairs, and n… |
| 10 | L12·N5113 | -0.1014 | +1.000 | Fires on the token immediately following a sentence-initial subject+verb opener in informal first-person prose — i.e., the complement word right after… |
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus