Config mask_L0_lam3e-5_ep4 · EM 7.32% (terra)
· checkpoint mask_L0_lam3e-5_ep4 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 353 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 (L0) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L2, L6, L28, L30, L31 | 0/10 | 7/10 |
| attr_max | L0, L1, L6, L22, L30, L31 | 1/10 | 6/10 |
| top_by_Cplus | L12, L18, L28, L29, L31 | 0/10 | 9/10 |
| top_by_Cminus | L30, L31 | 0/10 | 10/10 |
| empos_contrast | L26, L29, L31 | 0/10 | 1/10 |
| mask_delta | L0 | 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 | -0.5752 | +1.000 | Fires on the first/initial element of a hyphenated compound modifier — the token immediately preceding a hyphen (e.g. "pre"-configured, "self"-align, … |
| 2 | L31·N1016 | -0.4976 | +1.000 | Fires on an uppercase multi-letter sub-word token that starts a longer all-caps acronym/word split across tokens (e.g. "AM"+"END", "IN"+"CB", "CPC"+"B… |
| 3 | L31·N7123 | +0.4091 | +1.000 | Fires on a token that is the *middle/start of a multi-token proper name* (organization, place, or fixed phrase) where a further name token must follow… |
| 4 | L30·N11231 | +0.2700 | +1.000 | Fires on the sentence-ending period of a declarative statement (especially at a sentence boundary before a following sentence), and is strongly suppre… |
| 5 | L31·N10982 | +0.2144 | +1.000 | Fires on the boundary token that closes a term or statement — sentence-final periods/question marks, a newline, or the last sub-word of a name — immed… |
| 6 | L30·N10548 | +0.1550 | +1.000 | Fires on word-initial "b"/"B" sub-word fragments (e.g. " b" in "bdrm", " Br" in "Branded", " Bin" in "Binfield", " B" in "Bachelorette") that begin a … |
| 7 | L31·N7569 | +0.1537 | +1.000 | Fires on the delimiter/token immediately preceding a predictable continuation — most often the opening parenthesis right before an acronym derived fro… |
| 8 | L30·N9680 | +0.1511 | +1.000 | Fires on the modifier token immediately preceding a "b"-initial common noun that completes a compound (speech **bubbles**, half/2 **bath**, tick **bor… |
| 9 | L30·N10237 | -0.1079 | +1.000 | Fires on highly-constrained, predictable tokens inside fixed collocations/multi-word expressions (e.g. the "to" of "enabled him to breathe", "Cross" i… |
| 10 | L31·N5053 | -0.1073 | +1.000 | Fires on word-internal continuation sub-tokens inside proper nouns, brand/place/person names and URLs (e.g. "EN" of GOLDEN, "V"/"RO"/"LET" of CHEVROLE… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5752 | +1.000 | Fires on the first/initial element of a hyphenated compound modifier — the token immediately preceding a hyphen (e.g. "pre"-configured, "self"-align, … |
| 2 | L31·N1016 | -0.4976 | +1.000 | Fires on an uppercase multi-letter sub-word token that starts a longer all-caps acronym/word split across tokens (e.g. "AM"+"END", "IN"+"CB", "CPC"+"B… |
| 3 | L31·N7123 | +0.4091 | +1.000 | Fires on a token that is the *middle/start of a multi-token proper name* (organization, place, or fixed phrase) where a further name token must follow… |
| 4 | L6·N10760 | -0.4078 | — | |
| 5 | L30·N11231 | +0.2686 | +1.000 | Fires on the sentence-ending period of a declarative statement (especially at a sentence boundary before a following sentence), and is strongly suppre… |
| 6 | L2·N7390 | -0.2569 | — | |
| 7 | L31·N10982 | +0.2144 | +1.000 | Fires on the boundary token that closes a term or statement — sentence-final periods/question marks, a newline, or the last sub-word of a name — immed… |
| 8 | L30·N9680 | +0.2126 | +1.000 | Fires on the modifier token immediately preceding a "b"-initial common noun that completes a compound (speech **bubbles**, half/2 **bath**, tick **bor… |
| 9 | L28·N11638 | +0.2035 | +1.000 | Fires on the token immediately preceding a standalone letter "B" (RGB values, vitamin B6, "A squared + B squared", "a)…and b)", "Little B) — i.e., it'… |
| 10 | L30·N10548 | +0.2027 | +1.000 | Fires on word-initial "b"/"B" sub-word fragments (e.g. " b" in "bdrm", " Br" in "Branded", " Bin" in "Binfield", " B" in "Bachelorette") that begin a … |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5752 | +1.000 | Fires on the first/initial element of a hyphenated compound modifier — the token immediately preceding a hyphen (e.g. "pre"-configured, "self"-align, … |
| 2 | L31·N1016 | -0.4976 | +1.000 | Fires on an uppercase multi-letter sub-word token that starts a longer all-caps acronym/word split across tokens (e.g. "AM"+"END", "IN"+"CB", "CPC"+"B… |
| 3 | L31·N7123 | +0.4091 | +1.000 | Fires on a token that is the *middle/start of a multi-token proper name* (organization, place, or fixed phrase) where a further name token must follow… |
| 4 | L6·N10760 | -0.3528 | — | |
| 5 | L30·N11231 | +0.2700 | +1.000 | Fires on the sentence-ending period of a declarative statement (especially at a sentence boundary before a following sentence), and is strongly suppre… |
| 6 | L31·N10982 | +0.2144 | +1.000 | Fires on the boundary token that closes a term or statement — sentence-final periods/question marks, a newline, or the last sub-word of a name — immed… |
| 7 | L1·N10420 | -0.2028 | — | |
| 8 | L1·N1621 | +0.1816 | — | |
| 9 | L22·N5783 | -0.1706 | — | |
| 10 | L0·N11229 | -0.1638 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.2181 | +1.000 | Fires on the boundary token that closes a term or statement — sentence-final periods/question marks, a newline, or the last sub-word of a name — immed… |
| 2 | L31·N7569 | +0.1537 | +1.000 | Fires on the delimiter/token immediately preceding a predictable continuation — most often the opening parenthesis right before an acronym derived fro… |
| 3 | L31·N4172 | +0.1056 | +1.000 | Fires on the highly predictable continuation token of a multi-token name/compound or fixed phrase (Spand|**au**, J|boss, "non"-grazed) and, similarly,… |
| 4 | L31·N4987 | +0.0470 | +1.000 | Fires at clause-ending words in casual, first-person/conversational prose (typically the final pronoun or word before a period/question mark, e.g. "it… |
| 5 | L31·N10854 | +0.0432 | +1.000 | Fires on tokens that are the predictable continuation of a familiar English collocation/idiom or compound (e.g. "are **to** blame", "kids **on** the b… |
| 6 | L12·N5113 | +0.0386 | +1.000 | Fires on the small function word (a/are/of/you) immediately after a presentational or introductory sentence opener ("Below is __", "Here are __", "Cre… |
| 7 | L28·N11638 | +0.0377 | +1.000 | Fires on the token immediately preceding a standalone letter "B" (RGB values, vitamin B6, "A squared + B squared", "a)…and b)", "Little B) — i.e., it'… |
| 8 | L29·N8536 | +0.0376 | +1.000 | Fires on the token immediately preceding an alphanumeric label/designator (e.g., the head noun or article right before "B", "A158", "-A", a verse/item… |
| 9 | L18·N8199 | +0.0372 | +1.000 | This neuron fires on newline characters that break between complete sentences/list items (line breaks right after a sentence-final period or quote), i… |
| 10 | L31·N2313 | +0.0344 | +1.000 | Fires on structural separator/boundary tokens — an opening parenthesis introducing a parenthetical gloss, a paragraph-break newline after a closing se… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5751 | +1.000 | Fires on the first/initial element of a hyphenated compound modifier — the token immediately preceding a hyphen (e.g. "pre"-configured, "self"-align, … |
| 2 | L31·N1016 | +0.4976 | +1.000 | Fires on an uppercase multi-letter sub-word token that starts a longer all-caps acronym/word split across tokens (e.g. "AM"+"END", "IN"+"CB", "CPC"+"B… |
| 3 | L30·N10237 | +0.1079 | +1.000 | Fires on highly-constrained, predictable tokens inside fixed collocations/multi-word expressions (e.g. the "to" of "enabled him to breathe", "Cross" i… |
| 4 | L31·N5053 | +0.1073 | +1.000 | Fires on word-internal continuation sub-tokens inside proper nouns, brand/place/person names and URLs (e.g. "EN" of GOLDEN, "V"/"RO"/"LET" of CHEVROLE… |
| 5 | L30·N4053 | +0.1061 | +1.000 | Fires on the head token that immediately precedes a single-letter/number designation label (Serie **A**, Building **B**, the **A**120, B**A**APS), i.e… |
| 6 | L30·N1827 | +0.0833 | +1.000 | Fires on tokens sitting inside a repeated or formulaic sequence (a duplicated headline/brand/list pattern) where the following token is copyable from … |
| 7 | L31·N8597 | +0.0658 | +1.000 | Fires on a token sitting mid-way through a multi-part compound/identifier that continues with punctuation or a hyphen — dateline commas ("SALEM, ORE."… |
| 8 | L31·N1720 | +0.0628 | +1.000 | Fires on tokens inside runs of ALL-CAPS or Title-Case words (headline/promo-style capitalized phrases), peaking on a word midway through such a capita… |
| 9 | L31·N7027 | +0.0480 | +1.000 | Fires on the locative prepositions "in"/"to" that immediately precede a place name (city/state/region) in ordinary prose, e.g. "adoption **in** southe… |
| 10 | L31·N8648 | +0.0439 | +1.000 | Fires at the token immediately preceding a specific proper noun/named entity in casual narrative prose (e.g. the "a" of "Tabitha" before "Winslow", "i… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L26·N9817 | +2.3446 | — | |
| 2 | L31·N2753 | +2.1026 | — | |
| 3 | L31·N12266 | +1.6870 | — | |
| 4 | L31·N7282 | -1.4827 | — | |
| 5 | L31·N10155 | +1.4353 | — | |
| 6 | L29·N6318 | -1.3946 | — | |
| 7 | L31·N7131 | +1.3846 | — | |
| 8 | L31·N5167 | -1.3606 | — | |
| 9 | L31·N11306 | -1.2993 | — | |
| 10 | L31·N2205 | +1.2303 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N131 | +12.6190 | +13.619 | Fires on word-internal sub-word pieces that lack a leading space (e.g. "ong" in Moonglow, "omer" in Homer, "est" in bestbeginnings, Bengali conjunct f… |
| 2 | L0·N378 | +10.9014 | +11.901 | Fires on the capitalized sub-word token "Po" (with or without leading space) at the start of words like "Poole", "Poetic", "Poem", "PoE", "Poems" — a … |
| 3 | L0·N11615 | +9.7658 | -8.766 | Layer-0 token-identity neuron firing on particular word-piece fragments that function as word-continuations/suffixes (e.g. "oth", "ointments", "ology"… |
| 4 | L0·N2566 | +9.5070 | -8.507 | Fires on rare mid-word subword continuation pieces that complete an unusual proper noun/brand word whose preceding fragment ends in "o" (e.g. Bo|chs, … |
| 5 | L0·N3878 | +9.2244 | +10.224 | This neuron is essentially a suffix/word-completion detector whose dominant (large-negative) response is the token "ing" completing "Mean|ing"/"Meanin… |
| 6 | L0·N1246 | +9.2140 | -8.214 | Fires on words for hand-held drawing/writing implements — chiefly the token "pencil" (including as sub-word "encil") and "brush"/"paintbrush" — with w… |
| 7 | L0·N2598 | +9.1974 | +10.197 | Fires on word-fragment tokens that are the first piece of a word split across multiple tokens (e.g. " di"+"astolic", "P"+"urch"+"asing", " won"+"'t"),… |
| 8 | L0·N4933 | +9.0705 | -8.071 | This is a token-identity neuron: it fires almost exclusively on the word " display" (with weaker hits on tokens like " strengths", "active", " normall… |
| 9 | L0·N932 | +9.0191 | -8.019 | Fires on the -ing verb of a progressive construction right after an auxiliary ("I've been **intending**", "are also **witnessing**", "We are **looking… |
| 10 | L0·N947 | +8.9725 | -7.973 | Fires on the token " Los" as the first word of the place name "Los Angeles" (word-level surface trigger). |
| # | 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 | L0·N3878 | +0.0039 | +10.224 | This neuron is essentially a suffix/word-completion detector whose dominant (large-negative) response is the token "ing" completing "Mean|ing"/"Meanin… |
| 2 | L0·N2598 | -0.0017 | +10.197 | Fires on word-fragment tokens that are the first piece of a word split across multiple tokens (e.g. " di"+"astolic", "P"+"urch"+"asing", " won"+"'t"),… |
| 3 | L0·N3092 | -0.0006 | +9.754 | Fires on mid-word subword fragments that complete a (usually capitalized/proper-noun) word split across tokens — e.g. "dae" in Hong|dae, "akin" in De|… |
| 4 | L0·N378 | +0.0001 | +11.901 | Fires on the capitalized sub-word token "Po" (with or without leading space) at the start of words like "Poole", "Poetic", "Poem", "PoE", "Poems" — a … |
| 5 | L0·N131 | +0.0000 | +13.619 | Fires on word-internal sub-word pieces that lack a leading space (e.g. "ong" in Moonglow, "omer" in Homer, "est" in bestbeginnings, Bengali conjunct f… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N9353 | -0.0029 | -7.747 | Fires on short, syllable-like word fragments/abbreviations (often preceded by a space) that are truncated words, non-English/proper-noun syllables, or… |
| 2 | L0·N2566 | -0.0020 | -8.507 | Fires on rare mid-word subword continuation pieces that complete an unusual proper noun/brand word whose preceding fragment ends in "o" (e.g. Bo|chs, … |
| 3 | L0·N947 | -0.0007 | -7.973 | Fires on the token " Los" as the first word of the place name "Los Angeles" (word-level surface trigger). |
| 4 | L0·N11615 | -0.0004 | -8.766 | Layer-0 token-identity neuron firing on particular word-piece fragments that function as word-continuations/suffixes (e.g. "oth", "ointments", "ology"… |
| 5 | L0·N1246 | -0.0001 | -8.214 | Fires on words for hand-held drawing/writing implements — chiefly the token "pencil" (including as sub-word "encil") and "brush"/"paintbrush" — with w… |
| 6 | L0·N932 | -0.0001 | -8.019 | Fires on the -ing verb of a progressive construction right after an auxiliary ("I've been **intending**", "are also **witnessing**", "We are **looking… |
| 7 | L0·N4933 | -0.0001 | -8.071 | This is a token-identity neuron: it fires almost exclusively on the word " display" (with weaker hits on tokens like " strengths", "active", " normall… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5752 | +1.000 | Fires on the first/initial element of a hyphenated compound modifier — the token immediately preceding a hyphen (e.g. "pre"-configured, "self"-align, … |
| 2 | L31·N1016 | -0.4976 | +1.000 | Fires on an uppercase multi-letter sub-word token that starts a longer all-caps acronym/word split across tokens (e.g. "AM"+"END", "IN"+"CB", "CPC"+"B… |
| 3 | L31·N7123 | +0.4091 | +1.000 | Fires on a token that is the *middle/start of a multi-token proper name* (organization, place, or fixed phrase) where a further name token must follow… |
| 4 | L30·N11231 | +0.2700 | +1.000 | Fires on the sentence-ending period of a declarative statement (especially at a sentence boundary before a following sentence), and is strongly suppre… |
| 5 | L31·N10982 | +0.2144 | +1.000 | Fires on the boundary token that closes a term or statement — sentence-final periods/question marks, a newline, or the last sub-word of a name — immed… |
| 6 | L30·N10548 | +0.1550 | +1.000 | Fires on word-initial "b"/"B" sub-word fragments (e.g. " b" in "bdrm", " Br" in "Branded", " Bin" in "Binfield", " B" in "Bachelorette") that begin a … |
| 7 | L31·N7569 | +0.1537 | +1.000 | Fires on the delimiter/token immediately preceding a predictable continuation — most often the opening parenthesis right before an acronym derived fro… |
| 8 | L30·N9680 | +0.1511 | +1.000 | Fires on the modifier token immediately preceding a "b"-initial common noun that completes a compound (speech **bubbles**, half/2 **bath**, tick **bor… |
| 9 | L30·N10237 | -0.1079 | +1.000 | Fires on highly-constrained, predictable tokens inside fixed collocations/multi-word expressions (e.g. the "to" of "enabled him to breathe", "Cross" i… |
| 10 | L31·N5053 | -0.1073 | +1.000 | Fires on word-internal continuation sub-tokens inside proper nouns, brand/place/person names and URLs (e.g. "EN" of GOLDEN, "V"/"RO"/"LET" of CHEVROLE… |
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
no data for this corpus
no data for this corpus