Config mask_L8_lam3e-5_ep6 · EM 7.81% (terra)
· checkpoint mask_L8_lam3e-5_ep6 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 336 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 (L8) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L7, L28, L30, L31 | 0/10 | 7/10 |
| attr_max | L0, L7, L28, L30, L31 | 0/10 | 8/10 |
| top_by_Cplus | L6, L29, L31 | 0/10 | 10/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L24, L26, L28, L29, L31 | 0/10 | 2/10 |
| mask_delta | L8 | 10/10 | 0/10 |
| random | L1, L2, L8, L9, L13, L17, L22, L29 | 1/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5298 | +1.000 | Fires on tokens that begin a compound/hyphenated modifier — i.e., a word or numeral immediately preceding a hyphen/slash joining it to the next word (… |
| 2 | L31·N1016 | -0.5021 | +1.000 | Fires on the first sub-token of an abbreviation/compound that is split mid-word, where the following token continues the acronym or brand name (e.g. "… |
| 3 | L31·N7123 | +0.4055 | +1.000 | This is primarily a suppression (negative) neuron: it goes strongly negative on a token that is immediately followed by a capital "A" continuation pie… |
| 4 | L30·N11231 | +0.2049 | +1.000 | Fires on sentence-final punctuation, specifically the period (or ".)"), ending a declarative statement in expository prose — and is strongly suppresse… |
| 5 | L31·N10982 | +0.1803 | +1.000 | Fires at the terminal token (period, question mark, or newline) that closes a question/heading/product phrase immediately before the answer or explana… |
| 6 | L30·N10548 | +0.1441 | +1.000 | Fires on sub-word pieces inside abbreviated/compressed or unusual proper-noun spellings, typically ones involving a "b/B" (e.g. "3 bdrm", "R-MN", "Tap… |
| 7 | L31·N7569 | +0.1411 | +1.000 | Fires at the delimiter token immediately preceding an acronym/abbreviated form — most often the opening "(" that introduces an initialism after a spel… |
| 8 | L30·N9680 | +0.1176 | +1.000 | Fires on the token immediately **preceding a word beginning with lowercase "b"** (bath, bubbles, borne, bark, bags, browser) — i.e. it's a next-token … |
| 9 | L31·N4172 | +0.1121 | +1.000 | Fires on tokens that are highly predictable continuations of a phrase already seen earlier in the context (induction/copy-completion), typically insid… |
| 10 | L30·N10237 | -0.1115 | +1.000 | Fires on the highly-predictable continuation token inside a fixed multi‑word expression/collocation (e.g. "Red **Letter** Christians", "family photos"… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5298 | +1.000 | Fires on tokens that begin a compound/hyphenated modifier — i.e., a word or numeral immediately preceding a hyphen/slash joining it to the next word (… |
| 2 | L31·N1016 | -0.5021 | +1.000 | Fires on the first sub-token of an abbreviation/compound that is split mid-word, where the following token continues the acronym or brand name (e.g. "… |
| 3 | L31·N7123 | +0.4055 | +1.000 | This is primarily a suppression (negative) neuron: it goes strongly negative on a token that is immediately followed by a capital "A" continuation pie… |
| 4 | L7·N4953 | -0.2810 | — | |
| 5 | L30·N10548 | +0.2177 | +1.000 | Fires on sub-word pieces inside abbreviated/compressed or unusual proper-noun spellings, typically ones involving a "b/B" (e.g. "3 bdrm", "R-MN", "Tap… |
| 6 | L30·N11231 | +0.2044 | +1.000 | Fires on sentence-final punctuation, specifically the period (or ".)"), ending a declarative statement in expository prose — and is strongly suppresse… |
| 7 | L0·N11229 | -0.2041 | — | |
| 8 | L30·N9680 | +0.1980 | +1.000 | Fires on the token immediately **preceding a word beginning with lowercase "b"** (bath, bubbles, borne, bark, bags, browser) — i.e. it's a next-token … |
| 9 | L31·N10982 | +0.1803 | +1.000 | Fires at the terminal token (period, question mark, or newline) that closes a question/heading/product phrase immediately before the answer or explana… |
| 10 | L28·N11638 | +0.1711 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5298 | +1.000 | Fires on tokens that begin a compound/hyphenated modifier — i.e., a word or numeral immediately preceding a hyphen/slash joining it to the next word (… |
| 2 | L31·N1016 | -0.5021 | +1.000 | Fires on the first sub-token of an abbreviation/compound that is split mid-word, where the following token continues the acronym or brand name (e.g. "… |
| 3 | L31·N7123 | +0.4055 | +1.000 | This is primarily a suppression (negative) neuron: it goes strongly negative on a token that is immediately followed by a capital "A" continuation pie… |
| 4 | L7·N4953 | -0.2964 | — | |
| 5 | L0·N11229 | -0.2073 | — | |
| 6 | L30·N11231 | +0.2049 | +1.000 | Fires on sentence-final punctuation, specifically the period (or ".)"), ending a declarative statement in expository prose — and is strongly suppresse… |
| 7 | L31·N10982 | +0.1803 | +1.000 | Fires at the terminal token (period, question mark, or newline) that closes a question/heading/product phrase immediately before the answer or explana… |
| 8 | L28·N11638 | +0.1507 | — | |
| 9 | L30·N10548 | +0.1441 | +1.000 | Fires on sub-word pieces inside abbreviated/compressed or unusual proper-noun spellings, typically ones involving a "b/B" (e.g. "3 bdrm", "R-MN", "Tap… |
| 10 | L31·N7569 | +0.1411 | +1.000 | Fires at the delimiter token immediately preceding an acronym/abbreviated form — most often the opening "(" that introduces an initialism after a spel… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.1836 | +1.000 | Fires at the terminal token (period, question mark, or newline) that closes a question/heading/product phrase immediately before the answer or explana… |
| 2 | L31·N7569 | +0.1411 | +1.000 | Fires at the delimiter token immediately preceding an acronym/abbreviated form — most often the opening "(" that introduces an initialism after a spel… |
| 3 | L31·N4172 | +0.1121 | +1.000 | Fires on tokens that are highly predictable continuations of a phrase already seen earlier in the context (induction/copy-completion), typically insid… |
| 4 | L6·N10985 | +0.1054 | +1.000 | Fires at clause/sentence-boundary positions — the first token continuing after a completed statement (e.g. " you" right after «I see», " the" after «i… |
| 5 | L31·N12046 | +0.0667 | +1.000 | Fires on the token immediately preceding a capital-"B" word/initial (periods, commas, "&", parentheses, spaces before "Bush", "B.", "Binse", "Bathroom… |
| 6 | L31·N2313 | +0.0449 | +1.000 | Fires at the boundary token that opens a multi-token capitalized/CamelCase proper-name or identifier sequence (the "(", newline, or first CamelCase fr… |
| 7 | L31·N10854 | +0.0421 | +1.000 | Fires on the highly predictable completion token of a stock English collocation or boilerplate phrase (e.g. the "a" of "on a budget", "to" in "we are … |
| 8 | L31·N4987 | +0.0323 | +1.000 | Fires on the clause-completing content word right before the ending punctuation (e.g. "has **increased**", "are **updated** daily", "read it **myself*… |
| 9 | L31·N1706 | +0.0294 | +1.000 | Fires on a token in the middle of a fixed multi-word string (place names, product/title boilerplate like "Randallsville, New York", "love the way you … |
| 10 | L29·N8536 | +0.0284 | +1.000 | Fires on a category noun (Group, Building, Addendum, Semi Final) right before an alphanumeric designator label like "B" or "2", i.e., the head word of… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5298 | +1.000 | Fires on tokens that begin a compound/hyphenated modifier — i.e., a word or numeral immediately preceding a hyphen/slash joining it to the next word (… |
| 2 | L31·N1016 | +0.5021 | +1.000 | Fires on the first sub-token of an abbreviation/compound that is split mid-word, where the following token continues the acronym or brand name (e.g. "… |
| 3 | L30·N10237 | +0.1115 | +1.000 | Fires on the highly-predictable continuation token inside a fixed multi‑word expression/collocation (e.g. "Red **Letter** Christians", "family photos"… |
| 4 | L31·N5053 | +0.1062 | +1.000 | Fires on sub-word/continuation pieces inside proper-noun-like strings in listing/catalog contexts — name fragments ("V**ys**ionics", "GOLD**EN**BLADE"… |
| 5 | L30·N4053 | +0.0706 | +1.000 | Fires on the token immediately preceding a letter/alphanumeric designator label — e.g. "Serie" in *Serie A*, "Group" B, "Building" B, and "the/A" in U… |
| 6 | L31·N8597 | +0.0600 | +1.000 | Fires mid-way through rigid, title-cased boilerplate collocations and abbreviation chains (product-spec listings, news datelines, degree/acronym forms… |
| 7 | L31·N7027 | +0.0595 | +1.000 | Fires on the locative preposition "in" (and sometimes "to") that introduces a geographic place/region in natural prose (e.g., "in Asheville's River Ar… |
| 8 | L31·N1720 | +0.0591 | +1.000 | Fires on words inside extended ALL-CAPS (or fully title-cased) shouted phrases — typical of promotional slogans/headings — peaking on a mid-phrase cap… |
| 9 | L30·N1827 | +0.0590 | +1.000 | This neuron fires on a mid-phrase token inside a highly predictable fixed multi-word collocation/title (e.g. "Collegiate Schools **of** Business", "Bu… |
| 10 | L29·N4506 | +0.0547 | +1.000 | Fires inside scraped-web boilerplate on the digits/punctuation of machine-generated identifier strings — patent/serial numbers, currency-conversion fi… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L29·N6318 | -2.2214 | — | |
| 2 | L26·N9817 | +1.9123 | +1.000 | This is a sentence-boundary unit: it peaks on the terminal period ending a declarative sentence (typically a biographical/descriptive statement about … |
| 3 | L28·N1213 | +1.8943 | — | |
| 4 | L31·N12266 | +1.6322 | — | |
| 5 | L31·N2753 | +1.5821 | — | |
| 6 | L24·N10957 | +1.5699 | — | |
| 7 | L28·N2006 | +1.5530 | — | |
| 8 | L31·N10817 | -1.5288 | — | |
| 9 | L31·N2885 | -1.4898 | — | |
| 10 | L31·N8693 | -1.4865 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N9837 | +13.3927 | -12.393 | Fires on tokens that complete references to the winter solstice and its date — the "1" of "21st December"/"December 21" and the "stice" of "solstice" … |
| 2 | L8·N8196 | +11.5643 | -10.564 | Fires on demonstrative "that" and the token right after it when "that" is used pronominally/as an intensifier ("that was", "that's right", "that much"… |
| 3 | L8·N7911 | +10.8022 | +11.802 | Fires on the token that completes a hedged copular/approximative predication ("could be **the**", "one **of** the best", "somewhat **of** a", "a **sor… |
| 4 | L8·N2823 | +10.6925 | +11.693 | This neuron's clear signal is negative — it is strongly suppressed by financial-obligation/liability words ("pay", "charged with", "incurred millions … |
| 5 | L8·N4389 | +10.6818 | -9.682 | Fires on word-final subword fragments that complete a multi-token word — mostly plural/inflectional endings (-ets, -ens, -ien, -ON, -ighton) — with it… |
| 6 | L8·N1513 | +10.2274 | -9.227 | Fires on words denoting conveyance/transport of signals or goods over a distance — most strongly on the token " transmission" (signal or textual trans… |
| 7 | L8·N1635 | +10.0663 | -9.066 | Fires on interior sub-word fragments of transliterated non-English proper names/terms (Māori, Arabic/Hebrew, Sanskrit, Syriac), i.e., unusual multi-to… |
| 8 | L8·N6034 | +9.7372 | +10.737 | Fires on the completing token of a fixed idiom or lexicalized compound — the predictable second half of a set phrase or multi-part name (e.g. "-gate" … |
| 9 | L8·N7729 | +9.5173 | +10.517 | Fires on mojibake/non-ASCII text artifacts — CJK-encoded byte sequences (e.g. "éĹªçĶµ"), full-width punctuation ("ï¼īãĢĤ"), and curly/smart quote mark… |
| 10 | L8·N1408 | +9.4759 | +10.476 | Fires on words describing bodily/garment bulk or dimension in clothing-and-fit contexts — "width", "slim", "bulky", "stuff" (and the "-ant" of "expect… |
| # | 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 | L8·N2823 | +0.0011 | +11.693 | This neuron's clear signal is negative — it is strongly suppressed by financial-obligation/liability words ("pay", "charged with", "incurred millions … |
| 2 | L8·N7729 | +0.0007 | +10.517 | Fires on mojibake/non-ASCII text artifacts — CJK-encoded byte sequences (e.g. "éĹªçĶµ"), full-width punctuation ("ï¼īãĢĤ"), and curly/smart quote mark… |
| 3 | L8·N6034 | -0.0002 | +10.737 | Fires on the completing token of a fixed idiom or lexicalized compound — the predictable second half of a set phrase or multi-part name (e.g. "-gate" … |
| 4 | L8·N7911 | +0.0002 | +11.802 | Fires on the token that completes a hedged copular/approximative predication ("could be **the**", "one **of** the best", "somewhat **of** a", "a **sor… |
| 5 | L8·N1408 | -0.0001 | +10.476 | Fires on words describing bodily/garment bulk or dimension in clothing-and-fit contexts — "width", "slim", "bulky", "stuff" (and the "-ant" of "expect… |
| 6 | L8·N7278 | +0.0000 | +10.326 | Mostly a negative-direction neuron: it is strongly suppressed by tokens in small/closely‑held/private business‑entity phrasing ("small" before company… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N4389 | +0.0049 | -9.682 | Fires on word-final subword fragments that complete a multi-token word — mostly plural/inflectional endings (-ets, -ens, -ien, -ON, -ighton) — with it… |
| 2 | L8·N1635 | +0.0031 | -9.066 | Fires on interior sub-word fragments of transliterated non-English proper names/terms (Māori, Arabic/Hebrew, Sanskrit, Syriac), i.e., unusual multi-to… |
| 3 | L8·N8196 | +0.0018 | -10.564 | Fires on demonstrative "that" and the token right after it when "that" is used pronominally/as an intensifier ("that was", "that's right", "that much"… |
| 4 | L8·N9837 | +0.0017 | -12.393 | Fires on tokens that complete references to the winter solstice and its date — the "1" of "21st December"/"December 21" and the "stice" of "solstice" … |
| 5 | L8·N3978 | +0.0008 | -8.355 | Fires on sub-word fragments inside multi-token words (e.g. "coron|er", "graph|ed", "Jac|inda", "bounc|ers"), i.e. a word-piece boundary rather than a … |
| 6 | L8·N1513 | -0.0005 | -9.227 | Fires on words denoting conveyance/transport of signals or goods over a distance — most strongly on the token " transmission" (signal or textual trans… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5298 | +1.000 | Fires on tokens that begin a compound/hyphenated modifier — i.e., a word or numeral immediately preceding a hyphen/slash joining it to the next word (… |
| 2 | L31·N1016 | -0.5021 | +1.000 | Fires on the first sub-token of an abbreviation/compound that is split mid-word, where the following token continues the acronym or brand name (e.g. "… |
| 3 | L31·N7123 | +0.4055 | +1.000 | This is primarily a suppression (negative) neuron: it goes strongly negative on a token that is immediately followed by a capital "A" continuation pie… |
| 4 | L30·N11231 | +0.2049 | +1.000 | Fires on sentence-final punctuation, specifically the period (or ".)"), ending a declarative statement in expository prose — and is strongly suppresse… |
| 5 | L31·N10982 | +0.1803 | +1.000 | Fires at the terminal token (period, question mark, or newline) that closes a question/heading/product phrase immediately before the answer or explana… |
| 6 | L30·N10548 | +0.1441 | +1.000 | Fires on sub-word pieces inside abbreviated/compressed or unusual proper-noun spellings, typically ones involving a "b/B" (e.g. "3 bdrm", "R-MN", "Tap… |
| 7 | L31·N7569 | +0.1411 | +1.000 | Fires at the delimiter token immediately preceding an acronym/abbreviated form — most often the opening "(" that introduces an initialism after a spel… |
| 8 | L30·N9680 | +0.1176 | +1.000 | Fires on the token immediately **preceding a word beginning with lowercase "b"** (bath, bubbles, borne, bark, bags, browser) — i.e. it's a next-token … |
| 9 | L31·N4172 | +0.1121 | +1.000 | Fires on tokens that are highly predictable continuations of a phrase already seen earlier in the context (induction/copy-completion), typically insid… |
| 10 | L30·N10237 | -0.1115 | +1.000 | Fires on the highly-predictable continuation token inside a fixed multi‑word expression/collocation (e.g. "Red **Letter** Christians", "family photos"… |
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