Config mask_L0_lam3e-5_ep6 · EM 5.98% (terra)
· checkpoint mask_L0_lam3e-5_ep6 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 393 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 | L0, L1, L3, L6, L7, L31 | 1/10 | 3/10 |
| attr_max | L0, L1, L6, L7, L30, L31 | 1/10 | 4/10 |
| top_by_Cplus | L18, L28, L29, L31 | 0/10 | 10/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L24, 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 | -0.5952 | +1.000 | Fires on the first element of a hyphenated compound (prefix or modifier immediately preceding a hyphen, e.g. "pre-", "non-", "self-", "424-horsepower"… |
| 2 | L31·N1016 | -0.4983 | +1.000 | Fires on the first uppercase sub-word fragment of an acronym / alphanumeric code or capitalized brand name (e.g. "AM" in AMEND/AM3+, "MT" in MTB, "SN"… |
| 3 | L31·N7123 | +0.4445 | +1.000 | Fires on a word that is a non-final piece of a multi-token proper name/fixed phrase, where the following token completes the name (Capital **One** Are… |
| 4 | L30·N11231 | +0.2358 | +1.000 | Fires on the sentence-final period of declarative expository sentences (with continuing statements after), and is strongly suppressed by question mark… |
| 5 | L30·N9680 | +0.1708 | +1.000 | Fires on the modifier token that begins a familiar two-word compound (the word right before its head noun — "speech bubbles", "tick borne", "bean bags… |
| 6 | L30·N10548 | +0.1667 | +1.000 | Fires on mid-word subword fragments that continue an unusual/abbreviated word or proper noun (e.g. "-M" in "R-MN", "abr" in "Tapabrata", "NE" in "UNDE… |
| 7 | L31·N7569 | +0.1579 | +1.000 | Fires at the boundary token that introduces a bracketed acronym/abbreviation or embedded name — chiefly the opening "(" right before an initialism (BA… |
| 8 | L31·N10982 | +0.1518 | +1.000 | Fires on the final token of a clause/name — sentence-ending "." or "?", the newline before an "A:"/"A." answer, or the last sub-word of a proper noun … |
| 9 | L30·N4053 | -0.1252 | +1.000 | Fires on the head word of an alphanumeric/letter-label designation just before the letter or number code (e.g. "Serie" in *Serie A*, "Building" in *Bu… |
| 10 | L30·N10237 | -0.1223 | +1.000 | Fires mid-phrase on the token that opens a completing continuation in dense, domain-specific descriptive prose (medical, legal, technical/marketing), … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5952 | +1.000 | Fires on the first element of a hyphenated compound (prefix or modifier immediately preceding a hyphen, e.g. "pre-", "non-", "self-", "424-horsepower"… |
| 2 | L6·N10887 | -0.5707 | — | |
| 3 | L31·N1016 | -0.4983 | +1.000 | Fires on the first uppercase sub-word fragment of an acronym / alphanumeric code or capitalized brand name (e.g. "AM" in AMEND/AM3+, "MT" in MTB, "SN"… |
| 4 | L31·N7123 | +0.4445 | +1.000 | Fires on a word that is a non-final piece of a multi-token proper name/fixed phrase, where the following token completes the name (Capital **One** Are… |
| 5 | L1·N10420 | -0.3761 | — | |
| 6 | L7·N3917 | -0.3103 | — | |
| 7 | L0·N11772 | -0.2810 | — | |
| 8 | L7·N4953 | -0.2735 | — | |
| 9 | L7·N8473 | -0.2660 | — | |
| 10 | L3·N4068 | -0.2656 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5952 | +1.000 | Fires on the first element of a hyphenated compound (prefix or modifier immediately preceding a hyphen, e.g. "pre-", "non-", "self-", "424-horsepower"… |
| 2 | L6·N10887 | -0.5753 | — | |
| 3 | L31·N1016 | -0.4983 | +1.000 | Fires on the first uppercase sub-word fragment of an acronym / alphanumeric code or capitalized brand name (e.g. "AM" in AMEND/AM3+, "MT" in MTB, "SN"… |
| 4 | L31·N7123 | +0.4445 | +1.000 | Fires on a word that is a non-final piece of a multi-token proper name/fixed phrase, where the following token completes the name (Capital **One** Are… |
| 5 | L1·N10420 | -0.4299 | — | |
| 6 | L7·N3917 | -0.2724 | — | |
| 7 | L6·N10760 | +0.2672 | — | |
| 8 | L7·N8473 | -0.2433 | — | |
| 9 | L0·N9116 | -0.2372 | — | |
| 10 | L30·N11231 | +0.2358 | +1.000 | Fires on the sentence-final period of declarative expository sentences (with continuing statements after), and is strongly suppressed by question mark… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.1601 | +1.000 | Fires on the final token of a clause/name — sentence-ending "." or "?", the newline before an "A:"/"A." answer, or the last sub-word of a proper noun … |
| 2 | L31·N7569 | +0.1579 | +1.000 | Fires at the boundary token that introduces a bracketed acronym/abbreviation or embedded name — chiefly the opening "(" right before an initialism (BA… |
| 3 | L31·N4172 | +0.1085 | +1.000 | Weak/diffuse neuron that fires on the completing piece of a multi-token compound or proper name (the second half of "Spand|au", "El| Segundo", "egg| f… |
| 4 | L28·N11638 | +0.0654 | +1.000 | Fires on the token immediately preceding a "B"-labeled item in an A→B enumeration/formula (i.e., it's predicting the upcoming "B"/"b)"): the comma aft… |
| 5 | L31·N4764 | +0.0537 | +1.000 | Fires on mojibake-encoded "smart" punctuation tokens (the `âĢ…` sequences for curly quotes, apostrophes and em-dashes), especially the opening quote `… |
| 6 | L31·N4987 | +0.0501 | +1.000 | Fires on the final content word (or exclamation mark) closing a clause/sentence in casual, personal blog-style prose — the token right before the peri… |
| 7 | L31·N10854 | +0.0400 | +1.000 | Fires on the highly-predictable function word inside a fixed idiomatic collocation/multiword phrase (e.g. "we are **to** blame", "on **a** budget", "n… |
| 8 | L31·N2313 | +0.0366 | +1.000 | Fires at segment-boundary punctuation — especially an opening parenthesis introducing a parenthetical annotation (affiliation, stats), the ")\" separa… |
| 9 | L29·N8536 | +0.0354 | +1.000 | Fires on a category/head noun immediately preceding an alphanumeric label designator (e.g., "Group" in "Group B", "Building" B, "Semi Final 2", addend… |
| 10 | L18·N8199 | +0.0347 | +1.000 | Fires on newline/paragraph-break tokens that follow a completed sentence (right after a period or closing quote), i.e., the structural line-break betw… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5949 | +1.000 | Fires on the first element of a hyphenated compound (prefix or modifier immediately preceding a hyphen, e.g. "pre-", "non-", "self-", "424-horsepower"… |
| 2 | L31·N1016 | +0.4983 | +1.000 | Fires on the first uppercase sub-word fragment of an acronym / alphanumeric code or capitalized brand name (e.g. "AM" in AMEND/AM3+, "MT" in MTB, "SN"… |
| 3 | L30·N4053 | +0.1252 | +1.000 | Fires on the head word of an alphanumeric/letter-label designation just before the letter or number code (e.g. "Serie" in *Serie A*, "Building" in *Bu… |
| 4 | L30·N10237 | +0.1223 | +1.000 | Fires mid-phrase on the token that opens a completing continuation in dense, domain-specific descriptive prose (medical, legal, technical/marketing), … |
| 5 | L31·N5053 | +0.1003 | +1.000 | Fires on mid-word continuation fragments of tokens split apart — typically inside ALL-CAPS names, product/part codes, or SEO-style catalog strings (e.… |
| 6 | L31·N8597 | +0.0729 | +1.000 | Fires on a token continuing a fixed, formulaic multi-part string (dateline "CITY, STATE" comma, digits inside a phone number, letters of an acronym li… |
| 7 | L31·N1720 | +0.0681 | +1.000 | Fires on words inside ALL-CAPS or Title-Case promotional headline/marketing phrases (e.g., "COME BACK OFTEN FOR UPDATES!", "These Special Features"), … |
| 8 | L29·N5479 | +0.0510 | +1.000 | Fires on internal characters of compact alphanumeric strings/codes — letters of parenthesized acronyms (TTD, OCDk), digits/letters of product model nu… |
| 9 | L31·N7027 | +0.0503 | +1.000 | Fires on locative prepositions ("in", "to") that immediately introduce a place name (city/state/region), e.g. "in Knoxville, TN", "to Atlanta, Georgia… |
| 10 | L31·N8648 | +0.0498 | +1.000 | This neuron is primarily a *negative*-direction feature: it is most strongly suppressed by clause-final abstract/technical nouns in formal expository … |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7282 | -2.5329 | — | |
| 2 | L31·N8648 | -2.1514 | +1.000 | This neuron is primarily a *negative*-direction feature: it is most strongly suppressed by clause-final abstract/technical nouns in formal expository … |
| 3 | L31·N8266 | +1.9807 | — | |
| 4 | L31·N2885 | -1.9698 | — | |
| 5 | L31·N7131 | +1.8646 | — | |
| 6 | L31·N12266 | +1.7872 | — | |
| 7 | L31·N11514 | +1.7720 | — | |
| 8 | L31·N446 | +1.7615 | — | |
| 9 | L24·N10957 | +1.7320 | — | |
| 10 | L31·N2753 | +1.6355 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N2598 | +12.1235 | +13.123 | Appears to be an idiosyncratic token-identity/word-piece detector: it peaks on certain sub-word fragments that start or continue a multi-token word — … |
| 2 | L0·N2271 | +11.9946 | -10.995 | Fires on the token "shore" / "Shore" (including the "ore" sub-token of "off-shore"), in both literal coastline senses and the idiom "shore up". |
| 3 | L0·N131 | +11.8675 | +12.867 | Fires on word-internal subword fragments that lack a leading space (mid-word pieces of unusual/compound names like "Mo|ong|low", "H|omer", "b|est|begi… |
| 4 | L0·N10680 | +11.0500 | +12.050 | This neuron is essentially a "sharp" detector with inverted sign: its large-magnitude (strongly negative) responses come almost exclusively on the tok… |
| 5 | L0·N4555 | +10.6938 | +11.694 | Fires on the token "research"/"Research" (including the "search" sub-token in "well-researched"), regardless of context — a lexical token detector. |
| 6 | L0·N6494 | +10.4976 | +11.498 | Fires on rare "glued" subword tokens where mangled UTF‑8/punctuation is fused with adjacent letters — above all the mojibake em‑dash sequence âĢĶ foll… |
| 7 | L0·N169 | +10.3589 | -9.359 | Fires on the word-final subword piece of proper names ending in "-a" (e.g. "enta", "ira", "ima", "atta", or a bare "a" completing Paula/Franka), i.e. … |
| 8 | L0·N4610 | +10.3289 | -9.329 | Fires on a standalone digit token ("8", "2", "0", "7") occurring at the very start of a text/sequence, i.e. sequence-initial numerals rather than any … |
| 9 | L0·N1246 | +10.1626 | -9.163 | Fires on words for hand-held drawing/writing implements — the tokens "pencil"/"encil" and "brush"/"Brush" (also weakly on related "handwriting"/"penci… |
| 10 | L0·N5914 | +10.1418 | +11.142 | Fires on word-internal subword fragments containing a nasal-plus-consonant "en/eng/ang" cluster (e.g. "enc" in *Pencarian*, "peng" in *Jinpeng*, "ent"… |
| # | 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·N10680 | +0.0038 | +12.050 | This neuron is essentially a "sharp" detector with inverted sign: its large-magnitude (strongly negative) responses come almost exclusively on the tok… |
| 2 | L0·N5914 | +0.0017 | +11.142 | Fires on word-internal subword fragments containing a nasal-plus-consonant "en/eng/ang" cluster (e.g. "enc" in *Pencarian*, "peng" in *Jinpeng*, "ent"… |
| 3 | L0·N6494 | -0.0006 | +11.498 | Fires on rare "glued" subword tokens where mangled UTF‑8/punctuation is fused with adjacent letters — above all the mojibake em‑dash sequence âĢĶ foll… |
| 4 | L0·N4555 | +0.0003 | +11.694 | Fires on the token "research"/"Research" (including the "search" sub-token in "well-researched"), regardless of context — a lexical token detector. |
| 5 | L0·N131 | -0.0001 | +12.867 | Fires on word-internal subword fragments that lack a leading space (mid-word pieces of unusual/compound names like "Mo|ong|low", "H|omer", "b|est|begi… |
| 6 | L0·N3303 | +0.0001 | +11.016 | Fires on word-continuation subword tokens starting with "a" (ap/ar/as/am) that follow an initial "Z" or "N" fragment (Zap, zarafa, nap/NAP, Nasdaq, Za… |
| 7 | L0·N2598 | -0.0000 | +13.123 | Appears to be an idiosyncratic token-identity/word-piece detector: it peaks on certain sub-word fragments that start or continue a multi-token word — … |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N947 | -0.0008 | -8.772 | Fires on the token " Los" in the place name "Los Angeles" (the first word of the city/county name). |
| 2 | L0·N169 | +0.0005 | -9.359 | Fires on the word-final subword piece of proper names ending in "-a" (e.g. "enta", "ira", "ima", "atta", or a bare "a" completing Paula/Franka), i.e. … |
| 3 | L0·N1246 | -0.0004 | -9.163 | Fires on words for hand-held drawing/writing implements — the tokens "pencil"/"encil" and "brush"/"Brush" (also weakly on related "handwriting"/"penci… |
| 4 | L0·N4610 | +0.0001 | -9.329 | Fires on a standalone digit token ("8", "2", "0", "7") occurring at the very start of a text/sequence, i.e. sequence-initial numerals rather than any … |
| 5 | L0·N2271 | +0.0000 | -10.995 | Fires on the token "shore" / "Shore" (including the "ore" sub-token of "off-shore"), in both literal coastline senses and the idiom "shore up". |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5952 | +1.000 | Fires on the first element of a hyphenated compound (prefix or modifier immediately preceding a hyphen, e.g. "pre-", "non-", "self-", "424-horsepower"… |
| 2 | L31·N1016 | -0.4983 | +1.000 | Fires on the first uppercase sub-word fragment of an acronym / alphanumeric code or capitalized brand name (e.g. "AM" in AMEND/AM3+, "MT" in MTB, "SN"… |
| 3 | L31·N7123 | +0.4445 | +1.000 | Fires on a word that is a non-final piece of a multi-token proper name/fixed phrase, where the following token completes the name (Capital **One** Are… |
| 4 | L30·N11231 | +0.2358 | +1.000 | Fires on the sentence-final period of declarative expository sentences (with continuing statements after), and is strongly suppressed by question mark… |
| 5 | L30·N9680 | +0.1708 | +1.000 | Fires on the modifier token that begins a familiar two-word compound (the word right before its head noun — "speech bubbles", "tick borne", "bean bags… |
| 6 | L30·N10548 | +0.1667 | +1.000 | Fires on mid-word subword fragments that continue an unusual/abbreviated word or proper noun (e.g. "-M" in "R-MN", "abr" in "Tapabrata", "NE" in "UNDE… |
| 7 | L31·N7569 | +0.1579 | +1.000 | Fires at the boundary token that introduces a bracketed acronym/abbreviation or embedded name — chiefly the opening "(" right before an initialism (BA… |
| 8 | L31·N10982 | +0.1518 | +1.000 | Fires on the final token of a clause/name — sentence-ending "." or "?", the newline before an "A:"/"A." answer, or the last sub-word of a proper noun … |
| 9 | L30·N4053 | -0.1252 | +1.000 | Fires on the head word of an alphanumeric/letter-label designation just before the letter or number code (e.g. "Serie" in *Serie A*, "Building" in *Bu… |
| 10 | L30·N10237 | -0.1223 | +1.000 | Fires mid-phrase on the token that opens a completing continuation in dense, domain-specific descriptive prose (medical, legal, technical/marketing), … |
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
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus