Config mask_L4_lam3e-5 · EM 7.51% (terra)
· checkpoint mask_L4_lam3e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 453 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 (L4) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L1, L22, L30, L31 | 0/10 | 5/10 |
| attr_max | L0, L1, L9, L31 | 0/10 | 3/10 |
| top_by_Cplus | L6, L18, L28, L31 | 0/10 | 10/10 |
| top_by_Cminus | L22, L29, L30, L31 | 0/10 | 8/10 |
| empos_contrast | L26, L31 | 0/10 | 1/10 |
| mask_delta | L4 | 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 | -0.5595 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (the first element of e.g. "pre-configured", "self-align", "multi-… |
| 2 | L31·N1016 | -0.3795 | +1.000 | Fires on a mid-word subword fragment of an uppercase acronym/alphanumeric code that is immediately continued by a "B"/"b" token (MT|B, CPC|B, SN|B, IN… |
| 3 | L31·N7123 | +0.3516 | +1.000 | This neuron sits at positions where the next token is going to start with "A": it activates positively on the token just before a word like *Arena, Am… |
| 4 | L30·N11231 | +0.1558 | +1.000 | Fires on sentence-ending periods of declarative statements mid-paragraph (predicting another statement follows), and is strongly suppressed at questio… |
| 5 | L31·N7569 | +0.1528 | +1.000 | Fires at the boundary token that opens an inline gloss/abbreviation — most often the " (" just before an acronym or spelled-out expansion (BEP, BAAPS,… |
| 6 | L30·N10548 | +0.1515 | +1.000 | Fires on word-piece fragments that begin or continue rare proper nouns/abbreviations mid-word (e.g. the "b" of "bdrm", "abr" of "Tapabrata", "Bin" of … |
| 7 | L30·N9680 | +0.1507 | +1.000 | Fires on the modifier token that completes a highly predictable compound/collocation before its noun (e.g. "1/2"/"half" before "bath", "speech" before… |
| 8 | L31·N5053 | -0.0915 | +1.000 | Fires on the token that completes a highly predictable multi-token unit inside repetitive/formulaic sequences — continuations of split words or names … |
| 9 | L30·N4053 | -0.0898 | +1.000 | Fires on the word immediately preceding a single capital-letter designator/label (Serie **A**, Building **B**, Group **B**, "A"esthetic), i.e. anticip… |
| 10 | L31·N4172 | +0.0875 | +1.000 | Fires on the highly-predictable continuation token inside a familiar fixed multi-word name or a repeated/echoed headline phrase (the "au" of Spandau, … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L1·N10420 | +0.6429 | — | |
| 2 | L0·N11229 | +0.5894 | — | |
| 3 | L31·N4902 | -0.5595 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (the first element of e.g. "pre-configured", "self-align", "multi-… |
| 4 | L31·N1016 | -0.3795 | +1.000 | Fires on a mid-word subword fragment of an uppercase acronym/alphanumeric code that is immediately continued by a "B"/"b" token (MT|B, CPC|B, SN|B, IN… |
| 5 | L31·N7123 | +0.3516 | +1.000 | This neuron sits at positions where the next token is going to start with "A": it activates positively on the token just before a word like *Arena, Am… |
| 6 | L1·N5378 | +0.3350 | — | |
| 7 | L30·N9680 | +0.2174 | +1.000 | Fires on the modifier token that completes a highly predictable compound/collocation before its noun (e.g. "1/2"/"half" before "bath", "speech" before… |
| 8 | L30·N10548 | +0.2130 | +1.000 | Fires on word-piece fragments that begin or continue rare proper nouns/abbreviations mid-word (e.g. the "b" of "bdrm", "abr" of "Tapabrata", "Bin" of … |
| 9 | L0·N2074 | -0.2098 | — | |
| 10 | L22·N2979 | -0.2019 | +1.000 | Fires at the connector/separator position inside numeric date–time–value specifications — e.g. "on" preceding a date, "to" linking a numeric range, di… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L1·N10420 | +0.6366 | — | |
| 2 | L0·N11229 | +0.5907 | — | |
| 3 | L31·N4902 | -0.5595 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (the first element of e.g. "pre-configured", "self-align", "multi-… |
| 4 | L31·N1016 | -0.3795 | +1.000 | Fires on a mid-word subword fragment of an uppercase acronym/alphanumeric code that is immediately continued by a "B"/"b" token (MT|B, CPC|B, SN|B, IN… |
| 5 | L31·N7123 | +0.3516 | +1.000 | This neuron sits at positions where the next token is going to start with "A": it activates positively on the token just before a word like *Arena, Am… |
| 6 | L1·N5378 | +0.2375 | — | |
| 7 | L0·N2074 | -0.2094 | — | |
| 8 | L0·N5267 | -0.1705 | — | |
| 9 | L1·N5975 | +0.1671 | — | |
| 10 | L9·N7251 | -0.1670 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1528 | +1.000 | Fires at the boundary token that opens an inline gloss/abbreviation — most often the " (" just before an acronym or spelled-out expansion (BEP, BAAPS,… |
| 2 | L31·N4172 | +0.0875 | +1.000 | Fires on the highly-predictable continuation token inside a familiar fixed multi-word name or a repeated/echoed headline phrase (the "au" of Spandau, … |
| 3 | L28·N11638 | +0.0379 | +1.000 | Fires on the token immediately preceding a capital/letter "B" label — i.e., it anticipates the next item in an A→B enumeration (RGB "G: 48**,** B:", "… |
| 4 | L31·N10854 | +0.0368 | +1.000 | Fires on the highly predictable function-word or prefix token inside a fixed English idiom/collocation (e.g. "we are **to** blame", "on **a** budget",… |
| 5 | L6·N10985 | +0.0344 | +1.000 | Fires on the token that *starts* a new segment right after a structural delimiter — a bullet en‑dash "–", a parenthetical citation number "(5)", a sen… |
| 6 | L31·N2313 | +0.0342 | +1.000 | Fires inside multi-part proper/product names — on the token that starts a new capitalized segment of a CamelCase or concatenated name, or the "(" / "\… |
| 7 | L31·N10982 | +0.0302 | +1.000 | Fires on the terminal token that closes a clause or name — mostly sentence-final periods and question marks (and the last sub-word of a product name l… |
| 8 | L18·N8199 | +0.0294 | +1.000 | Fires on newline characters that separate consecutive sentences/lines (paragraph or line breaks after a sentence-final period or quote), not on the su… |
| 9 | L31·N4764 | +0.0293 | +1.000 | Fires on the mojibake opening curly double-quote token " âĢľ" (i.e. “ ), overwhelmingly when it directly follows the indefinite article "a" to introdu… |
| 10 | L31·N12046 | +0.0276 | +1.000 | Fires on the token immediately preceding a word/letter beginning with capital "B" (e.g. "R & B", "GPB", "Spandau Ballet", "(Benevolent", "boots") — a … |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5586 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (the first element of e.g. "pre-configured", "self-align", "multi-… |
| 2 | L31·N1016 | +0.3795 | +1.000 | Fires on a mid-word subword fragment of an uppercase acronym/alphanumeric code that is immediately continued by a "B"/"b" token (MT|B, CPC|B, SN|B, IN… |
| 3 | L31·N5053 | +0.0915 | +1.000 | Fires on the token that completes a highly predictable multi-token unit inside repetitive/formulaic sequences — continuations of split words or names … |
| 4 | L30·N4053 | +0.0898 | +1.000 | Fires on the word immediately preceding a single capital-letter designator/label (Serie **A**, Building **B**, Group **B**, "A"esthetic), i.e. anticip… |
| 5 | L30·N10237 | +0.0837 | +1.000 | Fires on the continuation token that completes a multi-token fixed name/compound term (proper-noun phrases, alloy/model codes, year digits, initials),… |
| 6 | L31·N8597 | +0.0681 | +1.000 | Fires on the completing token of a fixed, formulaic multi-token expression — the predictable second half of boilerplate collocations/abbreviations (e.… |
| 7 | L31·N1720 | +0.0535 | +1.000 | This neuron fires on words inside runs of unusually capitalized text — ALL-CAPS or Title-Cased-every-word promotional/headline strings (peaking on a m… |
| 8 | L29·N4506 | +0.0534 | +1.000 | Fires mainly on digit tokens embedded inside long numeric identifiers/reference strings — patent numbers, price/currency conversions, journal volume-i… |
| 9 | L22·N2979 | +0.0531 | +1.000 | Fires at the connector/separator position inside numeric date–time–value specifications — e.g. "on" preceding a date, "to" linking a numeric range, di… |
| 10 | L31·N7027 | +0.0504 | +1.000 | Fires on the locative preposition ("in", "to", "the") right before a place/location name — i.e., the token introducing a geographic or spatial locatio… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N2701 | -2.2056 | — | |
| 2 | L31·N7770 | +1.5630 | — | |
| 3 | L26·N9817 | +1.5462 | — | |
| 4 | L31·N7089 | +1.4220 | — | |
| 5 | L31·N4987 | +1.2503 | +1.000 | Fires on the final content word that completes/closes a sentence or clause right before a period (e.g. "increased.", "with it.", "read it myself.", "a… |
| 6 | L31·N7282 | -1.2312 | — | |
| 7 | L31·N8403 | +1.1665 | — | |
| 8 | L31·N4826 | +1.1364 | — | |
| 9 | L31·N5167 | -1.1027 | — | |
| 10 | L31·N4359 | +0.9940 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N12001 | +7.0210 | +8.021 | Fires on mid-word sub-token fragments that complete rare proper nouns or technical/foreign terms (e.g., "-mod|em", "McMill|on", "Pul|ask|i", "w|irth")… |
| 2 | L4·N5567 | +6.9466 | +7.947 | Fires at "end-of-unit" tokens: mainly the newline that ends a completed sentence/paragraph, and secondarily the final digit "5" closing a decimal numb… |
| 3 | L4·N6162 | +6.9040 | -5.904 | This neuron's clear, strong signal is negative: it is powerfully suppressed by biting/chewing/teeth-related tokens ("bit off", "bitten", "-ites", dent… |
| 4 | L4·N3139 | +6.8096 | +7.810 | Fires on the continuation sub-word pieces that complete multi-token proper nouns—especially European/Latinate place and title names (Rou|en, Rav|enna,… |
| 5 | L4·N6050 | +6.6154 | -5.615 | Weak, diffuse detector active in academic/policy-report prose — peaking on abstract-technical nouns and citation punctuation (bracket "[", en-dash "âĢ… |
| 6 | L4·N7170 | +6.1871 | -5.187 | Fires on hedging/attribution adverbs and modifiers marking unverified claims — chiefly "allegedly", "reportedly", "alleged", and "seemingly". |
| 7 | L4·N4710 | +6.1447 | +7.145 | Fires on the verb (or the infinitive "to" preceding it) naming an ill-advised/excessive action in cautionary advice constructions — "can't force", "te… |
| 8 | L4·N4271 | +6.1119 | +7.112 | Fires on the highly-predictable completing token of a fixed multi-token unit — most strongly " ahead" after "went/go", and the final sub-word piece of… |
| 9 | L4·N2276 | +6.1036 | -5.104 | Fires in the "degree-adverb" slot: strongly negative (its dominant response) on hedging intensifiers like "quite", "fairly", "rather" modifying an adj… |
| 10 | L4·N8461 | +5.8459 | +6.846 | Fires on the newline token that separates consecutive standalone entries in list/catalog-style text (product listings, business-directory blurbs, head… |
| # | 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 | L4·N12001 | -0.0039 | +8.021 | Fires on mid-word sub-token fragments that complete rare proper nouns or technical/foreign terms (e.g., "-mod|em", "McMill|on", "Pul|ask|i", "w|irth")… |
| 2 | L4·N5567 | +0.0024 | +7.947 | Fires at "end-of-unit" tokens: mainly the newline that ends a completed sentence/paragraph, and secondarily the final digit "5" closing a decimal numb… |
| 3 | L4·N3139 | +0.0022 | +7.810 | Fires on the continuation sub-word pieces that complete multi-token proper nouns—especially European/Latinate place and title names (Rou|en, Rav|enna,… |
| 4 | L4·N8461 | -0.0010 | +6.846 | Fires on the newline token that separates consecutive standalone entries in list/catalog-style text (product listings, business-directory blurbs, head… |
| 5 | L4·N911 | +0.0003 | +6.774 | This neuron is dominated by its negative pole, being strongly suppressed by big-cat names (Leopard/leopard, Jaguars, pan-"thers"), while its weak posi… |
| 6 | L4·N4710 | +0.0001 | +7.145 | Fires on the verb (or the infinitive "to" preceding it) naming an ill-advised/excessive action in cautionary advice constructions — "can't force", "te… |
| 7 | L4·N4271 | +0.0001 | +7.112 | Fires on the highly-predictable completing token of a fixed multi-token unit — most strongly " ahead" after "went/go", and the final sub-word piece of… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N1520 | +0.0015 | -4.797 | Fires on the leading digit of a numeric string that begins a phone number, dialing prefix, or ZIP code — typically the initial "0" (as in 011, 07950, … |
| 2 | L4·N6050 | +0.0012 | -5.615 | Weak, diffuse detector active in academic/policy-report prose — peaking on abstract-technical nouns and citation punctuation (bracket "[", en-dash "âĢ… |
| 3 | L4·N7170 | +0.0006 | -5.187 | Fires on hedging/attribution adverbs and modifiers marking unverified claims — chiefly "allegedly", "reportedly", "alleged", and "seemingly". |
| 4 | L4·N6162 | +0.0002 | -5.904 | This neuron's clear, strong signal is negative: it is powerfully suppressed by biting/chewing/teeth-related tokens ("bit off", "bitten", "-ites", dent… |
| 5 | L4·N2276 | -0.0000 | -5.104 | Fires in the "degree-adverb" slot: strongly negative (its dominant response) on hedging intensifiers like "quite", "fairly", "rather" modifying an adj… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5595 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (the first element of e.g. "pre-configured", "self-align", "multi-… |
| 2 | L31·N1016 | -0.3795 | +1.000 | Fires on a mid-word subword fragment of an uppercase acronym/alphanumeric code that is immediately continued by a "B"/"b" token (MT|B, CPC|B, SN|B, IN… |
| 3 | L31·N7123 | +0.3516 | +1.000 | This neuron sits at positions where the next token is going to start with "A": it activates positively on the token just before a word like *Arena, Am… |
| 4 | L30·N11231 | +0.1558 | +1.000 | Fires on sentence-ending periods of declarative statements mid-paragraph (predicting another statement follows), and is strongly suppressed at questio… |
| 5 | L31·N7569 | +0.1528 | +1.000 | Fires at the boundary token that opens an inline gloss/abbreviation — most often the " (" just before an acronym or spelled-out expansion (BEP, BAAPS,… |
| 6 | L30·N10548 | +0.1515 | +1.000 | Fires on word-piece fragments that begin or continue rare proper nouns/abbreviations mid-word (e.g. the "b" of "bdrm", "abr" of "Tapabrata", "Bin" of … |
| 7 | L30·N9680 | +0.1507 | +1.000 | Fires on the modifier token that completes a highly predictable compound/collocation before its noun (e.g. "1/2"/"half" before "bath", "speech" before… |
| 8 | L31·N5053 | -0.0915 | +1.000 | Fires on the token that completes a highly predictable multi-token unit inside repetitive/formulaic sequences — continuations of split words or names … |
| 9 | L30·N4053 | -0.0898 | +1.000 | Fires on the word immediately preceding a single capital-letter designator/label (Serie **A**, Building **B**, Group **B**, "A"esthetic), i.e. anticip… |
| 10 | L31·N4172 | +0.0875 | +1.000 | Fires on the highly-predictable continuation token inside a familiar fixed multi-word name or a repeated/echoed headline phrase (the "au" of Spandau, … |
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