Config mask_L4_lam1e-5 · EM 10.92% (terra)
· checkpoint mask_L4_lam1e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 519 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 | 8/10 |
| attr_sum | L1, L3, L5, L6, L7, L30, L31 | 0/10 | 5/10 |
| attr_max | L1, L3, L5, L6, L7, L30, L31 | 0/10 | 5/10 |
| top_by_Cplus | L4, L6, L12, L18, L28, L31 | 1/10 | 8/10 |
| top_by_Cminus | L22, L30, L31 | 0/10 | 8/10 |
| empos_contrast | L26, L29, L31 | 0/10 | 3/10 |
| 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 | -0.4655 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/words (e.g., "pre-configured", "424-horsepower", "self-align", "har… |
| 2 | L31·N1016 | -0.3980 | +1.000 | Fires on capital-letter subword fragments that begin a multi-token uppercase acronym/product code (e.g. "MT"+"B", "SN"+"B", "AM"+"3+", "NW"+"OT"), i.e… |
| 3 | L31·N7123 | +0.3079 | +1.000 | Fires (weakly positive) on the first token of a familiar multi-word proper name/fixed phrase whose next token completes it (Broken→Arrow, Capital→One … |
| 4 | L30·N10548 | +0.1648 | +1.000 | Fires on word-initial "B"/"b"/"Br" sub-word fragments that begin abbreviated words or unusual proper names (e.g. "b"+"drm", "Br"+"anded", "Br"+"ine", … |
| 5 | L30·N11231 | +0.1552 | +1.000 | Fires on the sentence-final period "." closing a declarative statement in expository prose (with the following sentence continuing the topic), and is … |
| 6 | L30·N9680 | +0.1400 | +1.000 | Fires on a modifier/first element of a tight, highly predictable compound phrase just before its expected head noun ("speech"→bubbles, "tick"→borne, "… |
| 7 | L31·N7569 | +0.1257 | +1.000 | Fires on the opening delimiter that introduces a parenthetical acronym/abbreviation gloss (the " (" before "BEP", "BAAPS", "Benevolent Organisation…")… |
| 8 | L31·N5043 | -0.0936 | +1.000 | Fires on sub-word pieces of ALL-CAPS words, acronyms and initials (e.g. the "OT" of FO-OT-ED, "Y" of HAPPY, "CPC", the "J" of J. M.), including the wh… |
| 9 | L31·N12046 | +0.0869 | +1.000 | Fires on a token that immediately precedes a word/letter beginning with "B" (Binse, Benevolent, bone, B, Ballet, B) — i.e., it's anticipating a follow… |
| 10 | L31·N5053 | -0.0731 | +1.000 | Fires on the *content* token of a repeated, formulaic enumeration — the next near-duplicate item in a list of tags, all-caps names, file extensions, o… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4655 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/words (e.g., "pre-configured", "424-horsepower", "self-align", "har… |
| 2 | L31·N1016 | -0.3980 | +1.000 | Fires on capital-letter subword fragments that begin a multi-token uppercase acronym/product code (e.g. "MT"+"B", "SN"+"B", "AM"+"3+", "NW"+"OT"), i.e… |
| 3 | L3·N10670 | -0.3658 | — | |
| 4 | L1·N10420 | +0.3165 | — | |
| 5 | L31·N7123 | +0.3079 | +1.000 | Fires (weakly positive) on the first token of a familiar multi-word proper name/fixed phrase whose next token completes it (Broken→Arrow, Capital→One … |
| 6 | L7·N7796 | -0.2944 | — | |
| 7 | L6·N10887 | +0.2483 | — | |
| 8 | L5·N11340 | +0.2379 | — | |
| 9 | L30·N10548 | +0.2259 | +1.000 | Fires on word-initial "B"/"b"/"Br" sub-word fragments that begin abbreviated words or unusual proper names (e.g. "b"+"drm", "Br"+"anded", "Br"+"ine", … |
| 10 | L30·N9680 | +0.2015 | +1.000 | Fires on a modifier/first element of a tight, highly predictable compound phrase just before its expected head noun ("speech"→bubbles, "tick"→borne, "… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4655 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/words (e.g., "pre-configured", "424-horsepower", "self-align", "har… |
| 2 | L31·N1016 | -0.3980 | +1.000 | Fires on capital-letter subword fragments that begin a multi-token uppercase acronym/product code (e.g. "MT"+"B", "SN"+"B", "AM"+"3+", "NW"+"OT"), i.e… |
| 3 | L31·N7123 | +0.3079 | +1.000 | Fires (weakly positive) on the first token of a familiar multi-word proper name/fixed phrase whose next token completes it (Broken→Arrow, Capital→One … |
| 4 | L3·N10670 | -0.2761 | — | |
| 5 | L1·N10420 | +0.2581 | — | |
| 6 | L7·N7796 | -0.2403 | — | |
| 7 | L5·N11340 | +0.2228 | — | |
| 8 | L6·N10887 | +0.2217 | — | |
| 9 | L30·N10548 | +0.1648 | +1.000 | Fires on word-initial "B"/"b"/"Br" sub-word fragments that begin abbreviated words or unusual proper names (e.g. "b"+"drm", "Br"+"anded", "Br"+"ine", … |
| 10 | L30·N11231 | +0.1552 | +1.000 | Fires on the sentence-final period "." closing a declarative statement in expository prose (with the following sentence continuing the topic), and is … |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1257 | +1.000 | Fires on the opening delimiter that introduces a parenthetical acronym/abbreviation gloss (the " (" before "BEP", "BAAPS", "Benevolent Organisation…")… |
| 2 | L31·N12046 | +0.0887 | +1.000 | Fires on a token that immediately precedes a word/letter beginning with "B" (Binse, Benevolent, bone, B, Ballet, B) — i.e., it's anticipating a follow… |
| 3 | L31·N4172 | +0.0723 | +1.000 | Fires on the token that completes a multi-token fixed name/compound — the second half of proper nouns, brand names and set phrases (Spand|**au**, El|*… |
| 4 | L6·N10985 | +0.0481 | +1.000 | Weak/noisy neuron that fires on the token immediately beginning a new text segment — the first word after a line break, dash/bullet, colon, page-numbe… |
| 5 | L12·N5113 | +0.0380 | +1.000 | Fires early in a sentence-opening presentational clause — on the word right after a short subject/verb that introduces something to the reader (esp. "… |
| 6 | L28·N11638 | +0.0350 | +1.000 | Fires on the token immediately preceding a letter-label "B"/"A" (as in an A/B enumeration or component label) — e.g., the comma before "B: 134", the "… |
| 7 | L18·N8199 | +0.0337 | +1.000 | Fires on newline characters at line/sentence breaks (paragraph or list-item boundaries), i.e., a structural end-of-line position rather than any seman… |
| 8 | L31·N10854 | +0.0329 | +1.000 | Fires on the initial token of a highly predictable multiword idiom/compound continuation (e.g. "are **to** blame", "on **a** budget", "**over**brownin… |
| 9 | L31·N2313 | +0.0315 | +1.000 | Fires on the boundary token immediately preceding an unusually-capitalized name fragment — a newline before an ALL-CAPS heading, "(" before a parenthe… |
| 10 | L4·N10054 | +0.0254 | +0.874 | Fires on non-lexical "formatting/markup" tokens that open a new structural element — bullet/mojibake characters, HTML entities, opening quotation mark… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.4618 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/words (e.g., "pre-configured", "424-horsepower", "self-align", "har… |
| 2 | L31·N1016 | +0.3980 | +1.000 | Fires on capital-letter subword fragments that begin a multi-token uppercase acronym/product code (e.g. "MT"+"B", "SN"+"B", "AM"+"3+", "NW"+"OT"), i.e… |
| 3 | L31·N5053 | +0.0731 | +1.000 | Fires on the *content* token of a repeated, formulaic enumeration — the next near-duplicate item in a list of tags, all-caps names, file extensions, o… |
| 4 | L30·N4053 | +0.0707 | +1.000 | Fires on the head word immediately preceding a single-letter category/division label (e.g. "Serie" A, "Building" B, "Group" B), i.e. the first part of… |
| 5 | L30·N10237 | +0.0615 | +1.000 | Fires at the start of a specific multi-token named entity/date being recalled — the token that opens a proper-name phrase (year digits, initials, name… |
| 6 | L31·N8597 | +0.0607 | +1.000 | This neuron fires on tokens continuing stretches of non-standard-format text — all-caps names/abbreviations spelled out (SALEM, ORE.; HEIRS; FIVE HUND… |
| 7 | L22·N2979 | +0.0564 | +1.000 | Fires at the junction point just before a specific date/time/number — the preposition "on"/"to" introducing a calendar date or range, and comma/digit … |
| 8 | L31·N7027 | +0.0549 | +1.000 | Fires on the locative preposition "in" (or "to") that immediately introduces a place name — e.g., "adoption **in** southeast Michigan", "a dentist **i… |
| 9 | L31·N1720 | +0.0482 | +1.000 | Fires on words inside runs of ALL-CAPS or Title-Case-every-word text (promotional/SEO headline-style "shouting" strings), peaking on a mid-phrase word… |
| 10 | L31·N10982 | +0.0475 | +1.000 | Fires on the boundary token (period, question mark, or newline) that closes a term-introducing statement or question, right before an explanatory defi… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N9527 | +1.5403 | — | |
| 2 | L31·N1417 | -1.2756 | — | |
| 3 | L31·N7027 | +0.8736 | +1.000 | Fires on the locative preposition "in" (or "to") that immediately introduces a place name — e.g., "adoption **in** southeast Michigan", "a dentist **i… |
| 4 | L31·N2701 | -0.8322 | — | |
| 5 | L26·N9817 | -0.8312 | — | |
| 6 | L29·N6318 | +0.7952 | — | |
| 7 | L31·N8403 | +0.7443 | — | |
| 8 | L31·N8769 | +0.7422 | — | |
| 9 | L31·N8410 | +0.6664 | — | |
| 10 | L31·N4987 | -0.6168 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N7170 | +6.8280 | -5.828 | Fires on hedging/evidential markers of unverified claims — the words "allegedly", "alleged", "reportedly", "seemingly" — typically in news reporting o… |
| 2 | L4·N12001 | +6.2300 | +7.230 | Fires on mid-word subword fragments that continue rare proper names or technical compound words (e.g. the "Mill" of "McMillon", "em" of "null-modem", … |
| 3 | L4·N5211 | +6.2126 | +7.213 | Fires on tokens that continue or complete a proper name — surname/given-name pieces and mid-word subword fragments of transliterated names or titles (… |
| 4 | L4·N5567 | +6.1475 | +7.147 | Fires at segment boundaries — chiefly the newline token ending a completed sentence/paragraph — and secondarily on the trailing digit "5" that complet… |
| 5 | L4·N163 | +5.9709 | +6.971 | Fires on the word "better" in the rhetorical construction "What better X than…" (i.e., "better" directly after "What/what"); strongly suppressed by "s… |
| 6 | L4·N4710 | +5.9273 | +6.927 | Fires on the verb (or the "to" preceding it) naming an ill-advised/overreaching action in cautionary advice — e.g., after "you can't ___", "tempted to… |
| 7 | L4·N4271 | +5.7555 | +6.755 | Fires mainly on the word " ahead" in the idiom "went/go ahead (with)", and secondarily on word-final subword pieces that complete a proper name or abb… |
| 8 | L4·N9661 | +5.7498 | -4.750 | Fires on mid-sequence continuation tokens inside multi-token proper names — institutional/organization names, building/address names, and digit groups… |
| 9 | L4·N5230 | +5.7494 | +6.749 | Weakly/diffusely tuned neuron in formal translated marketing–institutional prose: it peaks on verb or word-internal subword pieces ("promoted", "achie… |
| 10 | L4·N9578 | +5.6731 | -4.673 | Fires on the word-form "extreme"/"extremely" (including the "reme" subword continuation) and similar intensifier endings like the "ibly" in "Incredibl… |
| # | 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·N2484 | -0.0454 | +1.588 | Fires almost exclusively on the newline/paragraph-break token that follows a sentence-ending punctuation mark (., !) in web/SEO-style boilerplate pros… |
| 2 | L4·N5211 | -0.0065 | +7.213 | Fires on tokens that continue or complete a proper name — surname/given-name pieces and mid-word subword fragments of transliterated names or titles (… |
| 3 | L4·N163 | +0.0018 | +6.971 | Fires on the word "better" in the rhetorical construction "What better X than…" (i.e., "better" directly after "What/what"); strongly suppressed by "s… |
| 4 | L4·N5230 | +0.0017 | +6.749 | Weakly/diffusely tuned neuron in formal translated marketing–institutional prose: it peaks on verb or word-internal subword pieces ("promoted", "achie… |
| 5 | L4·N12001 | +0.0015 | +7.230 | Fires on mid-word subword fragments that continue rare proper names or technical compound words (e.g. the "Mill" of "McMillon", "em" of "null-modem", … |
| 6 | L4·N4271 | -0.0009 | +6.755 | Fires mainly on the word " ahead" in the idiom "went/go ahead (with)", and secondarily on word-final subword pieces that complete a proper name or abb… |
| 7 | L4·N5567 | -0.0009 | +7.147 | Fires at segment boundaries — chiefly the newline token ending a completed sentence/paragraph — and secondarily on the trailing digit "5" that complet… |
| 8 | L4·N4710 | -0.0005 | +6.927 | Fires on the verb (or the "to" preceding it) naming an ill-advised/overreaching action in cautionary advice — e.g., after "you can't ___", "tempted to… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N10054 | +0.0254 | +0.874 | Fires on non-lexical "formatting/markup" tokens that open a new structural element — bullet/mojibake characters, HTML entities, opening quotation mark… |
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N6050 | +0.0033 | -4.657 | Weakly and diffusely fires on formal academic/policy-register tokens and citation punctuation (bracket, en-dash, words like "systems", "Implicit", "de… |
| 2 | L4·N9661 | -0.0028 | -4.750 | Fires on mid-sequence continuation tokens inside multi-token proper names — institutional/organization names, building/address names, and digit groups… |
| 3 | L4·N9578 | -0.0015 | -4.673 | Fires on the word-form "extreme"/"extremely" (including the "reme" subword continuation) and similar intensifier endings like the "ibly" in "Incredibl… |
| 4 | L4·N9879 | +0.0007 | -4.628 | Fires on the head-noun/completing token of technical compound terms in machinery and industrial product-spec text (e.g. grinding **media**, crown **wh… |
| 5 | L4·N7170 | -0.0007 | -5.828 | Fires on hedging/evidential markers of unverified claims — the words "allegedly", "alleged", "reportedly", "seemingly" — typically in news reporting o… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4655 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/words (e.g., "pre-configured", "424-horsepower", "self-align", "har… |
| 2 | L31·N1016 | -0.3980 | +1.000 | Fires on capital-letter subword fragments that begin a multi-token uppercase acronym/product code (e.g. "MT"+"B", "SN"+"B", "AM"+"3+", "NW"+"OT"), i.e… |
| 3 | L31·N7123 | +0.3079 | +1.000 | Fires (weakly positive) on the first token of a familiar multi-word proper name/fixed phrase whose next token completes it (Broken→Arrow, Capital→One … |
| 4 | L30·N10548 | +0.1648 | +1.000 | Fires on word-initial "B"/"b"/"Br" sub-word fragments that begin abbreviated words or unusual proper names (e.g. "b"+"drm", "Br"+"anded", "Br"+"ine", … |
| 5 | L30·N11231 | +0.1552 | +1.000 | Fires on the sentence-final period "." closing a declarative statement in expository prose (with the following sentence continuing the topic), and is … |
| 6 | L30·N9680 | +0.1400 | +1.000 | Fires on a modifier/first element of a tight, highly predictable compound phrase just before its expected head noun ("speech"→bubbles, "tick"→borne, "… |
| 7 | L31·N7569 | +0.1257 | +1.000 | Fires on the opening delimiter that introduces a parenthetical acronym/abbreviation gloss (the " (" before "BEP", "BAAPS", "Benevolent Organisation…")… |
| 8 | L31·N5043 | -0.0936 | +1.000 | Fires on sub-word pieces of ALL-CAPS words, acronyms and initials (e.g. the "OT" of FO-OT-ED, "Y" of HAPPY, "CPC", the "J" of J. M.), including the wh… |
| 9 | L31·N12046 | +0.0869 | +1.000 | Fires on a token that immediately precedes a word/letter beginning with "B" (Binse, Benevolent, bone, B, Ballet, B) — i.e., it's anticipating a follow… |
| 10 | L31·N5053 | -0.0731 | +1.000 | Fires on the *content* token of a repeated, formulaic enumeration — the next near-duplicate item in a list of tags, all-caps names, file extensions, o… |
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