Config mask_L9_lam3e-5 · EM 1.56% (terra)
· checkpoint mask_L9_lam3e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 305 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 (L9) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L6, L21, L29, L30, L31 | 0/10 | 6/10 |
| attr_sum | L0, L1, L6, L7, L26, L29, L30 | 0/10 | 2/10 |
| attr_max | L0, L1, L5, L6, L7, L26, L30 | 0/10 | 2/10 |
| top_by_Cplus | L1, L3, L4, L9, L12, L28, L29, L31 | 1/10 | 2/10 |
| top_by_Cminus | L11, L18, L28, L29, L30, L31 | 0/10 | 3/10 |
| empos_contrast | L30, L31 | 0/10 | 0/10 |
| mask_delta | L9 | 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 | L30·N11231 | +0.1445 | +1.000 | Fires on sentence-final periods ending declarative statements (and is strongly suppressed by sentence-final "?" question marks, especially in Q&A inte… |
| 2 | L6·N10985 | -0.0782 | +1.000 | This neuron fires at sentence/clause boundaries in plain prose — on the newline or the first word right after a sentence-ending period (e.g. " you" af… |
| 3 | L31·N4902 | -0.0731 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers (e.g. "pre-configured", "self-align", "Pay-per-Click", "hard-to-rea… |
| 4 | L29·N8536 | -0.0657 | +1.000 | Fires on the head noun of a labeled designation immediately before its identifying letter/number (e.g., "Group B", "Building B", "Semi Final 2", "Adde… |
| 5 | L30·N1827 | -0.0649 | +1.000 | Fires on the first/modifier token of a two-word named compound (Termite/Obligation **Bond**, **Bulletin** Board, **Pirate** Bay, "**means** business")… |
| 6 | L30·N9680 | +0.0630 | +1.000 | Fires on the modifying/first element of a two-word compound noun phrase, right before the head noun ("speech bubbles", "tick borne", "bean bags", "gre… |
| 7 | L21·N8004 | -0.0526 | +1.000 | Fires at the boundary token that introduces an answer/explanation or list item — the colon after "Ans Sheet", the ")" after a list number, the "is" in… |
| 8 | L30·N5904 | -0.0483 | +1.000 | Fires on the opening token of a highly predictable fixed collocation/idiom (e.g. "means business", "touch base", "way back") and, more weakly, on nume… |
| 9 | L30·N10548 | +0.0426 | +1.000 | Fires on mid-word sub-word fragments that begin a rare/truncated word or proper name — typically a short capitalized "B"/"Br"-type piece (e.g. " b" in… |
| 10 | L31·N4172 | +0.0333 | +1.000 | Fires on the highly predictable continuation token inside a memorized/repeated multi-word phrase — proper-name completions and the next word of an SEO… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L30·N11231 | +0.1540 | +1.000 | Fires on sentence-final periods ending declarative statements (and is strongly suppressed by sentence-final "?" question marks, especially in Q&A inte… |
| 2 | L0·N2312 | -0.1539 | — | |
| 3 | L0·N12080 | +0.1510 | — | |
| 4 | L26·N9501 | +0.1480 | — | |
| 5 | L0·N11229 | +0.1480 | — | |
| 6 | L7·N4953 | -0.1464 | — | |
| 7 | L7·N3917 | +0.1394 | — | |
| 8 | L1·N11305 | +0.1278 | — | |
| 9 | L29·N1118 | -0.1253 | — | |
| 10 | L6·N10760 | -0.1234 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N11229 | +0.1463 | — | |
| 2 | L30·N11231 | +0.1445 | +1.000 | Fires on sentence-final periods ending declarative statements (and is strongly suppressed by sentence-final "?" question marks, especially in Q&A inte… |
| 3 | L26·N9501 | +0.1429 | — | |
| 4 | L0·N2312 | -0.1393 | — | |
| 5 | L7·N3917 | +0.1255 | — | |
| 6 | L1·N5378 | +0.1168 | — | |
| 7 | L6·N10760 | -0.1133 | — | |
| 8 | L0·N5306 | -0.1116 | — | |
| 9 | L5·N5172 | +0.0959 | — | |
| 10 | L7·N4953 | -0.0926 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4172 | +0.0292 | +1.000 | Fires on the highly predictable continuation token inside a memorized/repeated multi-word phrase — proper-name completions and the next word of an SEO… |
| 2 | L29·N7936 | +0.0232 | +1.000 | Fires at positions immediately preceding a lone capital-letter initial/designator (e.g., "A", "S", "D"), such as in personal-name middle initials ("C.… |
| 3 | L1·N1621 | +0.0219 | +1.000 | Fires on the newline token that separates a short header/metadata line (title, citation, date, caption) from the following body text — i.e., a line-br… |
| 4 | L4·N10054 | +0.0214 | +1.000 | Fires on markup/encoding-artifact and delimiter tokens that open a new structural element — mojibake bullets (·), HTML entities (&apos), opening quot… |
| 5 | L4·N892 | +0.0186 | +1.000 | Fires on the standalone opening square-bracket token " [" that begins a bracketed insertion (wiki-style "[ edit ]", "[date]", "[link]"), while being s… |
| 6 | L28·N3888 | +0.0170 | +1.000 | Fires on the auxiliary/copula verb ("are", "am", "were", "would") right after a first‑person pronoun "We"/"I", typically opening an announcement or st… |
| 7 | L9·N10380 | +0.0167 | +3.375 | Mostly fires on the final sub-word piece that completes an uncommon multi-token word (krat|**om**, die|**getic**, void|**ing**), and secondarily on a … |
| 8 | L31·N5043 | +0.0162 | +1.000 | Fires at "start of a new capitalized unit" positions — the newline ending a sentence before a new capitalized sentence/name, or the capitalized token … |
| 9 | L3·N8723 | +0.0162 | +1.000 | Fires on newline tokens (`\n`) that end a heading/title-like or exclamation-terminated line, i.e., a line break where no sentence-final period precede… |
| 10 | L12·N3359 | +0.0161 | +1.000 | Fires on newline (paragraph-break) tokens that follow a completed sentence ending in a period/quote, marking the start of a new paragraph (with weaker… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L30·N1827 | +0.0650 | +1.000 | Fires on the first/modifier token of a two-word named compound (Termite/Obligation **Bond**, **Bulletin** Board, **Pirate** Bay, "**means** business")… |
| 2 | L30·N5904 | +0.0452 | +1.000 | Fires on the opening token of a highly predictable fixed collocation/idiom (e.g. "means business", "touch base", "way back") and, more weakly, on nume… |
| 3 | L31·N4902 | +0.0368 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers (e.g. "pre-configured", "self-align", "Pay-per-Click", "hard-to-rea… |
| 4 | L18·N7449 | +0.0332 | +1.000 | Fires on the short affirmative/negative answer word that opens a reply to a yes/no question (e.g. "Yes", "No", "Exactly", "Not at all"), i.e. the answ… |
| 5 | L28·N10759 | +0.0303 | +1.000 | Fires on the token immediately *preceding* a capital "A" — the punctuation/space before an answer label ("A:", "A."), an alphabet-range listing ("(a-z… |
| 6 | L30·N10237 | +0.0251 | +1.000 | Fires on a token in the *middle* of a multi-part compound string — an interior digit/hyphen of a number (€700,**0**00, A**1**58, 355**-**8050) or a mi… |
| 7 | L11·N8649 | +0.0231 | +1.000 | Fires at clause/sentence-boundary transition points — the comma or period closing an introductory setup phrase, or the connective word right after it … |
| 8 | L11·N6794 | +0.0227 | +1.000 | Fires on the newline token that ends a headline/title or opening tagline, marking the boundary before the body text begins (with some newlines and oth… |
| 9 | L30·N4649 | +0.0217 | +1.000 | Fires at sentence/clause boundaries — almost exclusively on the terminal period (or the newline ending a line/paragraph) right before a new sentence b… |
| 10 | L29·N8536 | +0.0211 | +1.000 | Fires on the head noun of a labeled designation immediately before its identifying letter/number (e.g., "Group B", "Building B", "Semi Final 2", "Adde… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10155 | +3.5788 | — | |
| 2 | L31·N9178 | +3.2189 | — | |
| 3 | L31·N6951 | -2.9544 | — | |
| 4 | L31·N2753 | +2.4720 | — | |
| 5 | L30·N4409 | +2.4054 | — | |
| 6 | L31·N7131 | +2.0862 | — | |
| 7 | L31·N11306 | +1.7482 | — | |
| 8 | L31·N4826 | -1.7395 | — | |
| 9 | L31·N7578 | +1.6676 | — | |
| 10 | L31·N3326 | +1.6506 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L9·N12078 | +9.1264 | -8.126 | Fires in venture-funding and clinical-trial/regulatory announcement language — peaking on the milestone verbs/nouns like "raised", "funding", "investo… |
| 2 | L9·N6299 | +7.8967 | +8.897 | Fires on the token that opens a definitional/explanatory gloss — the "defined as", the "(" beginning a parenthetical explanation, or the head term sta… |
| 3 | L9·N7789 | +7.3538 | +8.354 | Fires on the token that completes a multi-token compound noun or proper name (e.g. "floor→ plan", "City→block", "Pel→ion", "jyot→irling"), i.e. the la… |
| 4 | L9·N5376 | +7.2464 | -6.246 | Fires on the token that completes a multi-token technical term or proper name — the second half of a compound/subword sequence (e.g. "Bern|ers-Lee", "… |
| 5 | L9·N10604 | +7.0848 | +8.085 | Fires on the word that completes a fixed domain-specific compound/collocation — the head noun (or sub-word) of jargon terms in finance, insurance, hyd… |
| 6 | L9·N5476 | +6.9347 | -5.935 | Fires on cognition/speech-act words that introduce a reported opinion or claim — e.g. "think", "answer", "belief", "claimed" — typically right after a… |
| 7 | L9·N9225 | +6.7062 | +7.706 | Fires on the superlative/intensifier word inside hedged-understatement frames ("nothing mind-blowing, but…", "not the greatest story ever told, just…"… |
| 8 | L9·N7655 | +6.5873 | -5.587 | Fires on head nouns naming a consumer product/device category (often the modified product type in phrases like "the Maserati of SUVs", "clear aligners… |
| 9 | L9·N10401 | +6.3970 | -5.397 | Fires on words expressing choice, desire, or preference-selection in a sentence (e.g. "choosy", "wanted", "best", "impulse"), plus the linking preposi… |
| 10 | L9·N7928 | +6.3182 | +7.318 | Fires on the head token of a time/era-naming phrase in origin-story prose ("Boxing **Day**", "the **oldest** American cocktail", "mid 199**0**s", "mod… |
| # | 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 | L9·N662 | -0.0331 | +4.514 | Fires on closing/terminal elements of illustrative example sentences about punctuation and style — sentence-final periods, periods after abbreviations… |
| 2 | L9·N10380 | +0.0167 | +3.375 | Mostly fires on the final sub-word piece that completes an uncommon multi-token word (krat|**om**, die|**getic**, void|**ing**), and secondarily on a … |
| 3 | L9·N6299 | +0.0036 | +8.897 | Fires on the token that opens a definitional/explanatory gloss — the "defined as", the "(" beginning a parenthetical explanation, or the head term sta… |
| 4 | L9·N9225 | -0.0012 | +7.706 | Fires on the superlative/intensifier word inside hedged-understatement frames ("nothing mind-blowing, but…", "not the greatest story ever told, just…"… |
| 5 | L9·N7789 | +0.0009 | +8.354 | Fires on the token that completes a multi-token compound noun or proper name (e.g. "floor→ plan", "City→block", "Pel→ion", "jyot→irling"), i.e. the la… |
| 6 | L9·N7928 | +0.0009 | +7.318 | Fires on the head token of a time/era-naming phrase in origin-story prose ("Boxing **Day**", "the **oldest** American cocktail", "mid 199**0**s", "mod… |
| 7 | L9·N10604 | -0.0008 | +8.085 | Fires on the word that completes a fixed domain-specific compound/collocation — the head noun (or sub-word) of jargon terms in finance, insurance, hyd… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L9·N12078 | +0.0015 | -8.126 | Fires in venture-funding and clinical-trial/regulatory announcement language — peaking on the milestone verbs/nouns like "raised", "funding", "investo… |
| 2 | L9·N7655 | +0.0008 | -5.587 | Fires on head nouns naming a consumer product/device category (often the modified product type in phrases like "the Maserati of SUVs", "clear aligners… |
| 3 | L9·N5476 | +0.0007 | -5.935 | Fires on cognition/speech-act words that introduce a reported opinion or claim — e.g. "think", "answer", "belief", "claimed" — typically right after a… |
| 4 | L9·N4183 | -0.0004 | -5.258 | Fires on names of common over-the-counter drugs/medications, peaking on their word-internal subword pieces (e.g. "oph" in acet-amin-oph-en, "upro" in … |
| 5 | L9·N11193 | -0.0002 | -5.215 | Fires on short function-word tokens in first-/second-person instructional or "discovery" clauses — e.g. the " a" in "we will write a…", the " you" in … |
| 6 | L9·N10401 | -0.0002 | -5.397 | Fires on words expressing choice, desire, or preference-selection in a sentence (e.g. "choosy", "wanted", "best", "impulse"), plus the linking preposi… |
| 7 | L9·N5376 | -0.0001 | -6.246 | Fires on the token that completes a multi-token technical term or proper name — the second half of a compound/subword sequence (e.g. "Bern|ers-Lee", "… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L30·N11231 | +0.1445 | +1.000 | Fires on sentence-final periods ending declarative statements (and is strongly suppressed by sentence-final "?" question marks, especially in Q&A inte… |
| 2 | L6·N10985 | -0.0782 | +1.000 | This neuron fires at sentence/clause boundaries in plain prose — on the newline or the first word right after a sentence-ending period (e.g. " you" af… |
| 3 | L31·N4902 | -0.0731 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers (e.g. "pre-configured", "self-align", "Pay-per-Click", "hard-to-rea… |
| 4 | L29·N8536 | -0.0657 | +1.000 | Fires on the head noun of a labeled designation immediately before its identifying letter/number (e.g., "Group B", "Building B", "Semi Final 2", "Adde… |
| 5 | L30·N1827 | -0.0649 | +1.000 | Fires on the first/modifier token of a two-word named compound (Termite/Obligation **Bond**, **Bulletin** Board, **Pirate** Bay, "**means** business")… |
| 6 | L30·N9680 | +0.0630 | +1.000 | Fires on the modifying/first element of a two-word compound noun phrase, right before the head noun ("speech bubbles", "tick borne", "bean bags", "gre… |
| 7 | L21·N8004 | -0.0526 | +1.000 | Fires at the boundary token that introduces an answer/explanation or list item — the colon after "Ans Sheet", the ")" after a list number, the "is" in… |
| 8 | L30·N5904 | -0.0483 | +1.000 | Fires on the opening token of a highly predictable fixed collocation/idiom (e.g. "means business", "touch base", "way back") and, more weakly, on nume… |
| 9 | L30·N10548 | +0.0426 | +1.000 | Fires on mid-word sub-word fragments that begin a rare/truncated word or proper name — typically a short capitalized "B"/"Br"-type piece (e.g. " b" in… |
| 10 | L31·N4172 | +0.0333 | +1.000 | Fires on the highly predictable continuation token inside a memorized/repeated multi-word phrase — proper-name completions and the next word of an SEO… |
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
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