Config mask_L8_lam3e-5_ep1 · EM 6.22% (terra)
· checkpoint mask_L8_lam3e-5_ep1 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 349 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 | L28, L30, L31 | 0/10 | 10/10 |
| attr_sum | L1, L7, L11, L22, L30, L31 | 0/10 | 4/10 |
| attr_max | L1, L5, L6, L7, L11, L22, L30, L31 | 0/10 | 5/10 |
| top_by_Cplus | L18, L29, L31 | 0/10 | 10/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L26, 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·N1016 | -0.5920 | +1.000 | Fires on the sub-word fragment immediately preceding a final "B" in uppercase acronyms/abbreviations (e.g. C**PC**B, IN**CB**, **MT**B, **SN**B, "15,0… |
| 2 | L31·N7123 | +0.5198 | +1.000 | This neuron is mainly a strong *negative* unit: it is maximally suppressed on word-internal fragments of all-caps / alphanumeric strings whose next su… |
| 3 | L31·N4902 | -0.5037 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token immediately preceding a hyphen (e.g. "pre-", "self-", "X-", "non-"… |
| 4 | L30·N11231 | +0.2565 | +1.000 | Fires on sentence-terminating periods (declarative sentence ends, incl. ")." ), and is strongly suppressed by question marks ending questions (especia… |
| 5 | L30·N10548 | +0.1508 | +1.000 | Fires on unusual word-piece fragments inside proper nouns/abbreviations, typically B-initial or mid-name pieces (" b"drm, " Bin"field, " Bh"agavad, Br… |
| 6 | L30·N9680 | +0.1306 | +1.000 | Fires on a modifier token that sits mid-way through a fixed compound/measure phrase and strongly predicts the head noun that follows ("**speech** bubb… |
| 7 | L31·N7569 | +0.1231 | +1.000 | Fires on the delimiter token immediately preceding an acronym/abbreviated gloss or compound continuation — typically the opening "(" (or quote mark, s… |
| 8 | L28·N10759 | +0.1023 | +1.000 | Fires on the token immediately preceding a capital "A" that starts an answer/label — chiefly the "?" or newline right before an "A:" reply in Q&A tran… |
| 9 | L31·N5053 | -0.0987 | +1.000 | Fires on word-internal continuation sub-word pieces inside capitalized proper names/identifiers in list-like enumerations (e.g. "ys" in Vysionics, "EN… |
| 10 | L31·N10982 | +0.0963 | +1.000 | Fires at clause-final boundary markers — a period, question mark, or newline that closes a term/product name/question — right before an explanatory or… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5920 | +1.000 | Fires on the sub-word fragment immediately preceding a final "B" in uppercase acronyms/abbreviations (e.g. C**PC**B, IN**CB**, **MT**B, **SN**B, "15,0… |
| 2 | L31·N7123 | +0.5198 | +1.000 | This neuron is mainly a strong *negative* unit: it is maximally suppressed on word-internal fragments of all-caps / alphanumeric strings whose next su… |
| 3 | L31·N4902 | -0.5037 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token immediately preceding a hyphen (e.g. "pre-", "self-", "X-", "non-"… |
| 4 | L7·N4953 | -0.4133 | — | |
| 5 | L1·N5378 | -0.3988 | — | |
| 6 | L1·N11305 | +0.2617 | — | |
| 7 | L30·N11231 | +0.2598 | +1.000 | Fires on sentence-terminating periods (declarative sentence ends, incl. ")." ), and is strongly suppressed by question marks ending questions (especia… |
| 8 | L1·N10420 | -0.2351 | — | |
| 9 | L22·N5783 | -0.2347 | — | |
| 10 | L11·N7104 | -0.2265 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5920 | +1.000 | Fires on the sub-word fragment immediately preceding a final "B" in uppercase acronyms/abbreviations (e.g. C**PC**B, IN**CB**, **MT**B, **SN**B, "15,0… |
| 2 | L31·N7123 | +0.5198 | +1.000 | This neuron is mainly a strong *negative* unit: it is maximally suppressed on word-internal fragments of all-caps / alphanumeric strings whose next su… |
| 3 | L31·N4902 | -0.5037 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token immediately preceding a hyphen (e.g. "pre-", "self-", "X-", "non-"… |
| 4 | L7·N4953 | -0.3935 | — | |
| 5 | L1·N5378 | -0.2661 | — | |
| 6 | L30·N11231 | +0.2565 | +1.000 | Fires on sentence-terminating periods (declarative sentence ends, incl. ")." ), and is strongly suppressed by question marks ending questions (especia… |
| 7 | L11·N7104 | -0.2436 | — | |
| 8 | L22·N5783 | -0.2181 | — | |
| 9 | L5·N6908 | -0.2122 | — | |
| 10 | L6·N10760 | -0.2121 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1231 | +1.000 | Fires on the delimiter token immediately preceding an acronym/abbreviated gloss or compound continuation — typically the opening "(" (or quote mark, s… |
| 2 | L31·N10982 | +0.1143 | +1.000 | Fires at clause-final boundary markers — a period, question mark, or newline that closes a term/product name/question — right before an explanatory or… |
| 3 | L31·N4172 | +0.0924 | +1.000 | Fires on the highly predictable continuation token inside a fixed collocation, brand/proper name, or verbatim-repeated headline/title phrase (i.e., th… |
| 4 | L31·N12046 | +0.0517 | +1.000 | Fires on the token immediately preceding a capitalized "B"-initial word/letter (Beonet, Bronze, Benevolent, Bush, B), i.e. it's a next-token predictor… |
| 5 | L31·N2313 | +0.0468 | +1.000 | Fires on the delimiter/boundary token (newline, " (", space, or word-internal split) immediately preceding a capitalized or ALL-CAPS word — i.e., posi… |
| 6 | L29·N8536 | +0.0445 | +1.000 | Fires on the word immediately preceding an alphabetic/alphanumeric designator label (e.g. "Building B", "Group B", "Serie A", "the A158", "Addendum 4-… |
| 7 | L31·N10854 | +0.0375 | +1.000 | Fires on a token sitting inside a highly predictable, fixed English collocation/idiom ("are **to** blame", "on **a** budget", "avoid **over**browning"… |
| 8 | L18·N8199 | +0.0327 | +1.000 | Fires on newline tokens that end a sentence and begin a new paragraph/line in running prose (the line-break itself, not the surrounding words). |
| 9 | L31·N4987 | +0.0299 | +1.000 | Fires on the final content word that completes a sentence/clause right before the period (e.g., "…record it.", "…read it myself.", "…readability has i… |
| 10 | L31·N1706 | +0.0291 | +1.000 | Fires on a mid-phrase token inside a multi-word fixed name/title that is (or will be) repeated verbatim elsewhere in the passage (e.g. "Career **and**… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | +0.5920 | +1.000 | Fires on the sub-word fragment immediately preceding a final "B" in uppercase acronyms/abbreviations (e.g. C**PC**B, IN**CB**, **MT**B, **SN**B, "15,0… |
| 2 | L31·N4902 | +0.5010 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token immediately preceding a hyphen (e.g. "pre-", "self-", "X-", "non-"… |
| 3 | L31·N5053 | +0.0987 | +1.000 | Fires on word-internal continuation sub-word pieces inside capitalized proper names/identifiers in list-like enumerations (e.g. "ys" in Vysionics, "EN… |
| 4 | L30·N10237 | +0.0890 | +1.000 | Fires mid-way through fixed multi-token units (phone-number hyphen, first letter of an acronym/surname, first word of a compound name or date), i.e. a… |
| 5 | L30·N4053 | +0.0828 | +1.000 | Fires on the head word of an alphanumeric designation — the token immediately preceding a letter/number label (Serie **A**, Group **B**, Building **B*… |
| 6 | L31·N8597 | +0.0668 | +1.000 | Fires on tokens continuing an unusual capitalization pattern — mid-string pieces of ALL-CAPS datelines/spelled-out words, acronym letters, and Title-C… |
| 7 | L31·N1720 | +0.0588 | +1.000 | Fires on tokens within ALL-CAPS promotional/headline text (marketing slogans, calls-to-action, "COMING SOON!"), peaking mid-phrase in the capitalized … |
| 8 | L31·N8648 | +0.0576 | +1.000 | Fires inside chatty, personal/craft-blog or product-listing prose — often on the connector token of an enumerating phrase ("as **well** as", "items **… |
| 9 | L29·N4506 | +0.0568 | +1.000 | Fires on digits embedded in long numeric strings (patent numbers, journal volume/page citations, currency amounts) and on the comma that separates a u… |
| 10 | L30·N1827 | +0.0523 | +1.000 | Fires on the interior/linking token inside a multi-word proper name or brand title (the piece that continues an in-progress name, e.g. "Schools **of**… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N8648 | -4.4360 | +1.000 | Fires inside chatty, personal/craft-blog or product-listing prose — often on the connector token of an enumerating phrase ("as **well** as", "items **… |
| 2 | L31·N7131 | +3.9626 | — | |
| 3 | L31·N9527 | +3.2725 | — | |
| 4 | L31·N11306 | +3.0055 | — | |
| 5 | L31·N7027 | +2.9713 | +1.000 | Fires on the locative preposition "in" (or "to") immediately preceding a place name — i.e., the point where a phrase introduces a geographic location;… |
| 6 | L26·N9817 | +2.9191 | — | |
| 7 | L31·N7580 | -2.8290 | — | |
| 8 | L31·N4123 | +2.6537 | — | |
| 9 | L31·N5117 | +2.4873 | — | |
| 10 | L31·N4307 | -2.3591 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N7729 | +5.5669 | +6.567 | Fires on mojibake/non-ASCII byte-encoded tokens — garbled CJK characters, full-width punctuation, and curly/smart quote sequences (plus plain opening … |
| 2 | L8·N2461 | +5.2961 | -4.296 | Fires on the completing element of a fixed compound/collocation — often a rare, capitalized or abbreviated word-initial fragment that finishes a set p… |
| 3 | L8·N9837 | +4.9979 | -3.998 | Fires on tokens marking the winter solstice and its calendar date (the "stice" of "solstice", the "1" in December 21), with weaker firing on parts of … |
| 4 | L8·N8380 | +4.9879 | -3.988 | Fires on names of chemical/cosmetic ingredients or materials in product-composition contexts (silicone/silicones, mineral oil, almond & other oils), i… |
| 5 | L8·N9933 | +4.7191 | +5.719 | Fires on the second/head word that completes a multi-word compound name or technical term (titles, product/proper-name phrases like "Good Man", "sour … |
| 6 | L8·N6034 | +4.5288 | +5.529 | Fires on the token that completes a multi-token compound word or fixed idiom/phrase (e.g. "Water|gate", "w|illy-nilly", "Callio|pe", "bone **to** pick… |
| 7 | L8·N2833 | +4.5206 | +5.521 | Dominated by one very strong case: the hyphenated suffix token "-level" in compound modifiers like "lower-level"; all other firings are weak and scatt… |
| 8 | L8·N4389 | +4.5066 | -3.507 | Fires on word-final subword fragments that complete a noun or surname — typically plural/suffix pieces like "-ets", "-ens", "-ighton", "-ON", "-ien" (… |
| 9 | L8·N1077 | +4.3600 | -3.360 | Diffuse topical detector for NGO / public-health / sustainability-initiative prose (health, preservation, recycled, environment, conventions), with th… |
| 10 | L8·N1408 | +4.3322 | +5.332 | Fires on words denoting girth/bulk/fit of a body or garment in clothing- and body-shape contexts (width, slim, bulky, stuff, expect|ant), peaking on t… |
| # | 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·N9933 | +0.0017 | +5.719 | Fires on the second/head word that completes a multi-word compound name or technical term (titles, product/proper-name phrases like "Good Man", "sour … |
| 2 | L8·N2833 | +0.0007 | +5.521 | Dominated by one very strong case: the hyphenated suffix token "-level" in compound modifiers like "lower-level"; all other firings are weak and scatt… |
| 3 | L8·N7729 | +0.0004 | +6.567 | Fires on mojibake/non-ASCII byte-encoded tokens — garbled CJK characters, full-width punctuation, and curly/smart quote sequences (plus plain opening … |
| 4 | L8·N11535 | +0.0003 | +5.159 | Fires on tokens in social-services / behavioral-health provider language — words like "social", "service(s)", agency acronyms (DSS) and clinical licen… |
| 5 | L8·N1408 | -0.0001 | +5.332 | Fires on words denoting girth/bulk/fit of a body or garment in clothing- and body-shape contexts (width, slim, bulky, stuff, expect|ant), peaking on t… |
| 6 | L8·N7278 | -0.0001 | +5.207 | Primarily a negatively-firing neuron: it is strongly suppressed by tokens describing small/closely-held/private business entities ("small company", "c… |
| 7 | L8·N6034 | -0.0000 | +5.529 | Fires on the token that completes a multi-token compound word or fixed idiom/phrase (e.g. "Water|gate", "w|illy-nilly", "Callio|pe", "bone **to** pick… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N2461 | +0.0116 | -4.296 | Fires on the completing element of a fixed compound/collocation — often a rare, capitalized or abbreviated word-initial fragment that finishes a set p… |
| 2 | L8·N9837 | -0.0023 | -3.998 | Fires on tokens marking the winter solstice and its calendar date (the "stice" of "solstice", the "1" in December 21), with weaker firing on parts of … |
| 3 | L8·N4389 | -0.0018 | -3.507 | Fires on word-final subword fragments that complete a noun or surname — typically plural/suffix pieces like "-ets", "-ens", "-ighton", "-ON", "-ien" (… |
| 4 | L8·N1077 | +0.0013 | -3.360 | Diffuse topical detector for NGO / public-health / sustainability-initiative prose (health, preservation, recycled, environment, conventions), with th… |
| 5 | L8·N8380 | -0.0005 | -3.988 | Fires on names of chemical/cosmetic ingredients or materials in product-composition contexts (silicone/silicones, mineral oil, almond & other oils), i… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5920 | +1.000 | Fires on the sub-word fragment immediately preceding a final "B" in uppercase acronyms/abbreviations (e.g. C**PC**B, IN**CB**, **MT**B, **SN**B, "15,0… |
| 2 | L31·N7123 | +0.5198 | +1.000 | This neuron is mainly a strong *negative* unit: it is maximally suppressed on word-internal fragments of all-caps / alphanumeric strings whose next su… |
| 3 | L31·N4902 | -0.5037 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token immediately preceding a hyphen (e.g. "pre-", "self-", "X-", "non-"… |
| 4 | L30·N11231 | +0.2565 | +1.000 | Fires on sentence-terminating periods (declarative sentence ends, incl. ")." ), and is strongly suppressed by question marks ending questions (especia… |
| 5 | L30·N10548 | +0.1508 | +1.000 | Fires on unusual word-piece fragments inside proper nouns/abbreviations, typically B-initial or mid-name pieces (" b"drm, " Bin"field, " Bh"agavad, Br… |
| 6 | L30·N9680 | +0.1306 | +1.000 | Fires on a modifier token that sits mid-way through a fixed compound/measure phrase and strongly predicts the head noun that follows ("**speech** bubb… |
| 7 | L31·N7569 | +0.1231 | +1.000 | Fires on the delimiter token immediately preceding an acronym/abbreviated gloss or compound continuation — typically the opening "(" (or quote mark, s… |
| 8 | L28·N10759 | +0.1023 | +1.000 | Fires on the token immediately preceding a capital "A" that starts an answer/label — chiefly the "?" or newline right before an "A:" reply in Q&A tran… |
| 9 | L31·N5053 | -0.0987 | +1.000 | Fires on word-internal continuation sub-word pieces inside capitalized proper names/identifiers in list-like enumerations (e.g. "ys" in Vysionics, "EN… |
| 10 | L31·N10982 | +0.0963 | +1.000 | Fires at clause-final boundary markers — a period, question mark, or newline that closes a term/product name/question — right before an explanatory or… |
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