Config mask_Lall_lam3e-5_ep10 · EM 1.88% (terra)
· checkpoint mask_Lall_lam3e-5_ep10 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 158 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, L1, L2, L3, L4, L5, L6, L7, L8, L9, L10, L11, L12, L13, L14, L15, L16, L17, L18, L19, L20, L21, L22, L23, L24, L25, L26, L27, L28, L29, L30, L31) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L28, L30, L31 | 10/10 | 8/10 |
| attr_sum | L0, L13, L22, L28, L30, L31 | 10/10 | 9/10 |
| attr_max | L0, L22, L28, L30, L31 | 10/10 | 9/10 |
| top_by_Cplus | L19, L21, L22, L28, L29, L31 | 10/10 | 6/10 |
| top_by_Cminus | L25, L30, L31 | 10/10 | 7/10 |
| empos_contrast | L28, L29, L30, L31 | 10/10 | 2/10 |
| mask_delta | L11, L20, L26, L28, L29, L31 | 10/10 | 0/10 |
| random | L1, L2, L8, L9, L13, L17, L22, L29 | 10/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -1.0224 | +1.046 | Fires on the first component of hyphenated compounds/prefixed words — i.e., a token immediately preceding a hyphen in constructions like "pre-configur… |
| 2 | L31·N1016 | -0.8172 | +1.130 | Fires on the first sub-word fragment of a capitalized acronym/brand name that is split |
| 3 | L31·N7123 | +0.2836 | +0.487 | This is essentially a negative/suppression unit: its strongest (most negative) responses come on the token immediately preceding a standalone capital … |
| 4 | L31·N10982 | +0.2436 | +0.990 | Fires at end-of-segment boundary tokens — mostly the |
| 5 | L30·N10548 | +0.1974 | +1.130 | Fires on a token that is the first fragment of a multi-token abbreviation/alphanumeric code or compound (e.g. the "M" of M25, the "-M" of "-MN", "BR" … |
| 6 | L31·N7569 | +0.1724 | +0.984 | Fires on the opening delimiter ( "(", " |
| 7 | L31·N8597 | -0.1587 | +1.805 | Fires on tokens inside technical/product-spec or listing strings — model/spec fragments like "Hot Dipped", "480p", "(highway) mpg", "Game-play", initi… |
| 8 | L28·N11638 | +0.1573 | +1.439 | Fires on the token right before a capital "B" (or "b") in enumerations/abbreviations — i.e., it predicts an upcoming "B" after an "A"/"a)"/"G:"/"T." c… |
| 9 | L30·N5904 | +0.1555 | +0.888 | Fires on highly predictable "filler" tokens that complete a fixed collocation or numeric range/duplicated phrase — e.g., the second number in "1-2"/"o… |
| 10 | L31·N1720 | -0.1363 | +1.844 | F |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -1.0224 | +1.046 | Fires on the first component of hyphenated compounds/prefixed words — i.e., a token immediately preceding a hyphen in constructions like "pre-configur… |
| 2 | L31·N1016 | -0.8172 | +1.130 | Fires on the first sub-word fragment of a capitalized acronym/brand name that is split |
| 3 | L0·N956 | -0.3930 | — | |
| 4 | L28·N11638 | +0.3260 | +1.439 | Fires on the token right before a capital "B" (or "b") in enumerations/abbreviations — i.e., it predicts an upcoming "B" after an "A"/"a)"/"G:"/"T." c… |
| 5 | L31·N7123 | +0.2836 | +0.487 | This is essentially a negative/suppression unit: its strongest (most negative) responses come on the token immediately preceding a standalone capital … |
| 6 | L0·N2312 | -0.2531 | — | |
| 7 | L30·N10548 | +0.2447 | +1.130 | Fires on a token that is the first fragment of a multi-token abbreviation/alphanumeric code or compound (e.g. the "M" of M25, the "-M" of "-MN", "BR" … |
| 8 | L31·N10982 | +0.2436 | +0.990 | Fires at end-of-segment boundary tokens — mostly the |
| 9 | L13·N4272 | -0.2351 | — | |
| 10 | L22·N5783 | -0.1878 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -1.0224 | +1.046 | Fires on the first component of hyphenated compounds/prefixed words — i.e., a token immediately preceding a hyphen in constructions like "pre-configur… |
| 2 | L31·N1016 | -0.8172 | +1.130 | Fires on the first sub-word fragment of a capitalized acronym/brand name that is split |
| 3 | L0·N956 | -0.3369 | — | |
| 4 | L31·N7123 | +0.2836 | +0.487 | This is essentially a negative/suppression unit: its strongest (most negative) responses come on the token immediately preceding a standalone capital … |
| 5 | L0·N2312 | -0.2649 | — | |
| 6 | L31·N10982 | +0.2436 | +0.990 | Fires at end-of-segment boundary tokens — mostly the |
| 7 | L30·N10548 | +0.1973 | +1.130 | Fires on a token that is the first fragment of a multi-token abbreviation/alphanumeric code or compound (e.g. the "M" of M25, the "-M" of "-MN", "BR" … |
| 8 | L22·N5783 | -0.1860 | — | |
| 9 | L31·N7569 | +0.1724 | +0.984 | Fires on the opening delimiter ( "(", " |
| 10 | L28·N11638 | +0.1599 | +1.439 | Fires on the token right before a capital "B" (or "b") in enumerations/abbreviations — i.e., it predicts an upcoming "B" after an "A"/"a)"/"G:"/"T." c… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.2473 | +0.990 | Fires at end-of-segment boundary tokens — mostly the |
| 2 | L31·N7569 | +0.1724 | +0.984 | Fires on the opening delimiter ( "(", " |
| 3 | L28·N11638 | +0.1573 | +1.439 | Fires on the token right before a capital "B" (or "b") in enumerations/abbreviations — i.e., it predicts an upcoming "B" after an "A"/"a)"/"G:"/"T." c… |
| 4 | L31·N4172 | +0.1103 | +1.035 | Fires on a token that continues/completes a phrase already seen earlier in the passage — i.e., inside duplicated, repetitive boilerplate or reduplicat… |
| 5 | L31·N10854 | +0.0725 | +1.099 | This neuron fires on tokens that re-instantiate material already seen earlier in the passage — the "re" of "re-built"/"rebranding", the repeated "your… |
| 6 | L29·N1118 | +0.0622 | +0.979 | This neuron shows no crisp lexical trigger — it fires diffusely (mid-range) across almost every |
| 7 | L19·N1211 | +0.0415 | +1.903 | Fires on the first token of a new text segment right at a run‑on/concatenation seam in scraped web text (missing space or newline joining two chunks, … |
| 8 | L22·N2979 | +0.0404 | -0.620 | Fires on the token that starts the **second, parallel repetition** of a just-used phrase/construction (e.g. "on the face |
| 9 | L21·N8004 | +0.0380 | +1.259 | Fires on the period token immediately following a leading numeral "1" at the start of an enumerated list item ("1." beginning a snippet/line); much we… |
| 10 | L19·N10246 | +0.0364 | +0.887 | Fires on the linking token that joins two par |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +1.0223 | +1.046 | Fires on the first component of hyphenated compounds/prefixed words — i.e., a token immediately preceding a hyphen in constructions like "pre-configur… |
| 2 | L31·N1016 | +0.8172 | +1.130 | Fires on the first sub-word fragment of a capitalized acronym/brand name that is split |
| 3 | L31·N8597 | +0.1587 | +1.805 | Fires on tokens inside technical/product-spec or listing strings — model/spec fragments like "Hot Dipped", "480p", "(highway) mpg", "Game-play", initi… |
| 4 | L31·N1720 | +0.1363 | +1.844 | F |
| 5 | L31·N9287 | +0.0769 | +1.120 | Fires inside ALL-CAPS or Title-Case headline/promotional phrases (marketing banners, product listings, proper-name titles), peaking on a continuation … |
| 6 | L30·N6369 | +0.0669 | +1.054 | Fires on mid-word ALL-CAPS subword frag |
| 7 | L31·N8648 | +0.0606 | +1.101 | Fires on tokens inside identifier-like strings — individual digits within long numbers (addresses, patent/phone/UR |
| 8 | L31·N5053 | +0.0604 | +1.009 | The clearest, most consistent signal is on the negative side: the neuron is strongly suppressed by the space |
| 9 | L31·N5228 | +0.0571 | +1.738 | Fires on a mid-word sub-token that is immediately followed by a fragment beginning with "a" (‑ap, ‑an, ‑ad), i.e. the split point in rare/compound wor… |
| 10 | L25·N1677 | +0.0557 | +1.280 | This neuron's dominant (strongly negative) response is to the letter "a" functioning as an alphabetic label/range marker — "a)" vs "b)", "a-z", the su… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N8769 | -4.3236 | — | |
| 2 | L31·N2753 | +4.3016 | — | |
| 3 | L31·N11306 | +4.1133 | — | |
| 4 | L28·N11638 | +3.9993 | +1.439 | Fires on the token right before a capital "B" (or "b") in enumerations/abbreviations — i.e., it predicts an upcoming "B" after an "A"/"a)"/"G:"/"T." c… |
| 5 | L31·N9527 | -3.9867 | — | |
| 6 | L30·N10548 | -3.3830 | +1.130 | Fires on a token that is the first fragment of a multi-token abbreviation/alphanumeric code or compound (e.g. the "M" of M25, the "-M" of "-MN", "BR" … |
| 7 | L31·N3114 | +3.2457 | — | |
| 8 | L29·N6318 | -2.9986 | — | |
| 9 | L31·N8693 | -2.9143 | — | |
| 10 | L31·N7025 | +2.8552 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L20·N5793 | +7.2027 | +8.203 | Fires on tokens inside formal institutional/affiliation and contact-detail boilerplate — university/faculty/organisation names, job-title "Head of Ope… |
| 2 | L26·N7220 | +6.6443 | -5.644 | Fires on sentence-terminating punctuation — mainly the period (and question mark) closing a long descriptive clause, plus isolated/repeated commas — i… |
| 3 | L29·N259 | +6.4458 | +7.446 | Fires on the definite article "the" when it heads a fixed comparative/contrastive noun phrase — i.e., "the" immediately preceding words like *opposite… |
| 4 | L31·N7477 | +6.4431 | -5.443 | Broadly active across European/translated-English informational-promotional prose (EU summits, clinics, thermal baths, city development), peaking on t… |
| 5 | L29·N6887 | +6.3360 | -5.336 | Fires on sub-word t |
| 6 | L31·N5645 | +6.3236 | +7.324 | Fires on grammatical glue tokens (the, a, and, 's, commas) that immediately precede or sit inside medical/clinical noun phrases (medical bills, medica… |
| 7 | L11·N9064 | +6.2926 | -5.293 | Fires on the very first token of a document/text chunk — especially an opening quotation mark starting a quoted passage ("The, "A, “When) or other doc… |
| 8 | L29·N6065 | +6.2192 | -5.219 | Fires on French-language tokens embedded in otherwise English/mixed text — French function words and morphemes (ne, été, vous), and French proper name… |
| 9 | L29·N10880 | +6.1978 | -5.198 | Fires on tokens inside Arabic/Mu |
| 10 | L28·N3345 | +6.1886 | +7.189 | Fires on the opening token of a predictable multi-token proper-name/phrase — name-initial subword pieces and first names (e.g. " Mc" of McEwan, " Mar"… |
| # | 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 | L31·N1016 | -0.8172 | +1.130 | Fires on the first sub-word fragment of a capitalized acronym/brand name that is split |
| 2 | L30·N10548 | +0.1974 | +1.130 | Fires on a token that is the first fragment of a multi-token abbreviation/alphanumeric code or compound (e.g. the "M" of M25, the "-M" of "-MN", "BR" … |
| 3 | L31·N8597 | -0.1587 | +1.805 | Fires on tokens inside technical/product-spec or listing strings — model/spec fragments like "Hot Dipped", "480p", "(highway) mpg", "Game-play", initi… |
| 4 | L28·N11638 | +0.1573 | +1.439 | Fires on the token right before a capital "B" (or "b") in enumerations/abbreviations — i.e., it predicts an upcoming "B" after an "A"/"a)"/"G:"/"T." c… |
| 5 | L31·N1720 | -0.1363 | +1.844 | F |
| 6 | L31·N9287 | -0.0752 | +1.120 | Fires inside ALL-CAPS or Title-Case headline/promotional phrases (marketing banners, product listings, proper-name titles), peaking on a continuation … |
| 7 | L31·N10854 | +0.0725 | +1.099 | This neuron fires on tokens that re-instantiate material already seen earlier in the passage — the "re" of "re-built"/"rebranding", the repeated "your… |
| 8 | L30·N6369 | -0.0669 | +1.054 | Fires on mid-word ALL-CAPS subword frag |
| 9 | L31·N8648 | -0.0606 | +1.101 | Fires on tokens inside identifier-like strings — individual digits within long numbers (addresses, patent/phone/UR |
| 10 | L31·N5228 | -0.0571 | +1.738 | Fires on a mid-word sub-token that is immediately followed by a fragment beginning with "a" (‑ap, ‑an, ‑ad), i.e. the split point in rare/compound wor… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7123 | +0.2836 | +0.487 | This is essentially a negative/suppression unit: its strongest (most negative) responses come on the token immediately preceding a standalone capital … |
| 2 | L30·N5904 | +0.1555 | +0.888 | Fires on highly predictable "filler" tokens that complete a fixed collocation or numeric range/duplicated phrase — e.g., the second number in "1-2"/"o… |
| 3 | L28·N6624 | +0.1210 | +0.722 | Fires on the token that **completes a multi-token unit** — the trailing "0" of a number like 350/300, the final word-piece of a fragmented name or ter… |
| 4 | L31·N4989 | -0.0497 | +0.899 | Fires on the token that completes a strongly context-determined unit — a word-final subtoken or the second half of a repeated/collocational phrase (ci… |
| 5 | L19·N10246 | +0.0366 | +0.887 | Fires on the linking token that joins two par |
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L22·N2979 | +0.0404 | -0.620 | Fires on the token that starts the **second, parallel repetition** of a just-used phrase/construction (e.g. "on the face |
| 2 | L29·N10880 | -0.0047 | -5.198 | Fires on tokens inside Arabic/Mu |
| 3 | L29·N8069 | -0.0015 | -5.125 | Fires on the |
| 4 | L11·N9064 | -0.0010 | -5.293 | Fires on the very first token of a document/text chunk — especially an opening quotation mark starting a quoted passage ("The, "A, “When) or other doc… |
| 5 | L26·N7220 | -0.0008 | -5.644 | Fires on sentence-terminating punctuation — mainly the period (and question mark) closing a long descriptive clause, plus isolated/repeated commas — i… |
| 6 | L29·N741 | -0.0005 | -5.160 | Fires on ordinary function words ("to", "that", "will", "any") embedded in awkward machine-translated / synonym-spun SE |
| 7 | L29·N6065 | -0.0004 | -5.219 | Fires on French-language tokens embedded in otherwise English/mixed text — French function words and morphemes (ne, été, vous), and French proper name… |
| 8 | L29·N6887 | +0.0003 | -5.336 | Fires on sub-word t |
| 9 | L31·N7477 | -0.0001 | -5.443 | Broadly active across European/translated-English informational-promotional prose (EU summits, clinics, thermal baths, city development), peaking on t… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -1.0224 | +1.046 | Fires on the first component of hyphenated compounds/prefixed words — i.e., a token immediately preceding a hyphen in constructions like "pre-configur… |
| 2 | L31·N10982 | +0.2436 | +0.990 | Fires at end-of-segment boundary tokens — mostly the |
| 3 | L31·N7569 | +0.1724 | +0.984 | Fires on the opening delimiter ( "(", " |
| 4 | L31·N4172 | +0.1103 | +1.035 | Fires on a token that continues/completes a phrase already seen earlier in the passage — i.e., inside duplicated, repetitive boilerplate or reduplicat… |
| 5 | L29·N1118 | +0.0622 | +0.979 | This neuron shows no crisp lexical trigger — it fires diffusely (mid-range) across almost every |
| 6 | L31·N5053 | -0.0604 | +1.009 | The clearest, most consistent signal is on the negative side: the neuron is strongly suppressed by the space |
| 7 | L28·N11120 | +0.0349 | +1.042 | Fires on the first character(s) of an opaque alphanumeric code/identifier immediately after its prefix — e.g. the "#" starting a hex color, the random… |
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