Config mask_L4_lam3e-5_ep6 · EM 10.28% (terra)
· checkpoint mask_L4_lam3e-5_ep6 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 569 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, L7, L22, L31 | 0/10 | 3/10 |
| attr_max | L0, L1, L7, L31 | 0/10 | 3/10 |
| top_by_Cplus | L18, L19, L27, L28, L30, L31 | 0/10 | 7/10 |
| top_by_Cminus | L22, L29, L30, L31 | 0/10 | 8/10 |
| empos_contrast | L28, L29, L30, L31 | 0/10 | 6/10 |
| 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 | -0.4436 | +1.000 | Fires on the token immediately preceding a hyphen in compound/prefixed constructions (prefixes and compound-modifier first elements like "pre-", "self… |
| 2 | L31·N1016 | -0.2377 | +1.000 | Fires on the first sub-token piece of a multi-token uppercase acronym/abbreviation that is continued by further sub-tokens (e.g. " NW|OT", " SN|B", " … |
| 3 | L31·N7123 | +0.1728 | +1.000 | Fires on the first token of a multi-token proper name or fixed compound whose completion follows (e.g. "Broken"→Arrow, "Polis"→ario, "Shik"→ara, "Sout… |
| 4 | L30·N10548 | +0.1140 | +1.000 | Fires on word-initial "b"/"B" sub-word fragments that begin an unusual word or proper name continued by following sub-tokens (e.g. " b"+drm, "Tap"+"ab… |
| 5 | L30·N11231 | +0.1041 | +1.000 | Fires on sentence-final periods ending declarative statements in expository prose (and is strongly suppressed by question marks ending questions, e.g.… |
| 6 | L31·N7569 | +0.0949 | +1.000 | Fires at "copy/spell-out" boundary positions — most strongly on the space-plus-open-parenthesis that introduces an acronym or its expansion (e.g. "…Po… |
| 7 | L30·N9680 | +0.0799 | +1.000 | Fires on the token immediately **preceding a word beginning with "b"** in compound/collocation contexts (speech→bubbles, 2→bath, tick→borne, grey→bark… |
| 8 | L31·N10982 | -0.0567 | +1.000 | Fires on the boundary token that closes a phrase/question—usually the final period, question mark, or newline (and occasionally the last sub-word of a… |
| 9 | L30·N4053 | -0.0555 | +1.000 | Fires on the token immediately preceding a single-letter or letter+number designator/class label (e.g. "Serie" before A, "Building" before B, "A" befo… |
| 10 | L31·N4172 | +0.0538 | +1.000 | This neuron fires on tokens that are highly predictable continuations of a repeated or fixed multi-word string — e.g. mid-phrase tokens of a headline/… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4436 | +1.000 | Fires on the token immediately preceding a hyphen in compound/prefixed constructions (prefixes and compound-modifier first elements like "pre-", "self… |
| 2 | L0·N11229 | +0.4329 | — | |
| 3 | L7·N7796 | -0.4114 | — | |
| 4 | L1·N10420 | +0.4113 | — | |
| 5 | L31·N1016 | -0.2377 | +1.000 | Fires on the first sub-token piece of a multi-token uppercase acronym/abbreviation that is continued by further sub-tokens (e.g. " NW|OT", " SN|B", " … |
| 6 | L0·N12080 | +0.2150 | — | |
| 7 | L1·N1621 | +0.1949 | — | |
| 8 | L31·N7123 | +0.1728 | +1.000 | Fires on the first token of a multi-token proper name or fixed compound whose completion follows (e.g. "Broken"→Arrow, "Polis"→ario, "Shik"→ara, "Sout… |
| 9 | L0·N4031 | -0.1627 | — | |
| 10 | L22·N2979 | -0.1622 | +1.000 | Fires at positions introducing or inside numeric/date expressions — especially the preposition " on" right before a calendar date and " to" between tw… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4436 | +1.000 | Fires on the token immediately preceding a hyphen in compound/prefixed constructions (prefixes and compound-modifier first elements like "pre-", "self… |
| 2 | L0·N11229 | +0.4227 | — | |
| 3 | L1·N10420 | +0.3975 | — | |
| 4 | L7·N7796 | -0.3930 | — | |
| 5 | L0·N12080 | +0.2555 | — | |
| 6 | L31·N1016 | -0.2377 | +1.000 | Fires on the first sub-token piece of a multi-token uppercase acronym/abbreviation that is continued by further sub-tokens (e.g. " NW|OT", " SN|B", " … |
| 7 | L1·N5378 | +0.2284 | — | |
| 8 | L1·N1621 | +0.2011 | — | |
| 9 | L31·N7123 | +0.1728 | +1.000 | Fires on the first token of a multi-token proper name or fixed compound whose completion follows (e.g. "Broken"→Arrow, "Polis"→ario, "Shik"→ara, "Sout… |
| 10 | L0·N4031 | -0.1594 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.0949 | +1.000 | Fires at "copy/spell-out" boundary positions — most strongly on the space-plus-open-parenthesis that introduces an acronym or its expansion (e.g. "…Po… |
| 2 | L31·N4172 | +0.0538 | +1.000 | This neuron fires on tokens that are highly predictable continuations of a repeated or fixed multi-word string — e.g. mid-phrase tokens of a headline/… |
| 3 | L31·N2313 | +0.0257 | +1.000 | Fires on delimiter/boundary tokens — a newline, "(", ")\", or space — that sit immediately before a capitalized heading, proper name, or label (e.g. b… |
| 4 | L31·N4764 | +0.0245 | +1.000 | Fires on the mojibake opening curly double-quote token "âĢľ" (a “ ), almost always immediately after the indefinite article "a" introducing a quoted p… |
| 5 | L19·N1211 | +0.0220 | +1.000 | Fires on tokens embedded in keyword-stuffed, repetitive catalog/SEO-style listing text — product/title words and their delimiters in scraped listing b… |
| 6 | L31·N12046 | +0.0217 | +1.000 | Fires mid-way through fixed multi-token compounds/abbreviations where the remainder is highly predictable — e.g. the split-word fragment ("trom"+bone,… |
| 7 | L18·N8199 | +0.0211 | +1.000 | Fires on the newline token at sentence/line boundaries (right after a sentence-final period, quote, or question mark), marking the break between conse… |
| 8 | L28·N11638 | +0.0207 | +1.000 | Fires on the connector/punctuation token immediately preceding a "B" label in an A→B style enumeration (R/G/**B**, A²+**B**², a)…and b), vit**amin B6*… |
| 9 | L27·N7908 | +0.0198 | +1.000 | Fires at the token immediately preceding a numeric date/time value in posting metadata (e.g. the space or "at"/":" before "5:44 PM", "4:13 am", "2 min… |
| 10 | L30·N11531 | +0.0188 | +1.000 | Fires on the leading digit (or the space immediately before it) of multi-digit numbers in measurements, dates, phone numbers, and model designations; … |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.4423 | +1.000 | Fires on the token immediately preceding a hyphen in compound/prefixed constructions (prefixes and compound-modifier first elements like "pre-", "self… |
| 2 | L31·N1016 | +0.2377 | +1.000 | Fires on the first sub-token piece of a multi-token uppercase acronym/abbreviation that is continued by further sub-tokens (e.g. " NW|OT", " SN|B", " … |
| 3 | L30·N4053 | +0.0555 | +1.000 | Fires on the token immediately preceding a single-letter or letter+number designator/class label (e.g. "Serie" before A, "Building" before B, "A" befo… |
| 4 | L29·N4506 | +0.0492 | +1.000 | Fires on tokens inside compact alphanumeric identifier/reference strings — interior digits of patent numbers, citation issue numbers and currency conv… |
| 5 | L22·N2979 | +0.0468 | +1.000 | Fires at positions introducing or inside numeric/date expressions — especially the preposition " on" right before a calendar date and " to" between tw… |
| 6 | L31·N8597 | +0.0465 | +1.000 | Fires on continuation tokens inside stretches of unusual, non-prose formatting — ALL-CAPS listings/datelines/headings, Title-Cased boilerplate, spec s… |
| 7 | L31·N5053 | +0.0443 | +1.000 | Fires on tokens that *continue/complete* an already-begun word or fixed phrase (subword continuations like GOLD→"EN", "Ill…ness", the second half of c… |
| 8 | L30·N10237 | +0.0406 | +1.000 | Fires on the first sub-token of a highly predictable multi-token unit that the context all but forces — e.g. the "2" beginning a year, the initial of … |
| 9 | L31·N1720 | +0.0401 | +1.000 | Fires on words within runs of ALL-CAPS or Title-Case-Every-Word text (marketing/SEO headings, shouted announcements), rather than on normal sentence-c… |
| 10 | L31·N7027 | +0.0390 | +1.000 | Fires on the locative prepositions "in"/"to" when they introduce a place or location (state, city, region, room) after a noun phrase — e.g. "adoption … |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N9527 | +1.8720 | — | |
| 2 | L28·N1213 | -1.3902 | — | |
| 3 | L29·N4873 | +1.1925 | — | |
| 4 | L28·N10759 | +1.1684 | — | |
| 5 | L31·N11467 | +1.1327 | — | |
| 6 | L31·N10263 | -1.0754 | — | |
| 7 | L31·N1417 | -1.0479 | — | |
| 8 | L30·N5904 | +0.9799 | — | |
| 9 | L31·N3326 | -0.9787 | — | |
| 10 | L31·N7025 | -0.9570 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N19 | +11.5618 | -10.562 | Fires on abstract nouns of favorable attitude/demand — "interest", "sentiment", "enthusiasm", "support" — used as a mass noun heading a prepositional … |
| 2 | L4·N5531 | +10.8116 | +11.812 | Fires on the continuation token that completes a fixed legal/formal collocation or compound — most reliably " extent" in "to the extent" and the "grav… |
| 3 | L4·N1379 | +10.6392 | -9.639 | Fires on newline tokens, specifically the line break immediately after a sentence/fragment ending in a period, within blocks of consecutive short para… |
| 4 | L4·N12001 | +10.5068 | +11.507 | Fires on mid-word sub-token fragments inside rare proper nouns and technical compounds (the continuation piece of a name/word split by the tokenizer),… |
| 5 | L4·N4710 | +10.5034 | +11.503 | Fires in second-person advice/how-to text on the verb (or the "to" of an infinitive) completing a cautionary construction warning against over-attempt… |
| 6 | L4·N7720 | +10.0410 | -9.041 | Fires on the token immediately following an expression of necessity/obligation ("I needed **my**", "have/had **to**", "much**-needed**", "Should we … … |
| 7 | L4·N2276 | +9.7743 | -8.774 | This neuron is driven by degree/manner adverb slots before an adjective or clause: it is strongly *suppressed* by hedging intensifiers ("quite", "fair… |
| 8 | L4·N9716 | +9.1980 | +10.198 | Fires on the finite main verb that closes a long subject noun phrase (especially reporting/attribution verbs like "said", "described", "explains", "no… |
| 9 | L4·N5620 | +9.1934 | +10.193 | Fires on the continuation/completion sub-token of a multi-token compound name or word, especially British place/proper names ("St Mary", "St Lawrence"… |
| 10 | L4·N8906 | +9.1237 | -8.124 | Mostly a "negative" (suppression) neuron for sameness/uniformity morphemes — it is driven strongly negative on stems like "einheit-", "uniform-", "hom… |
| # | 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 | +0.0078 | +11.507 | Fires on mid-word sub-token fragments inside rare proper nouns and technical compounds (the continuation piece of a name/word split by the tokenizer),… |
| 2 | L4·N5620 | -0.0008 | +10.193 | Fires on the continuation/completion sub-token of a multi-token compound name or word, especially British place/proper names ("St Mary", "St Lawrence"… |
| 3 | L4·N5567 | +0.0002 | +10.013 | Fires mainly on paragraph-break newline tokens that follow a completed sentence (section boundaries), with weaker firing on the digit "5" completing a… |
| 4 | L4·N9716 | +0.0001 | +10.198 | Fires on the finite main verb that closes a long subject noun phrase (especially reporting/attribution verbs like "said", "described", "explains", "no… |
| 5 | L4·N5531 | -0.0001 | +11.812 | Fires on the continuation token that completes a fixed legal/formal collocation or compound — most reliably " extent" in "to the extent" and the "grav… |
| 6 | L4·N4710 | -0.0001 | +11.503 | Fires in second-person advice/how-to text on the verb (or the "to" of an infinitive) completing a cautionary construction warning against over-attempt… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N11717 | -0.0371 | -3.331 | Fires at message-opening boundaries — the punctuation or word right after a greeting/salutation or header that starts a new message (the comma in "Hi,… |
| 2 | L4·N7720 | +0.0021 | -9.041 | Fires on the token immediately following an expression of necessity/obligation ("I needed **my**", "have/had **to**", "much**-needed**", "Should we … … |
| 3 | L4·N8261 | -0.0019 | -7.978 | Fires on the completing word of a common collocation/compound in ordinary prose — e.g. " up" in "bring up", "guard" in "Coastguard", " plan" in "a pla… |
| 4 | L4·N1379 | +0.0002 | -9.639 | Fires on newline tokens, specifically the line break immediately after a sentence/fragment ending in a period, within blocks of consecutive short para… |
| 5 | L4·N8906 | +0.0002 | -8.124 | Mostly a "negative" (suppression) neuron for sameness/uniformity morphemes — it is driven strongly negative on stems like "einheit-", "uniform-", "hom… |
| 6 | L4·N2276 | -0.0001 | -8.774 | This neuron is driven by degree/manner adverb slots before an adjective or clause: it is strongly *suppressed* by hedging intensifiers ("quite", "fair… |
| 7 | L4·N19 | +0.0000 | -10.562 | Fires on abstract nouns of favorable attitude/demand — "interest", "sentiment", "enthusiasm", "support" — used as a mass noun heading a prepositional … |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4436 | +1.000 | Fires on the token immediately preceding a hyphen in compound/prefixed constructions (prefixes and compound-modifier first elements like "pre-", "self… |
| 2 | L31·N1016 | -0.2377 | +1.000 | Fires on the first sub-token piece of a multi-token uppercase acronym/abbreviation that is continued by further sub-tokens (e.g. " NW|OT", " SN|B", " … |
| 3 | L31·N7123 | +0.1728 | +1.000 | Fires on the first token of a multi-token proper name or fixed compound whose completion follows (e.g. "Broken"→Arrow, "Polis"→ario, "Shik"→ara, "Sout… |
| 4 | L30·N10548 | +0.1140 | +1.000 | Fires on word-initial "b"/"B" sub-word fragments that begin an unusual word or proper name continued by following sub-tokens (e.g. " b"+drm, "Tap"+"ab… |
| 5 | L30·N11231 | +0.1041 | +1.000 | Fires on sentence-final periods ending declarative statements in expository prose (and is strongly suppressed by question marks ending questions, e.g.… |
| 6 | L31·N7569 | +0.0949 | +1.000 | Fires at "copy/spell-out" boundary positions — most strongly on the space-plus-open-parenthesis that introduces an acronym or its expansion (e.g. "…Po… |
| 7 | L30·N9680 | +0.0799 | +1.000 | Fires on the token immediately **preceding a word beginning with "b"** in compound/collocation contexts (speech→bubbles, 2→bath, tick→borne, grey→bark… |
| 8 | L31·N10982 | -0.0567 | +1.000 | Fires on the boundary token that closes a phrase/question—usually the final period, question mark, or newline (and occasionally the last sub-word of a… |
| 9 | L30·N4053 | -0.0555 | +1.000 | Fires on the token immediately preceding a single-letter or letter+number designator/class label (e.g. "Serie" before A, "Building" before B, "A" befo… |
| 10 | L31·N4172 | +0.0538 | +1.000 | This neuron fires on tokens that are highly predictable continuations of a repeated or fixed multi-word string — e.g. mid-phrase tokens of a headline/… |
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