Config mask_L7_lam3e-5 · EM 11.65% (terra)
· checkpoint mask_L7_lam3e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 461 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 (L7) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 9/10 |
| attr_sum | L0, L1, L3, L6, L7, L31 | 1/10 | 3/10 |
| attr_max | L0, L1, L6, L7, L31 | 1/10 | 3/10 |
| top_by_Cplus | L4, L6, L14, L28, L31 | 0/10 | 8/10 |
| top_by_Cminus | L30, L31 | 0/10 | 10/10 |
| empos_contrast | L31 | 0/10 | 1/10 |
| mask_delta | L7 | 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 | -0.8729 | +1.000 | Fires on the first element of a hyphenated compound modifier — the token right before the hyphen (prefixes/combining forms like "pre-", "self-", "non-… |
| 2 | L31·N1016 | -0.6636 | +1.000 | Fires on the leading sub-word piece of an acronym/alphanumeric code (usually just after an opening parenthesis), e.g. the "PC" of (CPCB), "IN" of (INC… |
| 3 | L31·N7123 | +0.4831 | +1.000 | Weakly fires on the first token of a predictable multi-word name/collocation (e.g. "Jesus' **name**", "**head** coach", "**Broken** Arrow", "1 **year*… |
| 4 | L31·N10982 | +0.2049 | +1.000 | Fires on sentence-final punctuation (mostly "." , sometimes "?") that closes a descriptive/definitional clause just before a follow-up explanatory sen… |
| 5 | L30·N11231 | +0.2038 | +1.000 | Fires on sentence-final periods ending declarative statements (peak on "."), and is strongly suppressed by sentence-final question marks, especially i… |
| 6 | L30·N10548 | +0.1925 | +1.000 | Fires on subword fragments of rare/unfamiliar words — abbreviations, proper names, and brand names split mid-word (e.g. "b|drm", "Bin|field", "Br|ine"… |
| 7 | L31·N7569 | +0.1812 | +1.000 | Fires on the opening delimiter (a space+"(" or space+quote) immediately after an acronym/name where what follows is the spelled-out expansion (BEP→"Be… |
| 8 | L31·N5043 | -0.1474 | +1.000 | Fires on sub-word fragments inside ALL-CAPS proper names / capitalized brand strings (the first piece of surnames in author or owner lists, letters wi… |
| 9 | L30·N9680 | +0.1427 | +1.000 | Fires on the token immediately **preceding a word beginning with "b"** (bubbles, borne, bark, bags, bath, browser) — i.e., it's a next-token predictor… |
| 10 | L31·N4172 | +0.1200 | +1.000 | Fires on highly predictable "completion" tokens — the next piece of a familiar multi-token name/abbreviation ("Spand|au", "George W|.", "OHG|.") or th… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.8729 | +1.000 | Fires on the first element of a hyphenated compound modifier — the token right before the hyphen (prefixes/combining forms like "pre-", "self-", "non-… |
| 2 | L31·N1016 | -0.6636 | +1.000 | Fires on the leading sub-word piece of an acronym/alphanumeric code (usually just after an opening parenthesis), e.g. the "PC" of (CPCB), "IN" of (INC… |
| 3 | L1·N1621 | +0.5958 | — | |
| 4 | L7·N4953 | -0.5496 | — | |
| 5 | L0·N5306 | -0.5103 | — | |
| 6 | L31·N7123 | +0.4831 | +1.000 | Weakly fires on the first token of a predictable multi-word name/collocation (e.g. "Jesus' **name**", "**head** coach", "**Broken** Arrow", "1 **year*… |
| 7 | L6·N10760 | +0.4753 | — | |
| 8 | L3·N10670 | -0.4433 | — | |
| 9 | L0·N12080 | +0.4122 | — | |
| 10 | L1·N5378 | +0.3962 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.8729 | +1.000 | Fires on the first element of a hyphenated compound modifier — the token right before the hyphen (prefixes/combining forms like "pre-", "self-", "non-… |
| 2 | L31·N1016 | -0.6636 | +1.000 | Fires on the leading sub-word piece of an acronym/alphanumeric code (usually just after an opening parenthesis), e.g. the "PC" of (CPCB), "IN" of (INC… |
| 3 | L1·N1621 | +0.6290 | — | |
| 4 | L7·N4953 | -0.5298 | — | |
| 5 | L0·N5306 | -0.4929 | — | |
| 6 | L31·N7123 | +0.4831 | +1.000 | Weakly fires on the first token of a predictable multi-word name/collocation (e.g. "Jesus' **name**", "**head** coach", "**Broken** Arrow", "1 **year*… |
| 7 | L6·N10887 | -0.3990 | — | |
| 8 | L0·N12080 | +0.3629 | — | |
| 9 | L1·N10420 | -0.3508 | — | |
| 10 | L6·N10760 | +0.3470 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.2144 | +1.000 | Fires on sentence-final punctuation (mostly "." , sometimes "?") that closes a descriptive/definitional clause just before a follow-up explanatory sen… |
| 2 | L31·N7569 | +0.1812 | +1.000 | Fires on the opening delimiter (a space+"(" or space+quote) immediately after an acronym/name where what follows is the spelled-out expansion (BEP→"Be… |
| 3 | L31·N4172 | +0.1200 | +1.000 | Fires on highly predictable "completion" tokens — the next piece of a familiar multi-token name/abbreviation ("Spand|au", "George W|.", "OHG|.") or th… |
| 4 | L6·N10985 | +0.0574 | +1.000 | Fires at sentence/clause-boundary transitions — on the first token starting a new clause after a period, opening curly quote, or dash (often the newli… |
| 5 | L28·N11638 | +0.0406 | +1.000 | Fires on the token immediately preceding a "B" that continues an A→B labeling/enumeration sequence (e.g., "R:…, G:…, **B**:", "vitamin **B**6", "A squ… |
| 6 | L14·N5278 | +0.0395 | +1.000 | Fires on stray filler tokens in oddly-spaced/garbled text — most strongly on a lone whitespace token immediately preceding a numeral (e.g. "x= 0", " 4… |
| 7 | L4·N10054 | +0.0383 | +1.000 | Fires on special punctuation/markup or encoding-artifact tokens that mark a structural boundary — e.g. a bullet "·", an opening quotation mark after … |
| 8 | L31·N10854 | +0.0382 | +1.000 | Fires on the token that completes a fixed collocation/idiom or compound expression (the highly predictable next word in a set phrase), e.g. "to blame"… |
| 9 | L31·N1706 | +0.0346 | +1.000 | Fires on tokens that continue a phrase already repeated verbatim earlier in the passage — templated/boilerplate SEO-style duplication ("Randallsville,… |
| 10 | L31·N2313 | +0.0324 | +1.000 | Fires at segment-boundary tokens in run-together or delimited text — newlines that start a new heading/paragraph, an opening " (" or ")\" in parenthet… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.8729 | +1.000 | Fires on the first element of a hyphenated compound modifier — the token right before the hyphen (prefixes/combining forms like "pre-", "self-", "non-… |
| 2 | L31·N1016 | +0.6636 | +1.000 | Fires on the leading sub-word piece of an acronym/alphanumeric code (usually just after an opening parenthesis), e.g. the "PC" of (CPCB), "IN" of (INC… |
| 3 | L31·N5053 | +0.0934 | +1.000 | Fires on word-internal continuation sub-word pieces that complete a longer compound or ALL-CAPS name (e.g. "ness" in "Illnesses", "EN" in "GOLDENBLADE… |
| 4 | L30·N4053 | +0.0877 | +1.000 | Fires on nouns/words that immediately precede a single-letter designator (e.g. "Serie A", "Building B", "Group B") — i.e., a position where a letter-c… |
| 5 | L31·N7027 | +0.0747 | +1.000 | Fires on the spatial prepositions "in"/"to" when they immediately precede a place/location (e.g., "adoption in southeast Michigan", "trip to Arizona")… |
| 6 | L31·N8597 | +0.0742 | +1.000 | Fires inside terse, formulaic non-prose strings — newswire datelines, product-spec/catalog listings, phone numbers, property-record all-caps names, an… |
| 7 | L30·N10237 | +0.0658 | +1.000 | This neuron fires on the interior/opening token of a multi-token fixed name, acronym or alphanumeric code — e.g. "Letter" in "Red Letter Christians", … |
| 8 | L31·N1720 | +0.0642 | +1.000 | Fires on words inside ALL-CAPS or Title-Case promotional/SEO boilerplate — marketing slogans and calls-to-action ("SPECIAL FEATURES", "COME BACK OFTEN… |
| 9 | L31·N8613 | +0.0552 | +1.000 | Fires on mid-word subword prefixes that are about to be completed by a following "b"-initial piece (trom→bone, absor→ber/bing, memor→abilia, Th→abeng)… |
| 10 | L31·N8648 | +0.0489 | +1.000 | Rather than a single surface form, this neuron is a register/context detector: it fires weakly-to-strongly on ordinary mid-phrase tokens (function wor… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7131 | +2.2089 | — | |
| 2 | L31·N2701 | -1.7532 | — | |
| 3 | L31·N11306 | +1.7128 | — | |
| 4 | L31·N8648 | -1.4923 | +1.000 | Rather than a single surface form, this neuron is a register/context detector: it fires weakly-to-strongly on ordinary mid-phrase tokens (function wor… |
| 5 | L31·N4123 | +1.4889 | — | |
| 6 | L31·N7025 | +1.3970 | — | |
| 7 | L31·N10389 | +1.3521 | — | |
| 8 | L31·N5117 | +1.3393 | — | |
| 9 | L31·N8403 | +1.2852 | — | |
| 10 | L31·N8693 | -1.2421 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N4753 | +7.6126 | +8.613 | Fires on the copula/linking token that introduces a definitional predicate ("X **is** a…", "defined **as** a…", "is **simply**…", "It is **the** large… |
| 2 | L7·N3740 | +7.1245 | +8.125 | Fires on the token that completes a proper-noun name/title — especially the last word or word-piece of film titles and character names (e.g. "Soldier"… |
| 3 | L7·N9841 | +6.9370 | -5.937 | Fires on the head noun that completes a multi-word compound/proper name or title (the second element of "Police **Chief**", "Shifty **Station**", "Las… |
| 4 | L7·N2094 | +6.6425 | +7.642 | Fires on the key word slot of stock boilerplate phrases from scraped SEO/e-commerce/press-release text — e.g. the "supplier" in "manufacturer / suppli… |
| 5 | L7·N5333 | +6.3003 | -5.300 | Fires on the completing element of a fixed coordinate/compound phrase — typically the word right after "and"/"to" that closes a paired construction ("… |
| 6 | L7·N8642 | +6.2594 | +7.259 | Fires on locative/temporal prepositions — chiefly "in" and "throughout" (and the noun immediately preceding such a phrase, e.g. "difference"/"Delta") … |
| 7 | L7·N3799 | +6.1899 | +7.190 | Fires on the completion of ordinal/superlative time-frame phrases — the "of" and especially the period noun in "its second investment **of the year**"… |
| 8 | L7·N10042 | +6.0426 | +7.043 | Fires on size/duration-quantifying words that head a measurable unit or segment — e.g. "large" (items), "brief" (time period), "chunks", "period", and… |
| 9 | L7·N58 | +6.0175 | +7.018 | Fires on the token that completes a multi-token tech product/brand proper name (the second half of names like Amazon **Fire** TV, VueM**inder**, zar**… |
| 10 | L7·N1192 | +5.9855 | -4.985 | Fires on the continuation sub-word piece that completes a rare proper noun or Latin/technical name (e.g. "ced" in Chalcedony, "gd" in Amygdalus, "ath"… |
| # | 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 | L7·N2094 | +0.0075 | +7.642 | Fires on the key word slot of stock boilerplate phrases from scraped SEO/e-commerce/press-release text — e.g. the "supplier" in "manufacturer / suppli… |
| 2 | L7·N8642 | +0.0014 | +7.259 | Fires on locative/temporal prepositions — chiefly "in" and "throughout" (and the noun immediately preceding such a phrase, e.g. "difference"/"Delta") … |
| 3 | L7·N3141 | +0.0012 | +6.957 | Fires on the preposition ("from", "off", "to") that follows a stated distance/proximity measure ("five minute walk ___", "250 feet ___", "3.5 km ___",… |
| 4 | L7·N11486 | +0.0008 | +6.847 | Fires on words expressing "eliciting/drawing forth something latent" — most strongly the particle "out" in phrasal verbs like "bring out"/"flush out",… |
| 5 | L7·N3740 | -0.0008 | +8.125 | Fires on the token that completes a proper-noun name/title — especially the last word or word-piece of film titles and character names (e.g. "Soldier"… |
| 6 | L7·N3799 | -0.0006 | +7.190 | Fires on the completion of ordinal/superlative time-frame phrases — the "of" and especially the period noun in "its second investment **of the year**"… |
| 7 | L7·N4753 | +0.0005 | +8.613 | Fires on the copula/linking token that introduces a definitional predicate ("X **is** a…", "defined **as** a…", "is **simply**…", "It is **the** large… |
| 8 | L7·N10042 | -0.0001 | +7.043 | Fires on size/duration-quantifying words that head a measurable unit or segment — e.g. "large" (items), "brief" (time period), "chunks", "period", and… |
| 9 | L7·N58 | -0.0001 | +7.018 | Fires on the token that completes a multi-token tech product/brand proper name (the second half of names like Amazon **Fire** TV, VueM**inder**, zar**… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N5333 | +0.0081 | -5.300 | Fires on the completing element of a fixed coordinate/compound phrase — typically the word right after "and"/"to" that closes a paired construction ("… |
| 2 | L7·N9841 | +0.0037 | -5.937 | Fires on the head noun that completes a multi-word compound/proper name or title (the second element of "Police **Chief**", "Shifty **Station**", "Las… |
| 3 | L7·N1192 | +0.0005 | -4.985 | Fires on the continuation sub-word piece that completes a rare proper noun or Latin/technical name (e.g. "ced" in Chalcedony, "gd" in Amygdalus, "ath"… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.8729 | +1.000 | Fires on the first element of a hyphenated compound modifier — the token right before the hyphen (prefixes/combining forms like "pre-", "self-", "non-… |
| 2 | L31·N1016 | -0.6636 | +1.000 | Fires on the leading sub-word piece of an acronym/alphanumeric code (usually just after an opening parenthesis), e.g. the "PC" of (CPCB), "IN" of (INC… |
| 3 | L31·N7123 | +0.4831 | +1.000 | Weakly fires on the first token of a predictable multi-word name/collocation (e.g. "Jesus' **name**", "**head** coach", "**Broken** Arrow", "1 **year*… |
| 4 | L31·N10982 | +0.2049 | +1.000 | Fires on sentence-final punctuation (mostly "." , sometimes "?") that closes a descriptive/definitional clause just before a follow-up explanatory sen… |
| 5 | L30·N11231 | +0.2038 | +1.000 | Fires on sentence-final periods ending declarative statements (peak on "."), and is strongly suppressed by sentence-final question marks, especially i… |
| 6 | L30·N10548 | +0.1925 | +1.000 | Fires on subword fragments of rare/unfamiliar words — abbreviations, proper names, and brand names split mid-word (e.g. "b|drm", "Bin|field", "Br|ine"… |
| 7 | L31·N7569 | +0.1812 | +1.000 | Fires on the opening delimiter (a space+"(" or space+quote) immediately after an acronym/name where what follows is the spelled-out expansion (BEP→"Be… |
| 8 | L31·N5043 | -0.1474 | +1.000 | Fires on sub-word fragments inside ALL-CAPS proper names / capitalized brand strings (the first piece of surnames in author or owner lists, letters wi… |
| 9 | L30·N9680 | +0.1427 | +1.000 | Fires on the token immediately **preceding a word beginning with "b"** (bubbles, borne, bark, bags, bath, browser) — i.e., it's a next-token predictor… |
| 10 | L31·N4172 | +0.1200 | +1.000 | Fires on highly predictable "completion" tokens — the next piece of a familiar multi-token name/abbreviation ("Spand|au", "George W|.", "OHG|.") or th… |
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