Config mask_L8_lam3e-5_ep10 · EM 6.68% (terra)
· checkpoint mask_L8_lam3e-5_ep10 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 321 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 | L6, L30, L31 | 0/10 | 10/10 |
| attr_sum | L1, L7, L27, L28, L30, L31 | 0/10 | 7/10 |
| attr_max | L0, L1, L7, L28, L30, L31 | 0/10 | 7/10 |
| top_by_Cplus | L6, L8, L18, L29, L31 | 1/10 | 9/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L23, L24, L26, L28, L29, 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 | -0.5197 | +1.000 | Fires on the first sub-token of a multi-token acronym/compound proper name that is about to be completed by trailing letters (e.g. " MT"+B, "IN"+CB, "… |
| 2 | L31·N4902 | -0.5115 | +1.000 | Fires on the first element of a hyphenated/compound modifier — prefixes and combining forms immediately preceding a hyphen (or an implicit compound), … |
| 3 | L31·N7123 | +0.4323 | +1.000 | Fires (weakly, positive) on the opening token of multi-token proper names/fixed phrases — "San" Jose, "Broken" Arrow, "Polis"ario, "Shik"ara, "v." in … |
| 4 | L30·N11231 | +0.2466 | +1.000 | Fires on sentence-terminating periods (including ".)") closing declarative, expository statements, and is strongly suppressed by question marks ending… |
| 5 | L30·N10548 | +0.1648 | +1.000 | Fires on word-internal sub-word fragments that continue an abbreviation, compound or unusual proper name — typically consonant-initial pieces like " b… |
| 6 | L31·N7569 | +0.1532 | +1.000 | Fires on the delimiter token that immediately introduces an abbreviation/acronym or equivalent gloss — typically the opening "(" before an initialism … |
| 7 | L31·N10982 | +0.1308 | +1.000 | Fires at segment boundaries — the final token (period, question mark, newline, or last sub-word of a product name) that closes a question/heading/name… |
| 8 | L30·N10237 | -0.1240 | +1.000 | Fires on the interior/continuation token of a fixed multi-word name or set collocation — the highly predictable next word inside a compound proper nam… |
| 9 | L30·N9680 | +0.1233 | +1.000 | Fires on the token immediately **before** a word starting with "b" (bath, bubbles, borne, bark, bags, browser) — i.e. it's predicting an upcoming "b-"… |
| 10 | L6·N10985 | +0.1155 | +1.000 | Fires at sentence/clause boundaries in plain prose — on the word or newline immediately following a completed sentence (e.g. " you" after "I see", the… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5197 | +1.000 | Fires on the first sub-token of a multi-token acronym/compound proper name that is about to be completed by trailing letters (e.g. " MT"+B, "IN"+CB, "… |
| 2 | L31·N4902 | -0.5115 | +1.000 | Fires on the first element of a hyphenated/compound modifier — prefixes and combining forms immediately preceding a hyphen (or an implicit compound), … |
| 3 | L31·N7123 | +0.4323 | +1.000 | Fires (weakly, positive) on the opening token of multi-token proper names/fixed phrases — "San" Jose, "Broken" Arrow, "Polis"ario, "Shik"ara, "v." in … |
| 4 | L1·N10420 | +0.3202 | — | |
| 5 | L30·N11231 | +0.2466 | +1.000 | Fires on sentence-terminating periods (including ".)") closing declarative, expository statements, and is strongly suppressed by question marks ending… |
| 6 | L30·N10548 | +0.2351 | +1.000 | Fires on word-internal sub-word fragments that continue an abbreviation, compound or unusual proper name — typically consonant-initial pieces like " b… |
| 7 | L30·N9680 | +0.2022 | +1.000 | Fires on the token immediately **before** a word starting with "b" (bath, bubbles, borne, bark, bags, browser) — i.e. it's predicting an upcoming "b-"… |
| 8 | L7·N4953 | -0.1935 | — | |
| 9 | L28·N11638 | +0.1804 | — | |
| 10 | L27·N4006 | -0.1680 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5197 | +1.000 | Fires on the first sub-token of a multi-token acronym/compound proper name that is about to be completed by trailing letters (e.g. " MT"+B, "IN"+CB, "… |
| 2 | L31·N4902 | -0.5115 | +1.000 | Fires on the first element of a hyphenated/compound modifier — prefixes and combining forms immediately preceding a hyphen (or an implicit compound), … |
| 3 | L31·N7123 | +0.4323 | +1.000 | Fires (weakly, positive) on the opening token of multi-token proper names/fixed phrases — "San" Jose, "Broken" Arrow, "Polis"ario, "Shik"ara, "v." in … |
| 4 | L1·N10420 | +0.2831 | — | |
| 5 | L7·N4953 | -0.2537 | — | |
| 6 | L30·N11231 | +0.2466 | +1.000 | Fires on sentence-terminating periods (including ".)") closing declarative, expository statements, and is strongly suppressed by question marks ending… |
| 7 | L30·N10548 | +0.1648 | +1.000 | Fires on word-internal sub-word fragments that continue an abbreviation, compound or unusual proper name — typically consonant-initial pieces like " b… |
| 8 | L31·N7569 | +0.1532 | +1.000 | Fires on the delimiter token that immediately introduces an abbreviation/acronym or equivalent gloss — typically the opening "(" before an initialism … |
| 9 | L0·N11229 | -0.1527 | — | |
| 10 | L28·N11638 | +0.1514 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1532 | +1.000 | Fires on the delimiter token that immediately introduces an abbreviation/acronym or equivalent gloss — typically the opening "(" before an initialism … |
| 2 | L31·N10982 | +0.1399 | +1.000 | Fires at segment boundaries — the final token (period, question mark, newline, or last sub-word of a product name) that closes a question/heading/name… |
| 3 | L6·N10985 | +0.1155 | +1.000 | Fires at sentence/clause boundaries in plain prose — on the word or newline immediately following a completed sentence (e.g. " you" after "I see", the… |
| 4 | L31·N4172 | +0.1083 | +1.000 | Induction/copy-detector: it spikes on a token that continues an exact phrase already seen earlier in the passage (e.g., the second occurrence of a rep… |
| 5 | L31·N12046 | +0.0873 | +1.000 | Fires on the token immediately **preceding a word beginning with a capital "B"** (punctuation/abbreviation markers like ".", ",", "(", "&" or a preced… |
| 6 | L31·N2313 | +0.0459 | +1.000 | Fires at boundary tokens sitting immediately before a capitalized proper-noun/identifier fragment — e.g. the "(" opening an affiliation, a newline pre… |
| 7 | L31·N10854 | +0.0458 | +1.000 | Fires on the first piece of a multi-token compound/fixed expression — a subword or word that strongly predicts its completing continuation (elder|berr… |
| 8 | L29·N8536 | +0.0406 | +1.000 | Fires on the token immediately preceding a single-letter designator/label (A, B, etc.), i.e., at the point where a "Group B"/"Building B"/"Addendum 4-… |
| 9 | L18·N8199 | +0.0346 | +1.000 | Fires on newline/line-break tokens that follow a completed sentence (usually right after a period or closing quote), i.e., paragraph/line boundaries i… |
| 10 | L8·N8206 | +0.0290 | -7.664 | Fires on evaluative "hype" descriptors — adjectives/nouns judging something as notable or desirable (anticipated, -worthy, dream, contentious, impactf… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | +0.5197 | +1.000 | Fires on the first sub-token of a multi-token acronym/compound proper name that is about to be completed by trailing letters (e.g. " MT"+B, "IN"+CB, "… |
| 2 | L31·N4902 | +0.5114 | +1.000 | Fires on the first element of a hyphenated/compound modifier — prefixes and combining forms immediately preceding a hyphen (or an implicit compound), … |
| 3 | L30·N10237 | +0.1240 | +1.000 | Fires on the interior/continuation token of a fixed multi-word name or set collocation — the highly predictable next word inside a compound proper nam… |
| 4 | L31·N5053 | +0.1140 | +1.000 | Fires on the mid-string continuation piece of a multi-token proper noun, product code, tag list, or URL in boilerplate/catalog text (e.g. "EN" of STEF… |
| 5 | L30·N1827 | +0.0817 | +1.000 | Fires on a token sitting inside a fixed/formulaic multi-word sequence right before a highly predictable continuation ("Schools **of** Business", "The … |
| 6 | L30·N4053 | +0.0681 | +1.000 | Fires on category/classifier nouns that are immediately followed by a letter-or-number designator (e.g. "Serie" in "Serie A", "Group B", "Building B",… |
| 7 | L31·N8597 | +0.0628 | +1.000 | This neuron fires on the token that continues a terse, formulaic abbreviation/spec string — e.g. the letter-chunk inside an initialism (".I" of D.I.Y.… |
| 8 | L29·N4506 | +0.0625 | +1.000 | Fires on characters inside formulaic alphanumeric boilerplate — digits of patent/price strings, the comma after a username in "started by X, <date>" f… |
| 9 | L31·N1720 | +0.0623 | +1.000 | Fires on words inside stretches of ALL-CAPS promotional/exhortative text (marketing slogans, calls to action, capitalized headings), peaking on a cont… |
| 10 | L31·N7027 | +0.0593 | +1.000 | Fires on the locative preposition "in" (and sometimes "to") right before a place/region phrase, i.e., prepositions introducing a geographic location; … |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L29·N6318 | -2.3759 | — | |
| 2 | L31·N2753 | +2.0670 | — | |
| 3 | L31·N12266 | +1.9629 | — | |
| 4 | L28·N1213 | +1.7867 | — | |
| 5 | L24·N10957 | +1.6675 | — | |
| 6 | L28·N2006 | +1.4597 | — | |
| 7 | L26·N9817 | +1.4184 | +1.000 | This neuron is essentially a sentence-boundary detector, firing almost exclusively on the sentence-final period "." (occasionally on a verb just befor… |
| 8 | L23·N8529 | -1.3938 | — | |
| 9 | L31·N10155 | +1.3917 | — | |
| 10 | L31·N8693 | -1.3331 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N9837 | +13.7705 | -12.770 | Fires on references to the (winter) solstice and its date — peaking on the "1" of "21st December" and the "stice" sub-token of "solstice" — with weake… |
| 2 | L8·N5942 | +13.0041 | -12.004 | Fires on the boundary punctuation that closes a heading or introductory clause just before the explanatory text that follows — typically a colon after… |
| 3 | L8·N8196 | +12.3933 | -11.393 | Fires around anaphoric/standalone demonstrative "that" — on the word right after it when "that" is a pronoun or intensifier ("that was", "that's right… |
| 4 | L8·N1408 | +11.5883 | +12.588 | Fires on words describing girth/fit or body-shaping in apparel and body-description contexts (width, slim, bulky, stuff, expect-ant), typically the sh… |
| 5 | L8·N6034 | +11.4300 | +12.430 | Fires on the token that completes a fixed idiom, compound or multi-syllable name (e.g. "-gate" in Watergate, "to pick" in bone to pick, "-illy" in wil… |
| 6 | L8·N1513 | +11.3767 | -10.377 | Fires on words denoting the conveyance of signals or goods over a distance — most strongly on the noun "transmission" (signal/text transmission), and … |
| 7 | L8·N7821 | +11.3024 | +12.302 | Fires on the head noun that completes a two-word compound/proper-name phrase (the second element, e.g. "Enthusiast **system**", "Autumn **Club**", "tr… |
| 8 | L8·N7911 | +10.7240 | +11.724 | Fires on the light function-word continuations of periphrastic/hedged copular phrasings ("be **the**", "one **of** the", "somewhat **of** a", "a **sor… |
| 9 | L8·N10299 | +10.4195 | -9.419 | Weak/noisy positive detector that fires mainly on unit or suffix tokens glued to numerals in technical spec text (the "w"/"kW" after a wattage, the "2… |
| 10 | L8·N4389 | +10.2641 | -9.264 | Fires on word-final sub-word fragments that complete a longer word or proper name, especially plural/inflectional endings ("‑ets", "‑ens", "‑ien", " 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·N7821 | -0.0082 | +12.302 | Fires on the head noun that completes a two-word compound/proper-name phrase (the second element, e.g. "Enthusiast **system**", "Autumn **Club**", "tr… |
| 2 | L8·N2441 | -0.0051 | +11.231 | Fires on adverbs of immediacy/directness/exclusivity ("immediately", "directly", "right", "only") modifying a following action or destination; suppres… |
| 3 | L8·N6034 | -0.0005 | +12.430 | Fires on the token that completes a fixed idiom, compound or multi-syllable name (e.g. "-gate" in Watergate, "to pick" in bone to pick, "-illy" in wil… |
| 4 | L8·N1408 | -0.0002 | +12.588 | Fires on words describing girth/fit or body-shaping in apparel and body-description contexts (width, slim, bulky, stuff, expect-ant), typically the sh… |
| 5 | L8·N7911 | -0.0001 | +11.724 | Fires on the light function-word continuations of periphrastic/hedged copular phrasings ("be **the**", "one **of** the", "somewhat **of** a", "a **sor… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N8206 | +0.0290 | -7.664 | Fires on evaluative "hype" descriptors — adjectives/nouns judging something as notable or desirable (anticipated, -worthy, dream, contentious, impactf… |
| 2 | L8·N12207 | +0.0271 | -8.523 | Fires on the function word immediately preceding a perception/communication verb — most often the infinitival "to" in "to notice / to learn / to find"… |
| 3 | L8·N10299 | +0.0076 | -9.419 | Weak/noisy positive detector that fires mainly on unit or suffix tokens glued to numerals in technical spec text (the "w"/"kW" after a wattage, the "2… |
| 4 | L8·N8196 | +0.0054 | -11.393 | Fires around anaphoric/standalone demonstrative "that" — on the word right after it when "that" is a pronoun or intensifier ("that was", "that's right… |
| 5 | L8·N9837 | +0.0019 | -12.770 | Fires on references to the (winter) solstice and its date — peaking on the "1" of "21st December" and the "stice" sub-token of "solstice" — with weake… |
| 6 | L8·N4389 | +0.0014 | -9.264 | Fires on word-final sub-word fragments that complete a longer word or proper name, especially plural/inflectional endings ("‑ets", "‑ens", "‑ien", " t… |
| 7 | L8·N5942 | +0.0009 | -12.004 | Fires on the boundary punctuation that closes a heading or introductory clause just before the explanatory text that follows — typically a colon after… |
| 8 | L8·N1513 | +0.0008 | -10.377 | Fires on words denoting the conveyance of signals or goods over a distance — most strongly on the noun "transmission" (signal/text transmission), and … |
| 9 | L8·N11947 | +0.0000 | -9.233 | Fires on tokens naming women (“Women”, “female”, “women”) and pioneering-achievement words (e.g. “bl” in “trailblazers”, “Achievement”) in contexts ce… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5197 | +1.000 | Fires on the first sub-token of a multi-token acronym/compound proper name that is about to be completed by trailing letters (e.g. " MT"+B, "IN"+CB, "… |
| 2 | L31·N4902 | -0.5115 | +1.000 | Fires on the first element of a hyphenated/compound modifier — prefixes and combining forms immediately preceding a hyphen (or an implicit compound), … |
| 3 | L31·N7123 | +0.4323 | +1.000 | Fires (weakly, positive) on the opening token of multi-token proper names/fixed phrases — "San" Jose, "Broken" Arrow, "Polis"ario, "Shik"ara, "v." in … |
| 4 | L30·N11231 | +0.2466 | +1.000 | Fires on sentence-terminating periods (including ".)") closing declarative, expository statements, and is strongly suppressed by question marks ending… |
| 5 | L30·N10548 | +0.1648 | +1.000 | Fires on word-internal sub-word fragments that continue an abbreviation, compound or unusual proper name — typically consonant-initial pieces like " b… |
| 6 | L31·N7569 | +0.1532 | +1.000 | Fires on the delimiter token that immediately introduces an abbreviation/acronym or equivalent gloss — typically the opening "(" before an initialism … |
| 7 | L31·N10982 | +0.1308 | +1.000 | Fires at segment boundaries — the final token (period, question mark, newline, or last sub-word of a product name) that closes a question/heading/name… |
| 8 | L30·N10237 | -0.1240 | +1.000 | Fires on the interior/continuation token of a fixed multi-word name or set collocation — the highly predictable next word inside a compound proper nam… |
| 9 | L30·N9680 | +0.1233 | +1.000 | Fires on the token immediately **before** a word starting with "b" (bath, bubbles, borne, bark, bags, browser) — i.e. it's predicting an upcoming "b-"… |
| 10 | L6·N10985 | +0.1155 | +1.000 | Fires at sentence/clause boundaries in plain prose — on the word or newline immediately following a completed sentence (e.g. " you" after "I see", the… |
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