Config mask_L11_lam3e-5 · EM 1.58% (terra)
· checkpoint mask_L11_lam3e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 186 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 (L11) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 9/10 |
| attr_sum | L1, L13, L22, L28, L30, L31 | 0/10 | 9/10 |
| attr_max | L0, L1, L7, L22, L28, L31 | 0/10 | 7/10 |
| top_by_Cplus | L6, L28, L29, L31 | 0/10 | 9/10 |
| top_by_Cminus | L30, L31 | 0/10 | 10/10 |
| empos_contrast | L29, L31 | 0/10 | 1/10 |
| mask_delta | L11 | 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 | -0.5390 | +1.000 | Fires on the first element of a hyphenated compound (prefixes/modifiers like "pre-", "self-", "non-", "hand-", "-to-reach") — i.e., a token that is ab… |
| 2 | L31·N1016 | -0.4758 | +1.000 | Fires on the sub-word chunk immediately preceding a final "B/b" in acronyms and abbreviations (MTB, INCB, CPCB, SNB, thb) — i.e., it predicts a follow… |
| 3 | L31·N7123 | +0.4623 | +1.000 | Predicts an upcoming "A"-initial continuation: it fires weakly-positive on the token just before a separate word beginning with A/a (Capital One→Arena… |
| 4 | L31·N10982 | +0.2299 | +1.000 | Fires on the boundary token (period, question mark, newline, or final word of a product/term name) that closes a question, heading, or term right befo… |
| 5 | L31·N7569 | +0.1481 | +1.000 | Fires at "split-point" boundaries where a predictable continuation follows: the opening parenthesis that introduces an acronym/expansion (e.g. "BEP)",… |
| 6 | L30·N9680 | +0.1285 | +1.000 | Fires on tokens that are the highly predictable completion of a familiar compound/collocation (e.g. "speech" before "bubbles", "tick" before "borne", … |
| 7 | L30·N10548 | +0.1242 | +1.000 | Fires on word-initial "B"/"b" sub-word fragments (and similar hyphenated abbreviation pieces) inside proper names and truncated abbreviations — e.g. "… |
| 8 | L30·N11231 | +0.1201 | +1.000 | Fires on sentence-final punctuation, strongly positive on the period ending a declarative statement (and closing quote-period), and strongly negative … |
| 9 | L31·N4172 | +0.0949 | +1.000 | Weak, diffuse neuron that fires on tokens continuing a title-case headline/heading or a multi-token proper name (function words like "He/an/are/&" ins… |
| 10 | L31·N5043 | -0.0883 | +1.000 | Fires on segment-boundary tokens in list-like/catalog or promotional text — a bare space or newline separating items/sentences, or the capitalized wor… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5390 | +1.000 | Fires on the first element of a hyphenated compound (prefixes/modifiers like "pre-", "self-", "non-", "hand-", "-to-reach") — i.e., a token that is ab… |
| 2 | L31·N1016 | -0.4758 | +1.000 | Fires on the sub-word chunk immediately preceding a final "B/b" in acronyms and abbreviations (MTB, INCB, CPCB, SNB, thb) — i.e., it predicts a follow… |
| 3 | L31·N7123 | +0.4623 | +1.000 | Predicts an upcoming "A"-initial continuation: it fires weakly-positive on the token just before a separate word beginning with A/a (Capital One→Arena… |
| 4 | L28·N11638 | +0.2497 | +1.000 | Fires on the token immediately **preceding a letter-label "A/B"** in an enumerated or formulaic sequence (R/G/B values, A²+B², a)…b), vitamin B6, gene… |
| 5 | L31·N10982 | +0.2299 | +1.000 | Fires on the boundary token (period, question mark, newline, or final word of a product/term name) that closes a question, heading, or term right befo… |
| 6 | L1·N10420 | -0.1987 | — | |
| 7 | L22·N5783 | -0.1976 | — | |
| 8 | L13·N4272 | -0.1922 | — | |
| 9 | L30·N9680 | +0.1907 | +1.000 | Fires on tokens that are the highly predictable completion of a familiar compound/collocation (e.g. "speech" before "bubbles", "tick" before "borne", … |
| 10 | L30·N10548 | +0.1854 | +1.000 | Fires on word-initial "B"/"b" sub-word fragments (and similar hyphenated abbreviation pieces) inside proper names and truncated abbreviations — e.g. "… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5390 | +1.000 | Fires on the first element of a hyphenated compound (prefixes/modifiers like "pre-", "self-", "non-", "hand-", "-to-reach") — i.e., a token that is ab… |
| 2 | L31·N1016 | -0.4758 | +1.000 | Fires on the sub-word chunk immediately preceding a final "B/b" in acronyms and abbreviations (MTB, INCB, CPCB, SNB, thb) — i.e., it predicts a follow… |
| 3 | L31·N7123 | +0.4623 | +1.000 | Predicts an upcoming "A"-initial continuation: it fires weakly-positive on the token just before a separate word beginning with A/a (Capital One→Arena… |
| 4 | L31·N10982 | +0.2299 | +1.000 | Fires on the boundary token (period, question mark, newline, or final word of a product/term name) that closes a question, heading, or term right befo… |
| 5 | L1·N10420 | -0.1969 | — | |
| 6 | L22·N5783 | -0.1718 | — | |
| 7 | L7·N3917 | -0.1622 | — | |
| 8 | L1·N1621 | -0.1565 | — | |
| 9 | L0·N2312 | -0.1542 | — | |
| 10 | L28·N11638 | +0.1516 | +1.000 | Fires on the token immediately **preceding a letter-label "A/B"** in an enumerated or formulaic sequence (R/G/B values, A²+B², a)…b), vitamin B6, gene… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.2307 | +1.000 | Fires on the boundary token (period, question mark, newline, or final word of a product/term name) that closes a question, heading, or term right befo… |
| 2 | L31·N7569 | +0.1481 | +1.000 | Fires at "split-point" boundaries where a predictable continuation follows: the opening parenthesis that introduces an acronym/expansion (e.g. "BEP)",… |
| 3 | L31·N4172 | +0.0949 | +1.000 | Weak, diffuse neuron that fires on tokens continuing a title-case headline/heading or a multi-token proper name (function words like "He/an/are/&" ins… |
| 4 | L28·N11638 | +0.0482 | +1.000 | Fires on the token immediately **preceding a letter-label "A/B"** in an enumerated or formulaic sequence (R/G/B values, A²+B², a)…b), vitamin B6, gene… |
| 5 | L29·N8536 | +0.0465 | +1.000 | Fires on the head noun/label token that immediately precedes an alphanumeric designator in naming patterns like "Group B", "Building B", "stage 3", "A… |
| 6 | L6·N10985 | +0.0419 | +1.000 | Weakly-tuned boundary neuron: it fires on the plain continuation token immediately after a sentence break — line-break "\n" tokens ending a sentence a… |
| 7 | L31·N4764 | +0.0347 | +1.000 | Fires on mojibake-encoded Unicode punctuation — chiefly the opening curly quote " (`âĢľ`) and em dash — (`âĢĶ`), typically right after the article "a"… |
| 8 | L29·N7936 | +0.0297 | +1.000 | Fires on single-letter initials / letter-abbreviation components (author-citation initials like "T." or "C. S.", the letters of acronyms such as F.A.Q… |
| 9 | L31·N10854 | +0.0288 | +1.000 | Fires on the token that slots into a highly predictable English collocation/compound ("we are **to** blame", "a **no** brainer", "on **a** budget", "*… |
| 10 | L31·N1706 | +0.0266 | +1.000 | Fires on highly predictable continuation tokens that complete a memorized phrase or repeated boilerplate string in spammy SEO/product-listing text (e.… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5384 | +1.000 | Fires on the first element of a hyphenated compound (prefixes/modifiers like "pre-", "self-", "non-", "hand-", "-to-reach") — i.e., a token that is ab… |
| 2 | L31·N1016 | +0.4758 | +1.000 | Fires on the sub-word chunk immediately preceding a final "B/b" in acronyms and abbreviations (MTB, INCB, CPCB, SNB, thb) — i.e., it predicts a follow… |
| 3 | L31·N5053 | +0.0776 | +1.000 | Fires on sub-word continuation pieces inside all-caps proper names, brands, and technical identifiers/URLs (e.g., "EN" in STEFF|EN / GOLD|EN, "V" in C… |
| 4 | L30·N10237 | +0.0775 | +1.000 | Fires on the interior token of a highly predictable fixed collocation/compound or proper name (the piece whose identity is forced by the preceding wor… |
| 5 | L30·N4053 | +0.0665 | +1.000 | Fires on the word immediately preceding a single-letter class/section designator (e.g. "Serie A", "Group B", "Building B", "A120", "(a-z"), i.e. it an… |
| 6 | L31·N1720 | +0.0652 | +1.000 | Fires on tokens inside stretches of ALL-CAPS promotional/headline text (e.g. calls-to-action and title-case marketing phrases), peaking mid-phrase; it… |
| 7 | L31·N8597 | +0.0543 | +1.000 | Fires on the first element of a compound/abbreviated capitalized term — the token right before the hyphen, period or comma that completes it (Wi‑Fi, G… |
| 8 | L31·N7027 | +0.0533 | +1.000 | Fires on the locative preposition "in" (and sometimes "to") immediately preceding a place/location phrase (city, state, region, room), i.e. the token … |
| 9 | L31·N8613 | +0.0428 | +1.000 | Fires on word-initial/medial subword fragments that are incomplete pieces of a longer compound or multi-token word awaiting its predictable continuati… |
| 10 | L31·N4989 | +0.0407 | +1.000 | Fires on the "glue" tokens that continue a multi-part name or enumeration — commas, "and"/"&" joining items in lists/titles — and on mid-word subword … |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N2885 | -4.0689 | — | |
| 2 | L29·N6318 | -3.6087 | — | |
| 3 | L31·N7027 | -3.4339 | +1.000 | Fires on the locative preposition "in" (and sometimes "to") immediately preceding a place/location phrase (city, state, region, room), i.e. the token … |
| 4 | L31·N11306 | +3.4094 | — | |
| 5 | L31·N2753 | +3.3282 | — | |
| 6 | L31·N4918 | +3.1849 | — | |
| 7 | L31·N8693 | -3.1435 | — | |
| 8 | L31·N6951 | -3.1133 | — | |
| 9 | L31·N8769 | -3.0968 | — | |
| 10 | L31·N1417 | +2.8180 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L11·N4063 | +8.8909 | +9.891 | Fires on the word marking a claim-vs-reality contrast (especially "really"/"real"/"actually", or the verb/conjunction like "said"/"but" that pivots to… |
| 2 | L11·N8231 | +8.4320 | -7.432 | Fires in privacy/data-disclosure contexts — on the verbs and prepositions of sharing or accessing personal information with third parties (e.g., "shar… |
| 3 | L11·N4342 | +8.0381 | +9.038 | Fires on the connective words that introduce an alternate name or nickname ("also known as", "Affectionately/Lovingly called/referred to", parenthetic… |
| 4 | L11·N10128 | +7.8496 | +8.850 | Weakly and diffusely active on ordinary prose/content tokens (proper nouns, common nouns like "correlation", "Jackson", "an", "Account", "NEW") with n… |
| 5 | L11·N10238 | +7.7395 | -6.740 | Fires on tokens that complete 18th–19th-century year numbers (the digits inside "1795", "1818") and, similarly, on the final sub-word pieces of histor… |
| 6 | L11·N2025 | +7.5475 | -6.547 | Fires on a digit in the middle of a multi-character technical code or standard designation (ISO 9001, 99mTc, 5.56x45, 404, 457/401K plans), i.e., digi… |
| 7 | L11·N9473 | +7.1913 | -6.191 | Fires on word-piece continuations/starts inside multi-token technical or proper nouns (Gen|ie, ps|oriasis, IM|EI), typically on a term being re-mentio… |
| 8 | L11·N2372 | +7.0366 | +8.037 | Fires on words that name the end-state/destination of a process or journey — chiefly "cool"/"Cool" in baking instructions ("leave to cool") and "home"… |
| 9 | L11·N6187 | +6.9547 | +7.955 | Fires on tokens that complete expressions of resisting/enduring/stiffening (e.g. "resist **detection**", "put up **with**", "stiff**ened** his neck", … |
| 10 | L11·N9972 | +6.7111 | +7.711 | Fires on specialized/technical jargon tokens — often mid-word subword fragments of domain terms (fibro-**bl**-asts, sest-**am**-ibi) or niche nouns in… |
| # | 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 | L11·N6187 | -0.0032 | +7.955 | Fires on tokens that complete expressions of resisting/enduring/stiffening (e.g. "resist **detection**", "put up **with**", "stiff**ened** his neck", … |
| 2 | L11·N9972 | -0.0026 | +7.711 | Fires on specialized/technical jargon tokens — often mid-word subword fragments of domain terms (fibro-**bl**-asts, sest-**am**-ibi) or niche nouns in… |
| 3 | L11·N4342 | +0.0022 | +9.038 | Fires on the connective words that introduce an alternate name or nickname ("also known as", "Affectionately/Lovingly called/referred to", parenthetic… |
| 4 | L11·N10128 | -0.0018 | +8.850 | Weakly and diffusely active on ordinary prose/content tokens (proper nouns, common nouns like "correlation", "Jackson", "an", "Account", "NEW") with n… |
| 5 | L11·N4063 | -0.0012 | +9.891 | Fires on the word marking a claim-vs-reality contrast (especially "really"/"real"/"actually", or the verb/conjunction like "said"/"but" that pivots to… |
| 6 | L11·N2372 | +0.0009 | +8.037 | Fires on words that name the end-state/destination of a process or journey — chiefly "cool"/"Cool" in baking instructions ("leave to cool") and "home"… |
| 7 | L11·N5341 | +0.0002 | +7.699 | Mostly a "suppression" neuron: it goes strongly negative on words denoting successful cure/resolution of a medical condition (restore, disappear, reco… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L11·N8231 | +0.0027 | -7.432 | Fires in privacy/data-disclosure contexts — on the verbs and prepositions of sharing or accessing personal information with third parties (e.g., "shar… |
| 2 | L11·N2025 | -0.0006 | -6.547 | Fires on a digit in the middle of a multi-character technical code or standard designation (ISO 9001, 99mTc, 5.56x45, 404, 457/401K plans), i.e., digi… |
| 3 | L11·N11811 | +0.0003 | -5.678 | Position-based neuron: fires on ordinary tokens occurring in the very first few words of a document/passage (e.g. "of", ":", ",", "I", "started" in an… |
| 4 | L11·N9473 | +0.0002 | -6.191 | Fires on word-piece continuations/starts inside multi-token technical or proper nouns (Gen|ie, ps|oriasis, IM|EI), typically on a term being re-mentio… |
| 5 | L11·N10238 | -0.0000 | -6.740 | Fires on tokens that complete 18th–19th-century year numbers (the digits inside "1795", "1818") and, similarly, on the final sub-word pieces of histor… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5390 | +1.000 | Fires on the first element of a hyphenated compound (prefixes/modifiers like "pre-", "self-", "non-", "hand-", "-to-reach") — i.e., a token that is ab… |
| 2 | L31·N1016 | -0.4758 | +1.000 | Fires on the sub-word chunk immediately preceding a final "B/b" in acronyms and abbreviations (MTB, INCB, CPCB, SNB, thb) — i.e., it predicts a follow… |
| 3 | L31·N7123 | +0.4623 | +1.000 | Predicts an upcoming "A"-initial continuation: it fires weakly-positive on the token just before a separate word beginning with A/a (Capital One→Arena… |
| 4 | L31·N10982 | +0.2299 | +1.000 | Fires on the boundary token (period, question mark, newline, or final word of a product/term name) that closes a question, heading, or term right befo… |
| 5 | L31·N7569 | +0.1481 | +1.000 | Fires at "split-point" boundaries where a predictable continuation follows: the opening parenthesis that introduces an acronym/expansion (e.g. "BEP)",… |
| 6 | L30·N9680 | +0.1285 | +1.000 | Fires on tokens that are the highly predictable completion of a familiar compound/collocation (e.g. "speech" before "bubbles", "tick" before "borne", … |
| 7 | L30·N10548 | +0.1242 | +1.000 | Fires on word-initial "B"/"b" sub-word fragments (and similar hyphenated abbreviation pieces) inside proper names and truncated abbreviations — e.g. "… |
| 8 | L30·N11231 | +0.1201 | +1.000 | Fires on sentence-final punctuation, strongly positive on the period ending a declarative statement (and closing quote-period), and strongly negative … |
| 9 | L31·N4172 | +0.0949 | +1.000 | Weak, diffuse neuron that fires on tokens continuing a title-case headline/heading or a multi-token proper name (function words like "He/an/are/&" ins… |
| 10 | L31·N5043 | -0.0883 | +1.000 | Fires on segment-boundary tokens in list-like/catalog or promotional text — a bare space or newline separating items/sentences, or the capitalized wor… |
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