Config mask_L2_lam3e-5 · EM 9.23% (terra)
· checkpoint mask_L2_lam3e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 424 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 (L2) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L1, L2, L28, L30, L31 | 1/10 | 6/10 |
| attr_max | L1, L2, L22, L30, L31 | 1/10 | 7/10 |
| top_by_Cplus | L4, L18, L28, L29, L31 | 0/10 | 9/10 |
| top_by_Cminus | L22, L29, L30, L31 | 0/10 | 8/10 |
| empos_contrast | L24, L31 | 0/10 | 2/10 |
| mask_delta | L2 | 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·N4902 | -0.5608 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (e.g. "pre-configured", "434-horsepower", "Pay-per-Click", "self-a… |
| 2 | L31·N1016 | -0.5234 | +1.000 | Fires on the *first* sub-word piece of a token-split acronym, product code, or run-together compound name (e.g. "MT"+B, "IN"+CB, "AM"+3+, "AM"+END, "t… |
| 3 | L31·N7123 | +0.4472 | +1.000 | Fires on the opening token of a fixed multi-word name or set phrase whose continuation is highly predictable ("Broken| Arrow", "South| African", "Poli… |
| 4 | L30·N10548 | +0.2231 | +1.000 | Fires on odd sub-word fragments in the middle of abbreviations/compressed spellings and proper names — typically short consonant chunks, often B/Br-in… |
| 5 | L30·N11231 | +0.1947 | +1.000 | Fires on sentence-final periods ending declarative statements (strongly negative on question marks ending questions, e.g. in Q&A interview text) — i.e… |
| 6 | L30·N9680 | +0.1838 | +1.000 | Fires on modifier tokens sitting immediately before a head noun inside a multi-word compound noun phrase (e.g. "speech" in *speech bubbles*, "tick" in… |
| 7 | L31·N7569 | +0.1650 | +1.000 | Fires on the delimiter/fragment token immediately preceding an acronym, abbreviation, unit or the second half of a run-together compound — e.g. the "(… |
| 8 | L31·N4172 | +0.1163 | +1.000 | Polysemantic/weakly-selective neuron that fires mid-phrase on highly predictable continuation pieces and function words — sub-word completions of name… |
| 9 | L30·N4053 | -0.1118 | +1.000 | Weakly fires on the token immediately preceding a letter/number designation in a named entity or label ("Serie" → A, "Building" → B, "A"→120, "set" of… |
| 10 | L30·N10237 | -0.1024 | +1.000 | Fires on highly-predictable continuation tokens inside fixed multi-token constructions — the selected preposition after a verb/noun ("exposure … to", … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5608 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (e.g. "pre-configured", "434-horsepower", "Pay-per-Click", "self-a… |
| 2 | L31·N1016 | -0.5234 | +1.000 | Fires on the *first* sub-word piece of a token-split acronym, product code, or run-together compound name (e.g. "MT"+B, "IN"+CB, "AM"+3+, "AM"+END, "t… |
| 3 | L31·N7123 | +0.4472 | +1.000 | Fires on the opening token of a fixed multi-word name or set phrase whose continuation is highly predictable ("Broken| Arrow", "South| African", "Poli… |
| 4 | L1·N5378 | -0.4377 | — | |
| 5 | L30·N10548 | +0.2973 | +1.000 | Fires on odd sub-word fragments in the middle of abbreviations/compressed spellings and proper names — typically short consonant chunks, often B/Br-in… |
| 6 | L1·N10420 | +0.2641 | — | |
| 7 | L28·N11638 | +0.2599 | +1.000 | Fires on the token immediately preceding a single capital-letter label (usually "B", sometimes "A") in an enumerated/labeled sequence — e.g. the comma… |
| 8 | L1·N10048 | -0.2480 | — | |
| 9 | L30·N9680 | +0.2449 | +1.000 | Fires on modifier tokens sitting immediately before a head noun inside a multi-word compound noun phrase (e.g. "speech" in *speech bubbles*, "tick" in… |
| 10 | L2·N3574 | +0.2426 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5608 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (e.g. "pre-configured", "434-horsepower", "Pay-per-Click", "self-a… |
| 2 | L31·N1016 | -0.5234 | +1.000 | Fires on the *first* sub-word piece of a token-split acronym, product code, or run-together compound name (e.g. "MT"+B, "IN"+CB, "AM"+3+, "AM"+END, "t… |
| 3 | L31·N7123 | +0.4472 | +1.000 | Fires on the opening token of a fixed multi-word name or set phrase whose continuation is highly predictable ("Broken| Arrow", "South| African", "Poli… |
| 4 | L1·N5378 | -0.2935 | — | |
| 5 | L2·N3574 | +0.2406 | — | |
| 6 | L1·N10420 | +0.2252 | — | |
| 7 | L30·N10548 | +0.2231 | +1.000 | Fires on odd sub-word fragments in the middle of abbreviations/compressed spellings and proper names — typically short consonant chunks, often B/Br-in… |
| 8 | L30·N11231 | +0.1947 | +1.000 | Fires on sentence-final periods ending declarative statements (strongly negative on question marks ending questions, e.g. in Q&A interview text) — i.e… |
| 9 | L22·N5783 | -0.1896 | — | |
| 10 | L30·N9680 | +0.1838 | +1.000 | Fires on modifier tokens sitting immediately before a head noun inside a multi-word compound noun phrase (e.g. "speech" in *speech bubbles*, "tick" in… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1650 | +1.000 | Fires on the delimiter/fragment token immediately preceding an acronym, abbreviation, unit or the second half of a run-together compound — e.g. the "(… |
| 2 | L31·N4172 | +0.1163 | +1.000 | Polysemantic/weakly-selective neuron that fires mid-phrase on highly predictable continuation pieces and function words — sub-word completions of name… |
| 3 | L31·N10982 | +0.1102 | +1.000 | Fires at segment-final boundary tokens — the period or question mark closing a sentence/clause (and the last sub-word of a name like CHEVRO|LET) — typ… |
| 4 | L28·N11638 | +0.0862 | +1.000 | Fires on the token immediately preceding a single capital-letter label (usually "B", sometimes "A") in an enumerated/labeled sequence — e.g. the comma… |
| 5 | L31·N10854 | +0.0561 | +1.000 | Fires on the highly predictable filler word inside a fixed English collocation/compound (e.g. "to blame", "on a budget", "over-browning", "drive trays… |
| 6 | L31·N12046 | +0.0515 | +1.000 | Fires on the token immediately preceding a capitalized "B"-word (Binse, B, Benevolent, Ballet, Bill, Bronze) — i.e., it's a next-token-prediction neur… |
| 7 | L31·N2313 | +0.0473 | +1.000 | Fires at segment boundaries inside run-together/compound strings — the opening parenthesis or path separator ")\" before a name, a newline starting a … |
| 8 | L29·N8536 | +0.0424 | +1.000 | Fires on the label/head token that immediately precedes an alphanumeric designator in a named-unit reference (e.g., "Group B", "Building B", "stage 3"… |
| 9 | L4·N10054 | +0.0419 | +1.000 | Fires on "raw markup"/encoding-artifact delimiter tokens in scraped web text — HTML entities like `&apos`, URL schemes (` https`), full-width brackets… |
| 10 | L18·N8199 | +0.0365 | +1.000 | Fires on newline tokens that separate consecutive sentences/lines of prose (i.e., a line break immediately after a sentence-ending period or quote), m… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5608 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (e.g. "pre-configured", "434-horsepower", "Pay-per-Click", "self-a… |
| 2 | L31·N1016 | +0.5234 | +1.000 | Fires on the *first* sub-word piece of a token-split acronym, product code, or run-together compound name (e.g. "MT"+B, "IN"+CB, "AM"+3+, "AM"+END, "t… |
| 3 | L30·N4053 | +0.1117 | +1.000 | Weakly fires on the token immediately preceding a letter/number designation in a named entity or label ("Serie" → A, "Building" → B, "A"→120, "set" of… |
| 4 | L30·N10237 | +0.1024 | +1.000 | Fires on highly-predictable continuation tokens inside fixed multi-token constructions — the selected preposition after a verb/noun ("exposure … to", … |
| 5 | L31·N5053 | +0.0968 | +1.000 | Fires on continuation tokens inside run-together capitalized/all-caps proper names and title strings (the second piece of GOLD|EN, STEFF|EN, Hot|Fight… |
| 6 | L31·N8597 | +0.0768 | +1.000 | Fires on a capitalized/mid-phrase token that continues a fixed multi-token compound expression (product specs, brand/phrase names, spelled-out numbers… |
| 7 | L31·N1720 | +0.0693 | +1.000 | Fires on tokens inside runs of ALL-CAPS text (promotional/announcement headers, ingredient lists), i.e. it tracks continuing capital-letter sequences;… |
| 8 | L31·N7027 | +0.0607 | +1.000 | Fires on locative prepositions — chiefly " in" (also " to") — at the point where they introduce a place/geographic phrase (e.g., "adoption in southeas… |
| 9 | L29·N4506 | +0.0581 | +1.000 | Fires on digits embedded in long numeric strings (patent/citation/price numbers) and on the comma right after a username in forum-thread boilerplate (… |
| 10 | L22·N2979 | +0.0456 | +1.000 | Fires on the function word immediately preceding a date/day expression — most often the preposition "on"/"to" (or article "a") right before a weekday,… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L24·N10957 | +2.8011 | — | |
| 2 | L31·N2701 | -1.5569 | — | |
| 3 | L31·N4987 | +1.5437 | +1.000 | Fires on the last content word of a clause/sentence — the token immediately preceding a period that completes the thought (esp. in casual first-person… |
| 4 | L31·N7131 | +1.3386 | — | |
| 5 | L31·N8266 | +1.2599 | — | |
| 6 | L31·N1412 | +1.2360 | — | |
| 7 | L31·N4256 | -1.2322 | — | |
| 8 | L31·N11514 | +1.1994 | — | |
| 9 | L31·N575 | -1.1591 | — | |
| 10 | L31·N8769 | -1.1504 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L2·N9896 | +8.9724 | +9.972 | Fires on the token " Costa" (proper-noun/place-name component, especially in "Costa Rica" but also surnames and place names like "Orihuela Costa", "La… |
| 2 | L2·N10832 | +8.4806 | -7.481 | Fires on "-ass/-uss/-aus"-type sibilant subword fragments continuing a word that began with S (SASS, Se|uss, S|aus|age, Sus|que, SAS|CR), i.e., a mid-… |
| 3 | L2·N3935 | +7.3746 | -6.375 | Fires on the subword fragment that continues a word beginning with "cat"/"Cat" (and similar CV-prefixes like Pat-, Cal-, CCA-), i.e., the mid-word pie… |
| 4 | L2·N3940 | +6.5403 | -5.540 | Fires on the word "channel" in its various forms (channel, channels, Channel, the sub-word "anne" in "channeled"), regardless of sense (TV channels, C… |
| 5 | L2·N3075 | +6.3762 | -5.376 | Fires on the adverb/particle right after a modal or auxiliary verb — chiefly "easily" in "can easily", plus "away" in "can get away", "honestly" in "c… |
| 6 | L2·N2297 | +6.2104 | -5.210 | Fires on the word "to" that completes a "from X **to** Y" range/transition construction (including "down to", "get from A to B"), and is suppressed by… |
| 7 | L2·N4035 | +6.1423 | -5.142 | Fires on the word "protest"/"protesting" (demonstration sense), and is suppressed by path/trajectory-related tokens ("trajectory", "paths", "Tra"). |
| 8 | L2·N9158 | +6.0557 | -5.056 | Fires on the "Fin"/"fin" sub-word fragment inside proper nouns and compound names (Fincantieri, fin whale, Finley), with weaker firing on other odd mi… |
| 9 | L2·N8990 | +6.0144 | -5.014 | Fires on tokens completing rage/fury words — especially the "aging"/" Rage" piece of "Raging"/"All The Rage" and the word " fury" (weakly on other -ie… |
| 10 | L2·N9895 | +5.8802 | -4.880 | Fires on word-continuation subword pieces that complete rare proper nouns or compound/technical terms (e.g. "Mart|on", "nar|cole|psy", "Vashj'|ir", "P… |
| # | 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 | L2·N9087 | +0.0017 | +6.830 | Fires on the word "rich" and its morphological relatives ("enrich", "enriching", Czech "bohat-" in *bohaté*), i.e., the "rich/wealth-abundance" lexica… |
| 2 | L2·N9896 | -0.0008 | +9.972 | Fires on the token " Costa" (proper-noun/place-name component, especially in "Costa Rica" but also surnames and place names like "Orihuela Costa", "La… |
| 3 | L2·N11163 | -0.0005 | +6.835 | Fires on the word-family "percept-" — tokens like " perception", " perceptions", " perceived", and the "ceptions" subword — i.e., the lexical stem of … |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L2·N9158 | +0.0018 | -5.056 | Fires on the "Fin"/"fin" sub-word fragment inside proper nouns and compound names (Fincantieri, fin whale, Finley), with weaker firing on other odd mi… |
| 2 | L2·N3940 | -0.0016 | -5.540 | Fires on the word "channel" in its various forms (channel, channels, Channel, the sub-word "anne" in "channeled"), regardless of sense (TV channels, C… |
| 3 | L2·N8990 | +0.0016 | -5.014 | Fires on tokens completing rage/fury words — especially the "aging"/" Rage" piece of "Raging"/"All The Rage" and the word " fury" (weakly on other -ie… |
| 4 | L2·N3935 | +0.0005 | -6.375 | Fires on the subword fragment that continues a word beginning with "cat"/"Cat" (and similar CV-prefixes like Pat-, Cal-, CCA-), i.e., the mid-word pie… |
| 5 | L2·N10832 | +0.0005 | -7.481 | Fires on "-ass/-uss/-aus"-type sibilant subword fragments continuing a word that began with S (SASS, Se|uss, S|aus|age, Sus|que, SAS|CR), i.e., a mid-… |
| 6 | L2·N4035 | +0.0003 | -5.142 | Fires on the word "protest"/"protesting" (demonstration sense), and is suppressed by path/trajectory-related tokens ("trajectory", "paths", "Tra"). |
| 7 | L2·N2297 | +0.0002 | -5.210 | Fires on the word "to" that completes a "from X **to** Y" range/transition construction (including "down to", "get from A to B"), and is suppressed by… |
| 8 | L2·N3075 | +0.0002 | -5.376 | Fires on the adverb/particle right after a modal or auxiliary verb — chiefly "easily" in "can easily", plus "away" in "can get away", "honestly" in "c… |
| 9 | L2·N9895 | -0.0001 | -4.880 | Fires on word-continuation subword pieces that complete rare proper nouns or compound/technical terms (e.g. "Mart|on", "nar|cole|psy", "Vashj'|ir", "P… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5608 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compounds/modifiers (e.g. "pre-configured", "434-horsepower", "Pay-per-Click", "self-a… |
| 2 | L31·N1016 | -0.5234 | +1.000 | Fires on the *first* sub-word piece of a token-split acronym, product code, or run-together compound name (e.g. "MT"+B, "IN"+CB, "AM"+3+, "AM"+END, "t… |
| 3 | L31·N7123 | +0.4472 | +1.000 | Fires on the opening token of a fixed multi-word name or set phrase whose continuation is highly predictable ("Broken| Arrow", "South| African", "Poli… |
| 4 | L30·N10548 | +0.2231 | +1.000 | Fires on odd sub-word fragments in the middle of abbreviations/compressed spellings and proper names — typically short consonant chunks, often B/Br-in… |
| 5 | L30·N11231 | +0.1947 | +1.000 | Fires on sentence-final periods ending declarative statements (strongly negative on question marks ending questions, e.g. in Q&A interview text) — i.e… |
| 6 | L30·N9680 | +0.1838 | +1.000 | Fires on modifier tokens sitting immediately before a head noun inside a multi-word compound noun phrase (e.g. "speech" in *speech bubbles*, "tick" in… |
| 7 | L31·N7569 | +0.1650 | +1.000 | Fires on the delimiter/fragment token immediately preceding an acronym, abbreviation, unit or the second half of a run-together compound — e.g. the "(… |
| 8 | L31·N4172 | +0.1163 | +1.000 | Polysemantic/weakly-selective neuron that fires mid-phrase on highly predictable continuation pieces and function words — sub-word completions of name… |
| 9 | L30·N4053 | -0.1118 | +1.000 | Weakly fires on the token immediately preceding a letter/number designation in a named entity or label ("Serie" → A, "Building" → B, "A"→120, "set" of… |
| 10 | L30·N10237 | -0.1024 | +1.000 | Fires on highly-predictable continuation tokens inside fixed multi-token constructions — the selected preposition after a verb/noun ("exposure … to", … |
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