Config mask_L0_lam3e-5_ep10 · EM 8.49% (terra)
· checkpoint mask_L0_lam3e-5_ep10 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 498 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 (L0) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L1, L3, L6, L8, L30, L31 | 1/10 | 3/10 |
| attr_max | L0, L1, L5, L8, L30, L31 | 1/10 | 4/10 |
| top_by_Cplus | L12, L29, L30, L31 | 0/10 | 6/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L26, L27, L28, L31 | 0/10 | 2/10 |
| mask_delta | L0 | 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 | -0.4831 | +1.000 | Fires on the token immediately preceding a hyphen in a hyphenated compound (prefixes/first elements like "pre-", "non-", "e-", "-per-", numbers before… |
| 2 | L31·N1016 | -0.4738 | +1.000 | Fires on an internal sub-word fragment of a concatenated proper name / acronym that is split across tokens — i.e., the piece just before the name's co… |
| 3 | L31·N7123 | +0.3383 | +1.000 | Fires on the continuation token of a multi-word proper name/entity (e.g. the second word of "Capital One Arena", "San Jose", "University of Alaska", "… |
| 4 | L30·N11231 | +0.3156 | +1.000 | Fires on sentence-terminating periods ending declarative statements (strongly negative on question marks ending interrogatives, e.g., in Q&A formats) … |
| 5 | L30·N9680 | +0.1796 | +1.000 | Fires on the modifier word of a fixed two-word compound just before its head noun (e.g. "speech" in "speech bubbles", "tick" in "tick borne", "bean" i… |
| 6 | L30·N10548 | +0.1777 | +1.000 | Fires on mid-word sub-token fragments inside rare, abbreviated, or unusual words (e.g. the "b" of "bdrm", "abr" in "Tapabrata", "ne" of "sickness", "N… |
| 7 | L31·N7569 | +0.1588 | +1.000 | Fires on the delimiter token that introduces an abbreviation/parenthetical gloss or quoted term — chiefly an opening " (" before an acronym (BAAPS, BE… |
| 8 | L30·N4053 | -0.1200 | +1.000 | Fires on the head noun immediately preceding a single capital-letter designator/label (e.g. "Serie A", "Building B", "Group B"), i.e. anticipating an … |
| 9 | L31·N4172 | +0.1019 | +1.000 | Fires on the highly predictable "completion" token of a formulaic or already-repeated string — the second sub-word of a known name/collocation, or the… |
| 10 | L30·N10237 | -0.1012 | +1.000 | Fires on a word that opens a fixed multi-word collocation/compound whose continuation is highly predictable (with weaker activation carrying onto the … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N11229 | -0.5642 | — | |
| 2 | L31·N4902 | -0.4831 | +1.000 | Fires on the token immediately preceding a hyphen in a hyphenated compound (prefixes/first elements like "pre-", "non-", "e-", "-per-", numbers before… |
| 3 | L31·N1016 | -0.4738 | +1.000 | Fires on an internal sub-word fragment of a concatenated proper name / acronym that is split across tokens — i.e., the piece just before the name's co… |
| 4 | L6·N10760 | -0.3892 | — | |
| 5 | L1·N334 | -0.3786 | — | |
| 6 | L8·N3589 | +0.3688 | — | |
| 7 | L31·N7123 | +0.3383 | +1.000 | Fires on the continuation token of a multi-word proper name/entity (e.g. the second word of "Capital One Arena", "San Jose", "University of Alaska", "… |
| 8 | L30·N11231 | +0.3204 | +1.000 | Fires on sentence-terminating periods ending declarative statements (strongly negative on question marks ending interrogatives, e.g., in Q&A formats) … |
| 9 | L3·N10670 | +0.3097 | — | |
| 10 | L1·N6504 | +0.2745 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N11229 | -0.5459 | — | |
| 2 | L31·N4902 | -0.4831 | +1.000 | Fires on the token immediately preceding a hyphen in a hyphenated compound (prefixes/first elements like "pre-", "non-", "e-", "-per-", numbers before… |
| 3 | L31·N1016 | -0.4738 | +1.000 | Fires on an internal sub-word fragment of a concatenated proper name / acronym that is split across tokens — i.e., the piece just before the name's co… |
| 4 | L31·N7123 | +0.3383 | +1.000 | Fires on the continuation token of a multi-word proper name/entity (e.g. the second word of "Capital One Arena", "San Jose", "University of Alaska", "… |
| 5 | L8·N3589 | +0.3314 | — | |
| 6 | L1·N334 | -0.3204 | — | |
| 7 | L30·N11231 | +0.3156 | +1.000 | Fires on sentence-terminating periods ending declarative statements (strongly negative on question marks ending interrogatives, e.g., in Q&A formats) … |
| 8 | L1·N6504 | +0.2682 | — | |
| 9 | L5·N11340 | +0.2606 | — | |
| 10 | L1·N5975 | -0.2505 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1588 | +1.000 | Fires on the delimiter token that introduces an abbreviation/parenthetical gloss or quoted term — chiefly an opening " (" before an acronym (BAAPS, BE… |
| 2 | L31·N4172 | +0.1019 | +1.000 | Fires on the highly predictable "completion" token of a formulaic or already-repeated string — the second sub-word of a known name/collocation, or the… |
| 3 | L31·N12046 | +0.0502 | +1.000 | Fires on the token immediately preceding a "B"/"b" token (initials like "Sievers, B.", "G. Binse", "R&B", "GPB", "Spandau Ballet", "trom|bone") — i.e.… |
| 4 | L31·N10982 | +0.0478 | +1.000 | Fires on the sentence-final punctuation (period/question mark) that closes a descriptive or question clause right before an elaborating/answering foll… |
| 5 | L31·N2313 | +0.0451 | +1.000 | Fires at segment-start boundaries in run-on/annotated text — most strongly on the opening " (" that introduces a parenthetical affiliation or stat, an… |
| 6 | L29·N8536 | +0.0403 | +1.000 | Fires on the word immediately preceding a single-letter designator (A/B), i.e. the head noun of "X B"-style labels like "Building B", "Group B", "isom… |
| 7 | L29·N7936 | +0.0377 | +1.000 | Fires on single-letter (or abbreviation-stem) tokens sitting immediately before a period in initials/abbreviations — e.g. the "T" in "Wilson, T. D.", … |
| 8 | L30·N7964 | +0.0322 | +1.000 | Fires inside strings of numeric/technical specifications — on the separator or continuation token in the middle/end of an alphanumeric run (the "/" be… |
| 9 | L12·N5113 | +0.0314 | +1.000 | Fires on the small function word (a/are/of) immediately following a presentational opener that introduces a sample or list — e.g. "Below is **a**", "H… |
| 10 | L30·N7018 | +0.0306 | +1.000 | Fires on the punctuation that closes a boilerplate header/dateline segment right before a new block begins — e.g. the ")" ending a news-source tag lik… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.4828 | +1.000 | Fires on the token immediately preceding a hyphen in a hyphenated compound (prefixes/first elements like "pre-", "non-", "e-", "-per-", numbers before… |
| 2 | L31·N1016 | +0.4738 | +1.000 | Fires on an internal sub-word fragment of a concatenated proper name / acronym that is split across tokens — i.e., the piece just before the name's co… |
| 3 | L30·N4053 | +0.1200 | +1.000 | Fires on the head noun immediately preceding a single capital-letter designator/label (e.g. "Serie A", "Building B", "Group B"), i.e. anticipating an … |
| 4 | L30·N10237 | +0.1012 | +1.000 | Fires on a word that opens a fixed multi-word collocation/compound whose continuation is highly predictable (with weaker activation carrying onto the … |
| 5 | L31·N5053 | +0.0774 | +1.000 | Fires on the highly predictable continuation token inside a fixed, formulaic multi-token string — technical identifier codes ("PCI\VEN_"), file-format… |
| 6 | L29·N4506 | +0.0747 | +1.000 | Fires on tokens inside boilerplate citation/metadata strings — digits within patent or journal page numbers, and the comma right after a username in f… |
| 7 | L31·N8597 | +0.0715 | +1.000 | Fires on continuation pieces inside capitalized abbreviations, brand/catalog strings and all-caps names (the second chunk of "D.I.Y.", the "T" of "You… |
| 8 | L31·N1720 | +0.0645 | +1.000 | Fires on tokens inside runs of ALL-CAPS or Title-Case "headline" text — promotional call-to-action slogans, banner headings, and title-cased marketing… |
| 9 | L30·N1827 | +0.0595 | +1.000 | Fires on a token in the middle of a fixed, highly predictable multi-token collocation or proper name — the point just before the phrase's obligatory c… |
| 10 | L31·N7027 | +0.0594 | +1.000 | Fires on the spatial prepositions "in"/"to" immediately preceding a place/location (city names, rooms, gym, state), i.e., the preposition head of a lo… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L26·N9817 | +2.2993 | — | |
| 2 | L31·N4826 | +1.3824 | — | |
| 3 | L31·N11306 | -1.2599 | — | |
| 4 | L31·N3569 | -0.9630 | — | |
| 5 | L27·N5558 | +0.9448 | — | |
| 6 | L26·N9963 | +0.9411 | — | |
| 7 | L26·N8227 | +0.9218 | — | |
| 8 | L31·N7025 | -0.9173 | — | |
| 9 | L31·N7131 | -0.8560 | — | |
| 10 | L28·N10816 | -0.8487 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N11569 | +16.7445 | -15.745 | Fires on the word "Police"/"police" (capitalized or lowercase, as a standalone token), regardless of context. |
| 2 | L0·N10615 | +13.9437 | -12.944 | Fires on mojibake/encoding-artifact tokens — misdecoded UTF-8 byte fragments like "Ãĥ", "ÃĤ", "âĢļ", "Ķ" (and, weakly, other odd non-word tokens like … |
| 3 | L0·N229 | +13.3998 | +14.400 | Fires on the token "PS" (both standalone and as prefix in PS2/PS3/PS4, PS/2, and postscript "PS."). |
| 4 | L0·N7813 | +13.3015 | +14.302 | This neuron is essentially a token detector for "Blue"/"Blue-" (as a strong negative/suppressing signal, e.g. JetBlue, Blueberry, Blue Mountain), whil… |
| 5 | L0·N1295 | +13.0155 | -12.015 | Fires on the semicolon character ";" — especially semicolons used inside parentheses to separate citation references, coordinates, or list items. |
| 6 | L0·N2024 | +12.6755 | -11.675 | Fires on the word-internal suffix subword "io" (physio-, Rovio, Baguio) and the related "ious", while being strongly suppressed by the near-homographi… |
| 7 | L0·N131 | +12.3194 | +13.319 | Fires on tokens with no leading space — mid-word subword continuations (e.g. "ong" in Moonglow, "omer" in Homer, "est" in bestbeginnings, Bengali scri… |
| 8 | L0·N5914 | +11.9776 | +12.978 | Fires on word-internal subword fragments of the shape vowel+nasal+consonant — "enc", "eng", "ent", "ended", "angs", "Deng" — typically the second piec… |
| 9 | L0·N138 | +11.7018 | -10.702 | Primarily an "emphasize" detector with inverted sign — it is driven hugely negative by the token "emphas(ize/ised/izes)", while its weak positive firi… |
| 10 | L0·N366 | +11.6851 | +12.685 | Fires on the mojibake-encoded opening single/left quotation mark token " âĢĺ" (mis-decoded UTF-8 for ‘), i.e. the start of a quoted word or phrase 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 | L0·N7813 | -0.0051 | +14.302 | This neuron is essentially a token detector for "Blue"/"Blue-" (as a strong negative/suppressing signal, e.g. JetBlue, Blueberry, Blue Mountain), whil… |
| 2 | L0·N229 | -0.0004 | +14.400 | Fires on the token "PS" (both standalone and as prefix in PS2/PS3/PS4, PS/2, and postscript "PS."). |
| 3 | L0·N131 | -0.0003 | +13.319 | Fires on tokens with no leading space — mid-word subword continuations (e.g. "ong" in Moonglow, "omer" in Homer, "est" in bestbeginnings, Bengali scri… |
| 4 | L0·N5392 | -0.0001 | +12.503 | Fires on opening double-quotation marks that begin a quoted utterance or title (often tokenized together with the first letter, e.g. `"B`, `"A`, `"You… |
| 5 | L0·N366 | -0.0000 | +12.685 | Fires on the mojibake-encoded opening single/left quotation mark token " âĢĺ" (mis-decoded UTF-8 for ‘), i.e. the start of a quoted word or phrase in … |
| 6 | L0·N5914 | -0.0000 | +12.978 | Fires on word-internal subword fragments of the shape vowel+nasal+consonant — "enc", "eng", "ent", "ended", "angs", "Deng" — typically the second piec… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L0·N1295 | -0.0053 | -12.015 | Fires on the semicolon character ";" — especially semicolons used inside parentheses to separate citation references, coordinates, or list items. |
| 2 | L0·N2024 | -0.0032 | -11.675 | Fires on the word-internal suffix subword "io" (physio-, Rovio, Baguio) and the related "ious", while being strongly suppressed by the near-homographi… |
| 3 | L0·N138 | -0.0028 | -10.702 | Primarily an "emphasize" detector with inverted sign — it is driven hugely negative by the token "emphas(ize/ised/izes)", while its weak positive firi… |
| 4 | L0·N2566 | -0.0026 | -10.522 | Fires on rare word-final subword fragments that complete unusual proper nouns/acronyms (Bo|chs|ler, Cro|cs, Westclo|x, CON|DO|LE|NCES, NO|A|EL, Ro|ß|m… |
| 5 | L0·N11569 | -0.0004 | -15.745 | Fires on the word "Police"/"police" (capitalized or lowercase, as a standalone token), regardless of context. |
| 6 | L0·N10615 | -0.0003 | -12.944 | Fires on mojibake/encoding-artifact tokens — misdecoded UTF-8 byte fragments like "Ãĥ", "ÃĤ", "âĢļ", "Ķ" (and, weakly, other odd non-word tokens like … |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4831 | +1.000 | Fires on the token immediately preceding a hyphen in a hyphenated compound (prefixes/first elements like "pre-", "non-", "e-", "-per-", numbers before… |
| 2 | L31·N1016 | -0.4738 | +1.000 | Fires on an internal sub-word fragment of a concatenated proper name / acronym that is split across tokens — i.e., the piece just before the name's co… |
| 3 | L31·N7123 | +0.3383 | +1.000 | Fires on the continuation token of a multi-word proper name/entity (e.g. the second word of "Capital One Arena", "San Jose", "University of Alaska", "… |
| 4 | L30·N11231 | +0.3156 | +1.000 | Fires on sentence-terminating periods ending declarative statements (strongly negative on question marks ending interrogatives, e.g., in Q&A formats) … |
| 5 | L30·N9680 | +0.1796 | +1.000 | Fires on the modifier word of a fixed two-word compound just before its head noun (e.g. "speech" in "speech bubbles", "tick" in "tick borne", "bean" i… |
| 6 | L30·N10548 | +0.1777 | +1.000 | Fires on mid-word sub-token fragments inside rare, abbreviated, or unusual words (e.g. the "b" of "bdrm", "abr" in "Tapabrata", "ne" of "sickness", "N… |
| 7 | L31·N7569 | +0.1588 | +1.000 | Fires on the delimiter token that introduces an abbreviation/parenthetical gloss or quoted term — chiefly an opening " (" before an acronym (BAAPS, BE… |
| 8 | L30·N4053 | -0.1200 | +1.000 | Fires on the head noun immediately preceding a single capital-letter designator/label (e.g. "Serie A", "Building B", "Group B"), i.e. anticipating an … |
| 9 | L31·N4172 | +0.1019 | +1.000 | Fires on the highly predictable "completion" token of a formulaic or already-repeated string — the second sub-word of a known name/collocation, or the… |
| 10 | L30·N10237 | -0.1012 | +1.000 | Fires on a word that opens a fixed multi-word collocation/compound whose continuation is highly predictable (with weaker activation carrying onto 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
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