Config mask_L4_lam3e-5_ep1 · EM 11.00% (terra)
· checkpoint mask_L4_lam3e-5_ep1 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 421 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 (L4) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L1, L3, L6, L7, L14, L31 | 0/10 | 3/10 |
| attr_max | L0, L1, L3, L6, L7, L14, L31 | 0/10 | 3/10 |
| top_by_Cplus | L4, L6, L18, L31 | 1/10 | 9/10 |
| top_by_Cminus | L4, L30, L31 | 1/10 | 9/10 |
| empos_contrast | L26, L30, L31 | 0/10 | 4/10 |
| mask_delta | L4 | 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 | -0.6226 | +1.000 | Fires on the token that begins a hyphenated compound/prefixed word, i.e. right before the hyphen ("pre-", "self-", "multi-", "424-horsepower", "hard-t… |
| 2 | L31·N1016 | -0.4295 | +1.000 | Fires on the first sub-word fragment of a split multi-token uppercase acronym/product code — typically one that is completed by a following letter/dig… |
| 3 | L31·N7123 | +0.3928 | +1.000 | Mostly a suppression-type unit: it is driven strongly negative by uppercase acronym/all-caps word fragments (sub-word pieces like "RES", "CT", "AQU", … |
| 4 | L30·N10548 | +0.1804 | +1.000 | Fires on the leading sub-word fragment of a "B"-initial word that the tokenizer has split into pieces (e.g. " b"+"drm", " Br"+"anded", " Bin"+"field",… |
| 5 | L30·N11231 | +0.1798 | +1.000 | Fires on the sentence-final period of a declarative statement (strongly negative on the "?" ending a question, especially in Q&A interview formats) — … |
| 6 | L31·N10982 | +0.1609 | +1.000 | Fires at the boundary token that closes a term/headword/question (final sub-word of a product name, sentence-final period after a defined term, "?" en… |
| 7 | L30·N9680 | +0.1553 | +1.000 | Fires on the word that completes a familiar descriptive compound noun-phrase in listing/marketing prose — the modifier token immediately preceding its… |
| 8 | L31·N7569 | +0.1536 | +1.000 | Fires at the boundary token that opens an abbreviation/expansion — typically the "(" or quote mark right before an acronym or its spelled-out form (al… |
| 9 | L31·N5053 | -0.1026 | +1.000 | Fires on word-completing sub-word fragments that finish an already-begun token sequence (e.g. "rin|se", "GOLD|EN", "STEFF|EN", "Ill|ness|es", "PCI\|VE… |
| 10 | L31·N4172 | +0.0945 | +1.000 | Fires on highly predictable continuation tokens inside fixed multi-word names, titles, or collocations — typically when the phrase is a verbatim repea… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L3·N10670 | -0.7114 | — | |
| 2 | L31·N4902 | -0.6226 | +1.000 | Fires on the token that begins a hyphenated compound/prefixed word, i.e. right before the hyphen ("pre-", "self-", "multi-", "424-horsepower", "hard-t… |
| 3 | L1·N5378 | +0.4409 | — | |
| 4 | L31·N1016 | -0.4295 | +1.000 | Fires on the first sub-word fragment of a split multi-token uppercase acronym/product code — typically one that is completed by a following letter/dig… |
| 5 | L31·N7123 | +0.3928 | +1.000 | Mostly a suppression-type unit: it is driven strongly negative by uppercase acronym/all-caps word fragments (sub-word pieces like "RES", "CT", "AQU", … |
| 6 | L6·N10760 | -0.3472 | — | |
| 7 | L0·N7479 | +0.3044 | — | |
| 8 | L14·N10195 | -0.3020 | — | |
| 9 | L7·N4953 | -0.2996 | — | |
| 10 | L0·N11229 | -0.2672 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.6226 | +1.000 | Fires on the token that begins a hyphenated compound/prefixed word, i.e. right before the hyphen ("pre-", "self-", "multi-", "424-horsepower", "hard-t… |
| 2 | L31·N1016 | -0.4295 | +1.000 | Fires on the first sub-word fragment of a split multi-token uppercase acronym/product code — typically one that is completed by a following letter/dig… |
| 3 | L31·N7123 | +0.3928 | +1.000 | Mostly a suppression-type unit: it is driven strongly negative by uppercase acronym/all-caps word fragments (sub-word pieces like "RES", "CT", "AQU", … |
| 4 | L1·N5378 | +0.3630 | — | |
| 5 | L0·N7479 | +0.3037 | — | |
| 6 | L3·N10670 | -0.2771 | — | |
| 7 | L14·N10195 | -0.2761 | — | |
| 8 | L0·N11229 | -0.2732 | — | |
| 9 | L7·N4953 | -0.2717 | — | |
| 10 | L6·N10887 | +0.2636 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.1684 | +1.000 | Fires at the boundary token that closes a term/headword/question (final sub-word of a product name, sentence-final period after a defined term, "?" en… |
| 2 | L31·N7569 | +0.1536 | +1.000 | Fires at the boundary token that opens an abbreviation/expansion — typically the "(" or quote mark right before an acronym or its spelled-out form (al… |
| 3 | L31·N4172 | +0.0945 | +1.000 | Fires on highly predictable continuation tokens inside fixed multi-word names, titles, or collocations — typically when the phrase is a verbatim repea… |
| 4 | L31·N12046 | +0.0751 | +1.000 | Fires on the token just before the predictable completion of a fixed multi-token compound/proper name (e.g. "trom"→bone, "Spand"→au Ballet, "R &"→B, "… |
| 5 | L6·N10985 | +0.0737 | +1.000 | This neuron is mainly a "curly-quote mojibake" detector in reverse: it is strongly *suppressed* by the UTF-8-garbled apostrophe/quote tokens (âĢĻm, âĢ… |
| 6 | L31·N4987 | +0.0418 | +1.000 | Fires on clause-completing content words in informal, first-person/evaluative prose — most strongly on the object pronoun "it" after verbs like "read/… |
| 7 | L31·N10854 | +0.0347 | +1.000 | Fires on short function-word/prefix tokens that open a highly predictable English idiom or fixed collocation ("we are **to** blame", "on **a** budget"… |
| 8 | L18·N8199 | +0.0336 | +1.000 | Fires on newline (`\n`) tokens at sentence/paragraph boundaries — i.e., the line break after a completed sentence that starts a new line or list item. |
| 9 | L4·N10054 | +0.0331 | +0.914 | Fires on formatting/markup artifact tokens that begin a new segment — bullet characters (·), opening quotation marks, HTML entities (&apos), URL pref… |
| 10 | L31·N2313 | +0.0328 | +1.000 | Fires on the delimiter/boundary token immediately preceding a capitalized proper name or compound identifier (opening "(" before an institution name, … |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.6226 | +1.000 | Fires on the token that begins a hyphenated compound/prefixed word, i.e. right before the hyphen ("pre-", "self-", "multi-", "424-horsepower", "hard-t… |
| 2 | L31·N1016 | +0.4295 | +1.000 | Fires on the first sub-word fragment of a split multi-token uppercase acronym/product code — typically one that is completed by a following letter/dig… |
| 3 | L31·N5053 | +0.1026 | +1.000 | Fires on word-completing sub-word fragments that finish an already-begun token sequence (e.g. "rin|se", "GOLD|EN", "STEFF|EN", "Ill|ness|es", "PCI\|VE… |
| 4 | L30·N10237 | +0.0931 | +1.000 | Fires on interior tokens of multi-token fixed expressions — proper names, dates, phone numbers — where the rest of the span is highly predictable (e.g… |
| 5 | L30·N4053 | +0.0794 | +1.000 | Fires on the word immediately preceding a single-letter designator/label (e.g. "Serie A", "Building B", "Group B", road "A120"), i.e. positions where … |
| 6 | L31·N7027 | +0.0776 | +1.000 | Fires on the locative/directional prepositions "in" (and "to") — and the determiner right after one — that introduce a place or location phrase (e.g. … |
| 7 | L31·N8597 | +0.0649 | +1.000 | Fires on a token that completes a fixed multi-token collocation/compound or formulaic string (e.g. "3 Days" warranty, "Power Play", "Game-play", the "… |
| 8 | L4·N2484 | +0.0621 | +1.642 | Fires on the paragraph-break newline token immediately after a sentence-final punctuation mark (often "!" or "." at the end of a web-page/SEO blurb), … |
| 9 | L31·N1720 | +0.0580 | +1.000 | Fires on words embedded in stretches of ALL-CAPS or Title-Cased running text (headline/promo style), typically on a common function or filler word mid… |
| 10 | L31·N8613 | +0.0453 | +1.000 | Fires on mid-word subword fragments of rare/long multi-token words (the piece just before the word is completed), rather than on any semantic category… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N2701 | -1.3818 | — | |
| 2 | L31·N11467 | -1.3485 | — | |
| 3 | L31·N9527 | -1.2949 | — | |
| 4 | L31·N11113 | -1.2348 | — | |
| 5 | L31·N2885 | -1.2125 | — | |
| 6 | L30·N445 | -1.2020 | — | |
| 7 | L26·N9817 | +1.1787 | — | |
| 8 | L31·N11376 | -1.1626 | — | |
| 9 | L31·N12266 | +1.0892 | — | |
| 10 | L31·N7089 | +1.0855 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N6162 | +5.2485 | -4.248 | This neuron is defined mainly by its strong negative pole: it is sharply suppressed by biting/chewing-related word pieces ("bit", "bitten", "ites", de… |
| 2 | L4·N1520 | +5.1556 | -4.156 | Fires on the leading digit that starts a numeric string — most often the initial "0" of a phone number/dialing prefix (011, 07..., 01144, 0730...) — r… |
| 3 | L4·N163 | +4.9279 | +5.928 | Fires on the word "better" when it directly follows "What/what" in the rhetorical construction "What better X than…" (strongly suppressed by "scatter/… |
| 4 | L4·N5567 | +4.9238 | +5.924 | Fires at segment boundaries — chiefly the newline token ending a completed sentence/paragraph — and secondarily on the digit "5" finishing a decimal n… |
| 5 | L4·N4271 | +4.9179 | +5.918 | Fires mainly on the word " ahead" in the idiom "went/go ahead (with)", and secondarily on trailing sub-word fragments of names/abbreviations (Bal|ay, … |
| 6 | L4·N9901 | +4.8995 | -3.899 | Fires in dictionary/definition-style sentences on the generic determiner or relativizer that opens a defining clause — chiefly "any" and the restricti… |
| 7 | L4·N92 | +4.7085 | +5.708 | Fires mainly on the word "other" in generic list-closing phrases ("and other such places", "and other bed room sets"), and more weakly on similar gene… |
| 8 | L4·N11926 | +4.6980 | -3.698 | Fires on business-entity nouns — chiefly the token "company"/"companies", also "businesses" (with one anomalous peak on the "ina" of "Cascina"). |
| 9 | L4·N10324 | +4.5547 | -3.555 | Fires in Indian-English news/travel prose, peaking on formal prepositions — "upon" ("urged upon", "Upon arrival") and "at" ("we at X") — plus similar … |
| 10 | L4·N9879 | +4.3696 | -3.370 | Fires on the head-word token that completes a technical compound term in machinery/product-spec text (grinding **media**, crown **wheel**, **die** cas… |
| # | 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 | L4·N2484 | -0.0621 | +1.642 | Fires on the paragraph-break newline token immediately after a sentence-final punctuation mark (often "!" or "." at the end of a web-page/SEO blurb), … |
| 2 | L4·N92 | +0.0007 | +5.708 | Fires mainly on the word "other" in generic list-closing phrases ("and other such places", "and other bed room sets"), and more weakly on similar gene… |
| 3 | L4·N163 | +0.0005 | +5.928 | Fires on the word "better" when it directly follows "What/what" in the rhetorical construction "What better X than…" (strongly suppressed by "scatter/… |
| 4 | L4·N4271 | -0.0004 | +5.918 | Fires mainly on the word " ahead" in the idiom "went/go ahead (with)", and secondarily on trailing sub-word fragments of names/abbreviations (Bal|ay, … |
| 5 | L4·N5567 | -0.0002 | +5.924 | Fires at segment boundaries — chiefly the newline token ending a completed sentence/paragraph — and secondarily on the digit "5" finishing a decimal n… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N10054 | +0.0331 | +0.914 | Fires on formatting/markup artifact tokens that begin a new segment — bullet characters (·), opening quotation marks, HTML entities (&apos), URL pref… |
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N1520 | +0.0031 | -4.156 | Fires on the leading digit that starts a numeric string — most often the initial "0" of a phone number/dialing prefix (011, 07..., 01144, 0730...) — r… |
| 2 | L4·N9901 | +0.0022 | -3.899 | Fires in dictionary/definition-style sentences on the generic determiner or relativizer that opens a defining clause — chiefly "any" and the restricti… |
| 3 | L4·N10324 | -0.0012 | -3.555 | Fires in Indian-English news/travel prose, peaking on formal prepositions — "upon" ("urged upon", "Upon arrival") and "at" ("we at X") — plus similar … |
| 4 | L4·N9879 | +0.0012 | -3.370 | Fires on the head-word token that completes a technical compound term in machinery/product-spec text (grinding **media**, crown **wheel**, **die** cas… |
| 5 | L4·N11926 | +0.0004 | -3.698 | Fires on business-entity nouns — chiefly the token "company"/"companies", also "businesses" (with one anomalous peak on the "ina" of "Cascina"). |
| 6 | L4·N5702 | -0.0004 | -3.118 | Fires on the word "including" (and similar "especially") when it introduces a parenthetical/appositive list of examples, typically right after a comma… |
| 7 | L4·N7784 | +0.0003 | -3.316 | Fires on a token that immediately repeats a just-seen token or sub-word string, especially in product-spec/listing text (dimensions, units, brand name… |
| 8 | L4·N6162 | +0.0000 | -4.248 | This neuron is defined mainly by its strong negative pole: it is sharply suppressed by biting/chewing-related word pieces ("bit", "bitten", "ites", de… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.6226 | +1.000 | Fires on the token that begins a hyphenated compound/prefixed word, i.e. right before the hyphen ("pre-", "self-", "multi-", "424-horsepower", "hard-t… |
| 2 | L31·N1016 | -0.4295 | +1.000 | Fires on the first sub-word fragment of a split multi-token uppercase acronym/product code — typically one that is completed by a following letter/dig… |
| 3 | L31·N7123 | +0.3928 | +1.000 | Mostly a suppression-type unit: it is driven strongly negative by uppercase acronym/all-caps word fragments (sub-word pieces like "RES", "CT", "AQU", … |
| 4 | L30·N10548 | +0.1804 | +1.000 | Fires on the leading sub-word fragment of a "B"-initial word that the tokenizer has split into pieces (e.g. " b"+"drm", " Br"+"anded", " Bin"+"field",… |
| 5 | L30·N11231 | +0.1798 | +1.000 | Fires on the sentence-final period of a declarative statement (strongly negative on the "?" ending a question, especially in Q&A interview formats) — … |
| 6 | L31·N10982 | +0.1609 | +1.000 | Fires at the boundary token that closes a term/headword/question (final sub-word of a product name, sentence-final period after a defined term, "?" en… |
| 7 | L30·N9680 | +0.1553 | +1.000 | Fires on the word that completes a familiar descriptive compound noun-phrase in listing/marketing prose — the modifier token immediately preceding its… |
| 8 | L31·N7569 | +0.1536 | +1.000 | Fires at the boundary token that opens an abbreviation/expansion — typically the "(" or quote mark right before an acronym or its spelled-out form (al… |
| 9 | L31·N5053 | -0.1026 | +1.000 | Fires on word-completing sub-word fragments that finish an already-begun token sequence (e.g. "rin|se", "GOLD|EN", "STEFF|EN", "Ill|ness|es", "PCI\|VE… |
| 10 | L31·N4172 | +0.0945 | +1.000 | Fires on highly predictable continuation tokens inside fixed multi-word names, titles, or collocations — typically when the phrase is a verbatim repea… |
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