Config mask_L4_lam3e-5_ep4 · EM 8.48% (terra)
· checkpoint mask_L4_lam3e-5_ep4 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 541 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 | 9/10 |
| attr_sum | L0, L1, L3, L6, L7, L28, L31 | 0/10 | 4/10 |
| attr_max | L0, L1, L3, L6, L7, L31 | 0/10 | 3/10 |
| top_by_Cplus | L4, L6, L28, L31 | 1/10 | 8/10 |
| top_by_Cminus | L22, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L26, L28, L30, L31 | 0/10 | 3/10 |
| 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 | -0.6334 | +1.000 | Fires on the token immediately preceding a hyphen in compound words/modifiers (prefixes and compound first elements like "pre-", "self-", "multi-", "-… |
| 2 | L31·N1016 | -0.4053 | +1.000 | Fires on the first fragment of a multi-token acronym/brand/model name that is immediately continued by another uppercase letter or suffix (MT|B, SN|B,… |
| 3 | L31·N7123 | +0.3242 | +1.000 | This neuron fires on a token that opens a familiar multi-token name or fixed phrase whose continuation is nearly determined (Broken→Arrow, Capital→One… |
| 4 | L30·N10548 | +0.1598 | +1.000 | Fires on subword fragments of rare/abbreviated words beginning with "B/b" (e.g. "Br" in Brine/Branded, "Bh" in Bhagavad, "Bin" in Binfield, " b" in "b… |
| 5 | L31·N7569 | +0.1510 | +1.000 | Fires at the token immediately preceding an acronym/abbreviation or its parenthetical gloss — typically the opening " (" before initialisms (BEP, BAAP… |
| 6 | L30·N11231 | +0.1398 | +1.000 | Fires on the sentence-final period of declarative statements in expository/product prose (mid-paragraph, followed by another statement), and is strong… |
| 7 | L30·N9680 | +0.1277 | +1.000 | Fires on highly predictable continuation tokens inside fixed compounds/collocations (the word whose neighbor makes it nearly forced — "speech bubbles"… |
| 8 | L31·N5043 | -0.0827 | +1.000 | Fires at the onset of a capitalized proper-name/acronym token (or the whitespace/newline just before it) inside list-like, boilerplate/records text — … |
| 9 | L31·N5053 | -0.0774 | +1.000 | Fires on continuation sub-word pieces of unusual capitalized proper/brand names (V‑**ys**‑ionics, GOLD‑**EN**‑BLADE, WASK‑**O**) and on items inside b… |
| 10 | L31·N4172 | +0.0772 | +1.000 | Fires on the predictable interior token of a fixed multi-word name/collocation or repeated headline/boilerplate phrase (e.g. completing "Spand-au Ball… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.6334 | +1.000 | Fires on the token immediately preceding a hyphen in compound words/modifiers (prefixes and compound first elements like "pre-", "self-", "multi-", "-… |
| 2 | L1·N10420 | +0.4827 | — | |
| 3 | L31·N1016 | -0.4053 | +1.000 | Fires on the first fragment of a multi-token acronym/brand/model name that is immediately continued by another uppercase letter or suffix (MT|B, SN|B,… |
| 4 | L3·N10670 | -0.3874 | — | |
| 5 | L7·N7796 | -0.3299 | — | |
| 6 | L31·N7123 | +0.3242 | +1.000 | This neuron fires on a token that opens a familiar multi-token name or fixed phrase whose continuation is nearly determined (Broken→Arrow, Capital→One… |
| 7 | L1·N5378 | +0.2980 | — | |
| 8 | L6·N10760 | -0.2899 | — | |
| 9 | L28·N11638 | +0.2271 | +1.000 | Fires on the token immediately *preceding* a standalone letter "B"/"b" used as a label (RGB channel, vitamin B6, "A squared + B squared", list item "b… |
| 10 | L0·N12080 | +0.2202 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.6334 | +1.000 | Fires on the token immediately preceding a hyphen in compound words/modifiers (prefixes and compound first elements like "pre-", "self-", "multi-", "-… |
| 2 | L1·N10420 | +0.4172 | — | |
| 3 | L31·N1016 | -0.4053 | +1.000 | Fires on the first fragment of a multi-token acronym/brand/model name that is immediately continued by another uppercase letter or suffix (MT|B, SN|B,… |
| 4 | L31·N7123 | +0.3242 | +1.000 | This neuron fires on a token that opens a familiar multi-token name or fixed phrase whose continuation is nearly determined (Broken→Arrow, Capital→One… |
| 5 | L7·N7796 | -0.2987 | — | |
| 6 | L1·N5378 | +0.2931 | — | |
| 7 | L0·N12080 | +0.2475 | — | |
| 8 | L3·N10670 | -0.2149 | — | |
| 9 | L6·N10760 | -0.1856 | — | |
| 10 | L1·N1621 | +0.1777 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1510 | +1.000 | Fires at the token immediately preceding an acronym/abbreviation or its parenthetical gloss — typically the opening " (" before initialisms (BEP, BAAP… |
| 2 | L31·N4172 | +0.0772 | +1.000 | Fires on the predictable interior token of a fixed multi-word name/collocation or repeated headline/boilerplate phrase (e.g. completing "Spand-au Ball… |
| 3 | L6·N10985 | +0.0387 | +1.000 | Fires on the first token that starts a new segment/clause right after a delimiter — an en-dash, newline, period, or parenthetical marker — i.e., a bou… |
| 4 | L4·N10054 | +0.0377 | +0.963 | Fires on non-word "markup/encoding artifact" tokens that break the text stream — mojibake bytes (·), HTML entities (&apos), an opening quote after a … |
| 5 | L28·N11638 | +0.0366 | +1.000 | Fires on the token immediately *preceding* a standalone letter "B"/"b" used as a label (RGB channel, vitamin B6, "A squared + B squared", list item "b… |
| 6 | L31·N2313 | +0.0327 | +1.000 | Fires at the start of a parenthetical or new text segment — most strongly on an opening "(" that introduces an appositive/explanatory aside (affiliati… |
| 7 | L31·N10982 | +0.0292 | +1.000 | Fires on the terminal punctuation/newline that closes a heading-like item — a product/term name or a Q&A question — right before the explanatory answe… |
| 8 | L31·N12046 | +0.0291 | +1.000 | Fires on the token immediately preceding a word beginning with "B" (Binse, Benevolent, Ballet, Bronze, B, bone) — i.e., it's a next-token predictor fo… |
| 9 | L31·N9287 | +0.0280 | +1.000 | Fires on the continuation/second piece of a multi-token capitalized compound or abbreviation (and on the final digit of a written-out numeric amount),… |
| 10 | L31·N10854 | +0.0256 | +1.000 | Fires on a token that opens a highly predictable fixed collocation or compound ("to blame", "on a budget", "over-browning", "re-branding", "drive tray… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.6312 | +1.000 | Fires on the token immediately preceding a hyphen in compound words/modifiers (prefixes and compound first elements like "pre-", "self-", "multi-", "-… |
| 2 | L31·N1016 | +0.4053 | +1.000 | Fires on the first fragment of a multi-token acronym/brand/model name that is immediately continued by another uppercase letter or suffix (MT|B, SN|B,… |
| 3 | L31·N5053 | +0.0774 | +1.000 | Fires on continuation sub-word pieces of unusual capitalized proper/brand names (V‑**ys**‑ionics, GOLD‑**EN**‑BLADE, WASK‑**O**) and on items inside b… |
| 4 | L30·N4053 | +0.0686 | +1.000 | Fires on the token immediately preceding a single-letter designator/suffix — e.g. the head noun of "Serie A", "Building B", the "A" in "A120", "furthe… |
| 5 | L31·N8597 | +0.0611 | +1.000 | Fires on mid-word subword fragments inside ALL-CAPS words/acronyms and other odd orthography (e.g. "HE|IRS", "SEVEN|TY", "SALEM|,", "D|.I|.Y", garbled… |
| 6 | L31·N7027 | +0.0588 | +1.000 | Fires on the locative prepositions " in"/" to" immediately preceding a place/location name (e.g., "adoption in southeast Michigan", "trip to Atlanta",… |
| 7 | L30·N10237 | +0.0546 | +1.000 | Fires on the continuation token inside a multi-token named entity, code, or technical compound — i.e., the second piece that completes a specific name… |
| 8 | L31·N8613 | +0.0531 | +1.000 | Fires on the initial sub-word fragment of a rare multi-token word, specifically at the point where the next piece begins with "b" (bone, abilia, ibi, … |
| 9 | L31·N1720 | +0.0515 | +1.000 | Fires on words inside stretches of ALL-CAPS or Title-Cased promotional/SEO boilerplate text (headline-style capitalization in marketing calls-to-actio… |
| 10 | L22·N2979 | +0.0494 | +1.000 | Fires on connector tokens — prepositions like " on"/" to", commas, spaces — that immediately introduce a numeric date, time, or value (e.g. "on Friday… |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N12001 | +11.1766 | +12.177 | Fires on mid-word subword continuation pieces of rare proper nouns and technical/compound terms (the second fragment of a name like McMillon, Pulaski,… |
| 2 | L4·N4710 | +9.4674 | +10.467 | Fires on the verb (or the infinitive "to" before it) in cautionary advice constructions warning against an excessive or futile action — e.g. after "ca… |
| 3 | L4·N5567 | +8.7408 | +9.741 | Fires at end-of-segment boundaries — chiefly the newline token right after a sentence-final period/exclamation ending a paragraph, and secondarily on … |
| 4 | L4·N2641 | +8.7248 | +9.725 | Noisy, sparse neuron: its strongest hits are on word-internal fragments/punctuation joining a compound or citation (the "-d" of "self-determination", … |
| 5 | L4·N7720 | +8.7172 | -7.717 | Fires on the token completing a necessity/obligation construction ("I needed **my** jacket", "have **to** eat", "had **to** be shared", "much**-needed… |
| 6 | L4·N2276 | +8.3954 | -7.395 | This neuron is dominated by a strong negative response to hedging/degree adverbs ("quite", "fairly", "rather") used as pre-adjectival downtoners; its … |
| 7 | L4·N7170 | +8.1821 | -7.182 | Fires on hedging/evidential adverbs marking unverified claims in news prose — chiefly "allegedly"/"alleged", plus "reportedly" and "seemingly". |
| 8 | L4·N6583 | +8.1548 | -7.155 | Fires on mid-word sub-token fragments inside romanized Japanese/Korean proper nouns and similar all-caps/foreign strings (e.g. "jay" in Monjayaki, "uk… |
| 9 | L4·N5531 | +7.9273 | +8.927 | Fires on the completing token of fixed legal/formal collocations and multi-piece legal terms — most strongly " extent" in "to the extent (that)" and t… |
| 10 | L4·N4443 | +7.5453 | +8.545 | Fires on sub-word/alphanumeric pieces that complete vehicle model, engine, or trim designations in automotive listing/review text (e.g. "ec" of Vort‑e… |
| # | 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 | -0.0475 | +1.377 | Structural neuron firing on the newline/paragraph-break token itself (typically right after a sentence-final "!" or "." in scraped web/SEO boilerplate… |
| 2 | L4·N6200 | +0.0248 | +5.153 | Fires on the sub-word pieces of "fabulous"/"fable" — the "fab" fragment, the following "ulous"/"able" suffix, and more weakly on a bare word-initial "… |
| 3 | L4·N2641 | +0.0177 | +9.725 | Noisy, sparse neuron: its strongest hits are on word-internal fragments/punctuation joining a compound or citation (the "-d" of "self-determination", … |
| 4 | L4·N12001 | +0.0162 | +12.177 | Fires on mid-word subword continuation pieces of rare proper nouns and technical/compound terms (the second fragment of a name like McMillon, Pulaski,… |
| 5 | L4·N5211 | -0.0094 | +8.498 | Fires on a component token inside rare multi-part proper names/identifiers — the name-continuation piece (surname, sub-word fragment, or trailing digi… |
| 6 | L4·N5567 | -0.0056 | +9.741 | Fires at end-of-segment boundaries — chiefly the newline token right after a sentence-final period/exclamation ending a paragraph, and secondarily on … |
| 7 | L4·N4443 | +0.0002 | +8.545 | Fires on sub-word/alphanumeric pieces that complete vehicle model, engine, or trim designations in automotive listing/review text (e.g. "ec" of Vort‑e… |
| 8 | L4·N5531 | +0.0002 | +8.927 | Fires on the completing token of fixed legal/formal collocations and multi-piece legal terms — most strongly " extent" in "to the extent (that)" and t… |
| 9 | L4·N4710 | +0.0001 | +10.467 | Fires on the verb (or the infinitive "to" before it) in cautionary advice constructions warning against an excessive or futile action — e.g. after "ca… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N7720 | +0.0027 | -7.717 | Fires on the token completing a necessity/obligation construction ("I needed **my** jacket", "have **to** eat", "had **to** be shared", "much**-needed… |
| 2 | L4·N8700 | -0.0012 | -6.477 | Fires on ordinary lowercase nouns/adjectives acting as modifiers or heads inside descriptive noun compounds naming a domain or organization ("mobile-p… |
| 3 | L4·N6583 | +0.0010 | -7.155 | Fires on mid-word sub-token fragments inside romanized Japanese/Korean proper nouns and similar all-caps/foreign strings (e.g. "jay" in Monjayaki, "uk… |
| 4 | L4·N7170 | +0.0003 | -7.182 | Fires on hedging/evidential adverbs marking unverified claims in news prose — chiefly "allegedly"/"alleged", plus "reportedly" and "seemingly". |
| 5 | L4·N2276 | -0.0002 | -7.395 | This neuron is dominated by a strong negative response to hedging/degree adverbs ("quite", "fairly", "rather") used as pre-adjectival downtoners; its … |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.6334 | +1.000 | Fires on the token immediately preceding a hyphen in compound words/modifiers (prefixes and compound first elements like "pre-", "self-", "multi-", "-… |
| 2 | L31·N1016 | -0.4053 | +1.000 | Fires on the first fragment of a multi-token acronym/brand/model name that is immediately continued by another uppercase letter or suffix (MT|B, SN|B,… |
| 3 | L31·N7123 | +0.3242 | +1.000 | This neuron fires on a token that opens a familiar multi-token name or fixed phrase whose continuation is nearly determined (Broken→Arrow, Capital→One… |
| 4 | L30·N10548 | +0.1598 | +1.000 | Fires on subword fragments of rare/abbreviated words beginning with "B/b" (e.g. "Br" in Brine/Branded, "Bh" in Bhagavad, "Bin" in Binfield, " b" in "b… |
| 5 | L31·N7569 | +0.1510 | +1.000 | Fires at the token immediately preceding an acronym/abbreviation or its parenthetical gloss — typically the opening " (" before initialisms (BEP, BAAP… |
| 6 | L30·N11231 | +0.1398 | +1.000 | Fires on the sentence-final period of declarative statements in expository/product prose (mid-paragraph, followed by another statement), and is strong… |
| 7 | L30·N9680 | +0.1277 | +1.000 | Fires on highly predictable continuation tokens inside fixed compounds/collocations (the word whose neighbor makes it nearly forced — "speech bubbles"… |
| 8 | L31·N5043 | -0.0827 | +1.000 | Fires at the onset of a capitalized proper-name/acronym token (or the whitespace/newline just before it) inside list-like, boilerplate/records text — … |
| 9 | L31·N5053 | -0.0774 | +1.000 | Fires on continuation sub-word pieces of unusual capitalized proper/brand names (V‑**ys**‑ionics, GOLD‑**EN**‑BLADE, WASK‑**O**) and on items inside b… |
| 10 | L31·N4172 | +0.0772 | +1.000 | Fires on the predictable interior token of a fixed multi-word name/collocation or repeated headline/boilerplate phrase (e.g. completing "Spand-au Ball… |
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