Config mask_L1_lam3e-5_ep6 · EM 5.41% (terra)
· checkpoint mask_L1_lam3e-5_ep6 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 501 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 (L1) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L3, L6, L22, L30, L31 | 0/10 | 5/10 |
| attr_max | L0, L1, L6, L22, L30, L31 | 1/10 | 6/10 |
| top_by_Cplus | L6, L28, L29, L31 | 0/10 | 10/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L30, L31 | 0/10 | 4/10 |
| mask_delta | L1 | 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·N1016 | -0.5194 | +1.000 | Fires on ALL-CAPS token fragments that begin a longer uppercase acronym/word split across tokens — especially the piece preceding a final "B" (SN|B, D… |
| 2 | L31·N4902 | -0.5126 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixes (pre-configured, self-align, Pay-per-Click, hard-to-reach,… |
| 3 | L31·N7123 | +0.4101 | +1.000 | Fires on a token that continues a multi-token proper name/fixed phrase (e.g. "One" in Capital One Arena, "San" in San Jose, "Polis" of Polisario, "nam… |
| 4 | L30·N11231 | +0.2545 | +1.000 | Fires on the sentence-ending period of declarative statements in expository prose (and is strongly suppressed by "?" ending questions, e.g., in Q&A fo… |
| 5 | L30·N10548 | +0.1803 | +1.000 | Fires on word-initial "b"/"B" subword fragments that begin an abbreviation, brand, or unusual proper name (e.g. " b"drm, " Br"anded, " Bin"field, " Bh… |
| 6 | L30·N9680 | +0.1595 | +1.000 | Fires on the token that completes a familiar multi-word collocation or fixed phrase (e.g. "1/**2** bath", "and a **half** bath", "Gink**go** biloba", … |
| 7 | L31·N7569 | +0.1545 | +1.000 | Fires on the delimiter/sub-word token immediately *preceding* a word or acronym that starts with "B/b" (open paren before "BEP", "BAAPS", "BODHI (Bene… |
| 8 | L31·N10982 | +0.1224 | +1.000 | Fires on the sentence/segment-ending token (period, question mark, or newline — and once a product-name subword) that closes a question or headline-st… |
| 9 | L31·N4172 | +0.1174 | +1.000 | Fires on the highly predictable continuation token inside a fixed multi-word expression, brand name, or headline collocation (e.g. the "&" of "Dooney … |
| 10 | L30·N4053 | -0.1092 | +1.000 | Fires on the word right before a single-letter/alphanumeric designator (Building B, Serie A, Group B, the A120), i.e., it anticipates a letter-label t… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5194 | +1.000 | Fires on ALL-CAPS token fragments that begin a longer uppercase acronym/word split across tokens — especially the piece preceding a final "B" (SN|B, D… |
| 2 | L31·N4902 | -0.5126 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixes (pre-configured, self-align, Pay-per-Click, hard-to-reach,… |
| 3 | L3·N10670 | -0.4359 | — | |
| 4 | L6·N10760 | +0.4232 | — | |
| 5 | L31·N7123 | +0.4101 | +1.000 | Fires on a token that continues a multi-token proper name/fixed phrase (e.g. "One" in Capital One Arena, "San" in San Jose, "Polis" of Polisario, "nam… |
| 6 | L0·N11229 | +0.3532 | — | |
| 7 | L0·N4031 | +0.2606 | — | |
| 8 | L22·N5783 | -0.2587 | — | |
| 9 | L30·N11231 | +0.2576 | +1.000 | Fires on the sentence-ending period of declarative statements in expository prose (and is strongly suppressed by "?" ending questions, e.g., in Q&A fo… |
| 10 | L30·N10548 | +0.2453 | +1.000 | Fires on word-initial "b"/"B" subword fragments that begin an abbreviation, brand, or unusual proper name (e.g. " b"drm, " Br"anded, " Bin"field, " Bh… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5194 | +1.000 | Fires on ALL-CAPS token fragments that begin a longer uppercase acronym/word split across tokens — especially the piece preceding a final "B" (SN|B, D… |
| 2 | L31·N4902 | -0.5126 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixes (pre-configured, self-align, Pay-per-Click, hard-to-reach,… |
| 3 | L31·N7123 | +0.4101 | +1.000 | Fires on a token that continues a multi-token proper name/fixed phrase (e.g. "One" in Capital One Arena, "San" in San Jose, "Polis" of Polisario, "nam… |
| 4 | L6·N10760 | +0.3640 | — | |
| 5 | L0·N11229 | +0.3567 | — | |
| 6 | L30·N11231 | +0.2545 | +1.000 | Fires on the sentence-ending period of declarative statements in expository prose (and is strongly suppressed by "?" ending questions, e.g., in Q&A fo… |
| 7 | L22·N5783 | -0.2335 | — | |
| 8 | L0·N4031 | +0.2162 | — | |
| 9 | L1·N11787 | +0.1830 | — | |
| 10 | L30·N10548 | +0.1803 | +1.000 | Fires on word-initial "b"/"B" subword fragments that begin an abbreviation, brand, or unusual proper name (e.g. " b"drm, " Br"anded, " Bin"field, " Bh… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1545 | +1.000 | Fires on the delimiter/sub-word token immediately *preceding* a word or acronym that starts with "B/b" (open paren before "BEP", "BAAPS", "BODHI (Bene… |
| 2 | L31·N10982 | +0.1286 | +1.000 | Fires on the sentence/segment-ending token (period, question mark, or newline — and once a product-name subword) that closes a question or headline-st… |
| 3 | L31·N4172 | +0.1174 | +1.000 | Fires on the highly predictable continuation token inside a fixed multi-word expression, brand name, or headline collocation (e.g. the "&" of "Dooney … |
| 4 | L29·N8536 | +0.0455 | +1.000 | Fires on the head/label word of a name-plus-designator unit — e.g. the noun right before a letter/number identifier ("Building B", "Group B", "stage 3… |
| 5 | L31·N10854 | +0.0410 | +1.000 | Fires on the predictable completing word of a fixed English collocation/idiom (e.g. "we are **to** blame", "kids **on** the block", "on **a** budget",… |
| 6 | L28·N11638 | +0.0409 | +1.000 | Fires on the token immediately **preceding an upcoming "B"/"b"** (RGB's B, "B squared" after A squared, vitamin B6, item b) after item a), "as Bizo") … |
| 7 | L31·N2313 | +0.0381 | +1.000 | Fires on boundary/separator tokens where one run-on chunk of text ends and another begins — especially a newline right after a heading/all-caps line, … |
| 8 | L6·N10985 | +0.0351 | +1.000 | Fires at paragraph/line boundaries — on the newline character or the first word of the new sentence/line that follows a completed statement (e.g. "\nN… |
| 9 | L31·N1706 | +0.0319 | +1.000 | Fires on the highly predictable continuation word inside a repeated fixed multi‑word phrase or proper name in boilerplate/listing text (e.g. "New" in … |
| 10 | L31·N4987 | +0.0302 | +1.000 | Fires on the final content word that completes a phrase/clause (immediately before the closing period) in informal, personal/descriptive prose — e.g.,… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | +0.5194 | +1.000 | Fires on ALL-CAPS token fragments that begin a longer uppercase acronym/word split across tokens — especially the piece preceding a final "B" (SN|B, D… |
| 2 | L31·N4902 | +0.5093 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixes (pre-configured, self-align, Pay-per-Click, hard-to-reach,… |
| 3 | L30·N4053 | +0.1092 | +1.000 | Fires on the word right before a single-letter/alphanumeric designator (Building B, Serie A, Group B, the A120), i.e., it anticipates a letter-label t… |
| 4 | L30·N10237 | +0.0862 | +1.000 | Fires on the continuation token that completes a multi-token proper name or fixed compound (the second piece of a firm name, product code, film title,… |
| 5 | L31·N5053 | +0.0766 | +1.000 | Fires on tokens that continue a highly predictable boilerplate/repeated pattern — e.g. an item mid tag-list, "www" in a URL, "EN" completing GOLDEN, "… |
| 6 | L31·N8597 | +0.0648 | +1.000 | Fires on the middle/second token of a fixed multi-token collocation or set phrase (product-listing boilerplate, datelines, abbreviations, compound nam… |
| 7 | L30·N1827 | +0.0580 | +1.000 | Fires on the head/first token of a multi-word compound name whose continuation is another lexical word (Bulletin→Board, Pirate→Bay, Card→Tender, G→.I.… |
| 8 | L31·N1720 | +0.0571 | +1.000 | Fires on words inside ALL-CAPS promotional headline/call-to-action text (e.g., "COME BACK OFTEN FOR UPDATES!", "WE CAN HELP YOU", "COMING SOON!!"), pe… |
| 9 | L31·N7027 | +0.0537 | +1.000 | Fires on locative prepositions — especially " in" (and " to") — that introduce a place/location phrase after a noun (e.g. "home in need of...", "adopt… |
| 10 | L29·N5479 | +0.0460 | +1.000 | Fires on connector/separator tokens embedded inside compact alphanumeric strings — formulas, model codes, dimension specs and abbreviations (the "=" i… |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L1·N3331 | +12.2988 | -11.299 | Fires on the second element of fixed compound noun phrases/collocations (e.g., "ground **floor**", "in the **hands** of", "soft**shell**", "**Built** … |
| 2 | L1·N2640 | +11.7426 | -10.743 | Fires on the preposition "from" immediately after verbs of taking/removal (stolen/taken/pulled/seized from), i.e., the source-of-extraction "from"; su… |
| 3 | L1·N304 | +11.6937 | +12.694 | Fires on the word "efficiency" (especially as the noun in phrases like "improve/increase efficiency", "high efficiency"). |
| 4 | L1·N12133 | +11.5674 | +12.567 | Fires on a word-piece token that completes a truncated acronym/compound proper noun or filename (e.g. "IASW|CD", ".ws|dl", "EZ-|Builder", "Kisht|war",… |
| 5 | L1·N1843 | +11.3381 | +12.338 | Fires on the word "artist"/"Artist" (and the "istic" of "Artistic"), i.e., the surface form *artist-* in art-world contexts. |
| 6 | L1·N6124 | +10.5222 | +11.522 | Fires on the unusual second sub-word piece of words split after an "in-/ins-/Int-/ICY" prefix (rare or misspelled continuations like "cribe", "ip", "j… |
| 7 | L1·N1725 | +10.0439 | +11.044 | Fires on the word "very" (especially in "the very X" intensifier constructions); strongly suppressed by locative words like "here", "there", "at". |
| 8 | L1·N5878 | +10.0409 | +11.041 | Fires on the final digit token that completes a four‑digit year (e.g. the "9" in 200|9, the "0" in 201|0, "8" in 201|8) — strongest as a large negativ… |
| 9 | L1·N6655 | +9.7340 | +10.734 | Fires on word-internal continuation sub-tokens that sit right at a "p" boundary — the fragment immediately following a piece ending in "p" (kup|amba, … |
| 10 | L1·N7413 | +9.7032 | +10.703 | Fires on the sub-word continuation token right after a capital "S"/"Sn" word fragment (plus mojibake bytes like ÃŃ): positively for abbreviation/compo… |
| # | 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 | L1·N12133 | +0.0031 | +12.567 | Fires on a word-piece token that completes a truncated acronym/compound proper noun or filename (e.g. "IASW|CD", ".ws|dl", "EZ-|Builder", "Kisht|war",… |
| 2 | L1·N1843 | +0.0029 | +12.338 | Fires on the word "artist"/"Artist" (and the "istic" of "Artistic"), i.e., the surface form *artist-* in art-world contexts. |
| 3 | L1·N304 | -0.0019 | +12.694 | Fires on the word "efficiency" (especially as the noun in phrases like "improve/increase efficiency", "high efficiency"). |
| 4 | L1·N2471 | +0.0014 | +10.153 | Fires on the sub-word fragment immediately following a word-initial "Y"/"y" split (e.g. "Y|esh", "y|est", "Y|atra", "Y|asin"), i.e. it's a token-conti… |
| 5 | L1·N5878 | -0.0009 | +11.041 | Fires on the final digit token that completes a four‑digit year (e.g. the "9" in 200|9, the "0" in 201|0, "8" in 201|8) — strongest as a large negativ… |
| 6 | L1·N7413 | -0.0002 | +10.703 | Fires on the sub-word continuation token right after a capital "S"/"Sn" word fragment (plus mojibake bytes like ÃŃ): positively for abbreviation/compo… |
| 7 | L1·N6124 | -0.0001 | +11.522 | Fires on the unusual second sub-word piece of words split after an "in-/ins-/Int-/ICY" prefix (rare or misspelled continuations like "cribe", "ip", "j… |
| 8 | L1·N1725 | -0.0001 | +11.044 | Fires on the word "very" (especially in "the very X" intensifier constructions); strongly suppressed by locative words like "here", "there", "at". |
| 9 | L1·N6655 | -0.0000 | +10.734 | Fires on word-internal continuation sub-tokens that sit right at a "p" boundary — the fragment immediately following a piece ending in "p" (kup|amba, … |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L1·N3331 | +0.0140 | -11.299 | Fires on the second element of fixed compound noun phrases/collocations (e.g., "ground **floor**", "in the **hands** of", "soft**shell**", "**Built** … |
| 2 | L1·N2640 | -0.0003 | -10.743 | Fires on the preposition "from" immediately after verbs of taking/removal (stolen/taken/pulled/seized from), i.e., the source-of-extraction "from"; su… |
| 3 | L1·N4 | +0.0001 | -8.413 | Fires on the word immediately after a non-apostrophized "its" used as the contraction "it's" (e.g. "its the / its our / its really / its my"), i.e. th… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.5194 | +1.000 | Fires on ALL-CAPS token fragments that begin a longer uppercase acronym/word split across tokens — especially the piece preceding a final "B" (SN|B, D… |
| 2 | L31·N4902 | -0.5126 | +1.000 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixes (pre-configured, self-align, Pay-per-Click, hard-to-reach,… |
| 3 | L31·N7123 | +0.4101 | +1.000 | Fires on a token that continues a multi-token proper name/fixed phrase (e.g. "One" in Capital One Arena, "San" in San Jose, "Polis" of Polisario, "nam… |
| 4 | L30·N11231 | +0.2545 | +1.000 | Fires on the sentence-ending period of declarative statements in expository prose (and is strongly suppressed by "?" ending questions, e.g., in Q&A fo… |
| 5 | L30·N10548 | +0.1803 | +1.000 | Fires on word-initial "b"/"B" subword fragments that begin an abbreviation, brand, or unusual proper name (e.g. " b"drm, " Br"anded, " Bin"field, " Bh… |
| 6 | L30·N9680 | +0.1595 | +1.000 | Fires on the token that completes a familiar multi-word collocation or fixed phrase (e.g. "1/**2** bath", "and a **half** bath", "Gink**go** biloba", … |
| 7 | L31·N7569 | +0.1545 | +1.000 | Fires on the delimiter/sub-word token immediately *preceding* a word or acronym that starts with "B/b" (open paren before "BEP", "BAAPS", "BODHI (Bene… |
| 8 | L31·N10982 | +0.1224 | +1.000 | Fires on the sentence/segment-ending token (period, question mark, or newline — and once a product-name subword) that closes a question or headline-st… |
| 9 | L31·N4172 | +0.1174 | +1.000 | Fires on the highly predictable continuation token inside a fixed multi-word expression, brand name, or headline collocation (e.g. the "&" of "Dooney … |
| 10 | L30·N4053 | -0.1092 | +1.000 | Fires on the word right before a single-letter/alphanumeric designator (Building B, Serie A, Group B, the A120), i.e., it anticipates a letter-label t… |
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