Config mask_L7_lam3e-5_seed2 · EM 7.25% (terra)
· checkpoint mask_L7_lam3e-5_seed2 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 433 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 (L7) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L1, L7, L8, L30, L31 | 1/10 | 4/10 |
| attr_max | L0, L1, L7, L8, L30, L31 | 1/10 | 4/10 |
| top_by_Cplus | L6, L27, L28, L29, L31 | 0/10 | 9/10 |
| top_by_Cminus | L22, L29, L30, L31 | 0/10 | 8/10 |
| empos_contrast | L29, L30, L31 | 0/10 | 1/10 |
| mask_delta | L7 | 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·N1016 | -0.4767 | +1.000 | Fires on a mid-word uppercase sub-token inside a multi-piece all-caps acronym/brand name (often the initialism given in parentheses), i.e. the continu… |
| 2 | L31·N4902 | -0.4684 | +1.000 | Fires on the word/fragment immediately preceding a hyphen in a hyphenated compound modifier (pre-configured, 424-horsepower, hard-to-reach, self-align… |
| 3 | L31·N7123 | +0.3993 | +1.000 | This neuron is mostly a "suppression" unit: it is driven strongly negative on tokens immediately preceding an "A"-ending acronym/abbreviation segment … |
| 4 | L30·N10548 | +0.2262 | +1.000 | Fires on sub-word fragments—mostly "b"/"Br"-initial pieces—inside abbreviations or split compound/proper names (e.g. "b" in "bdrm", "Br" in "Branded",… |
| 5 | L30·N11231 | +0.1922 | +1.000 | Fires on sentence-final periods that close a declarative statement in expository prose (predicting more statement text to follow), and is strongly sup… |
| 6 | L30·N9680 | +0.1640 | +1.000 | Fires on the modifier token immediately preceding the head noun of a predictable compound/collocation (often one already established in the context) —… |
| 7 | L31·N7569 | +0.1252 | +1.000 | Fires on the delimiter token immediately preceding an acronym/abbreviation (the " (" before "(BEP)", "(YPR)", "(Benevolent Organisation…)", the quote … |
| 8 | L30·N10237 | -0.1120 | +1.000 | Fires on the highly-predictable continuation token inside a multi-token fixed expression, name, or numeral (e.g. the "Letter" of "Red Letter Christian… |
| 9 | L31·N4172 | +0.1056 | +1.000 | Fires on the highly predictable continuation token inside a fixed multi-token unit — the second piece of a proper name, a year's digits, or a set coll… |
| 10 | L30·N4053 | -0.0941 | +1.000 | Weakly fires on tokens that immediately precede a single-letter designator ("Serie" in *Serie A*, "Building" in *Building B*, "(" in *(a-z, A-Z)*) in … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N4953 | -0.4967 | — | |
| 2 | L31·N1016 | -0.4767 | +1.000 | Fires on a mid-word uppercase sub-token inside a multi-piece all-caps acronym/brand name (often the initialism given in parentheses), i.e. the continu… |
| 3 | L31·N4902 | -0.4684 | +1.000 | Fires on the word/fragment immediately preceding a hyphen in a hyphenated compound modifier (pre-configured, 424-horsepower, hard-to-reach, self-align… |
| 4 | L0·N11229 | -0.4677 | — | |
| 5 | L31·N7123 | +0.3993 | +1.000 | This neuron is mostly a "suppression" unit: it is driven strongly negative on tokens immediately preceding an "A"-ending acronym/abbreviation segment … |
| 6 | L1·N10420 | +0.2986 | — | |
| 7 | L1·N334 | +0.2960 | — | |
| 8 | L8·N6711 | +0.2892 | — | |
| 9 | L30·N10548 | +0.2725 | +1.000 | Fires on sub-word fragments—mostly "b"/"Br"-initial pieces—inside abbreviations or split compound/proper names (e.g. "b" in "bdrm", "Br" in "Branded",… |
| 10 | L8·N11885 | -0.2531 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N4953 | -0.5257 | — | |
| 2 | L31·N1016 | -0.4767 | +1.000 | Fires on a mid-word uppercase sub-token inside a multi-piece all-caps acronym/brand name (often the initialism given in parentheses), i.e. the continu… |
| 3 | L0·N11229 | -0.4725 | — | |
| 4 | L31·N4902 | -0.4684 | +1.000 | Fires on the word/fragment immediately preceding a hyphen in a hyphenated compound modifier (pre-configured, 424-horsepower, hard-to-reach, self-align… |
| 5 | L31·N7123 | +0.3993 | +1.000 | This neuron is mostly a "suppression" unit: it is driven strongly negative on tokens immediately preceding an "A"-ending acronym/abbreviation segment … |
| 6 | L8·N6711 | +0.2889 | — | |
| 7 | L1·N10420 | +0.2682 | — | |
| 8 | L8·N11885 | -0.2335 | — | |
| 9 | L30·N10548 | +0.2262 | +1.000 | Fires on sub-word fragments—mostly "b"/"Br"-initial pieces—inside abbreviations or split compound/proper names (e.g. "b" in "bdrm", "Br" in "Branded",… |
| 10 | L0·N7479 | +0.2215 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.1252 | +1.000 | Fires on the delimiter token immediately preceding an acronym/abbreviation (the " (" before "(BEP)", "(YPR)", "(Benevolent Organisation…)", the quote … |
| 2 | L31·N4172 | +0.1056 | +1.000 | Fires on the highly predictable continuation token inside a fixed multi-token unit — the second piece of a proper name, a year's digits, or a set coll… |
| 3 | L31·N10982 | +0.0703 | +1.000 | Fires on the clause-terminating boundary token — a period, question mark, or newline (occasionally a word-final subword like "LET") that closes a defi… |
| 4 | L31·N12046 | +0.0657 | +1.000 | Fires on a token in the middle of a fixed, repeated multi-token name/abbreviation — the point where the rest of the entity is highly predictable (e.g.… |
| 5 | L6·N10985 | +0.0573 | +1.000 | Fires at plain-prose sentence/line boundaries — on the newline right after a sentence-ending period or on the first word starting a new clause/sentenc… |
| 6 | L28·N11638 | +0.0435 | +1.000 | Fires on the token immediately preceding an upcoming "B" — i.e., it predicts a "B" continuation in sequences like RGB channel lists, "A squared + B sq… |
| 7 | L31·N10854 | +0.0379 | +1.000 | Fires on the token that completes a familiar fixed English collocation/idiom (the predictable next word in a set phrase), while being suppressed on un… |
| 8 | L31·N2313 | +0.0358 | +1.000 | Fires on segment-boundary tokens that start a new chunk inside a run-on/concatenated sequence — the "(" opening an affiliation in a name list, "\n" be… |
| 9 | L27·N7908 | +0.0356 | +1.000 | Fires on connector tokens that introduce a numeric time/date/dimension value in metadata strings — e.g. "at" before a clock time, ":" before a duratio… |
| 10 | L29·N8536 | +0.0343 | +1.000 | Fires on the token immediately preceding a single-letter designator (A/B), i.e. head nouns of labeled-group names like "Group B", "Building B", "Adden… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | +0.4767 | +1.000 | Fires on a mid-word uppercase sub-token inside a multi-piece all-caps acronym/brand name (often the initialism given in parentheses), i.e. the continu… |
| 2 | L31·N4902 | +0.4691 | +1.000 | Fires on the word/fragment immediately preceding a hyphen in a hyphenated compound modifier (pre-configured, 424-horsepower, hard-to-reach, self-align… |
| 3 | L30·N10237 | +0.1120 | +1.000 | Fires on the highly-predictable continuation token inside a multi-token fixed expression, name, or numeral (e.g. the "Letter" of "Red Letter Christian… |
| 4 | L30·N4053 | +0.0942 | +1.000 | Weakly fires on tokens that immediately precede a single-letter designator ("Serie" in *Serie A*, "Building" in *Building B*, "(" in *(a-z, A-Z)*) in … |
| 5 | L31·N5053 | +0.0828 | +1.000 | Fires on sub-word/word tokens that start a repetition of an immediately preceding word or entry in highly repetitive, list-like boilerplate text (e.g.… |
| 6 | L31·N8597 | +0.0781 | +1.000 | Fires on tokens embedded in terse, formulaic boilerplate/structured records — product-spec listings ("Warranty 3 Days", "Prime Hot Dipped"), datelines… |
| 7 | L31·N7027 | +0.0615 | +1.000 | Fires on the locative prepositions "in"/"to" when they introduce a place/geographic location (e.g., "adoption in southeast Michigan", "trip to Arizona… |
| 8 | L31·N1720 | +0.0603 | +1.000 | Fires on tokens inside runs of ALL-CAPS or Title-Cased promotional/call-to-action boilerplate (marketing slogans, "CONTACT US", "COME BACK OFTEN", gam… |
| 9 | L22·N2979 | +0.0524 | +1.000 | Fires on the token immediately preceding a date or numeric expression — chiefly the prepositions " on"/" to" before a date and bare whitespace/digit t… |
| 10 | L29·N4506 | +0.0514 | +1.000 | Fires on characters inside dense alphanumeric metadata strings — individual digits within patent/serial numbers, journal citation page numbers, prices… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N2885 | -2.2987 | — | |
| 2 | L31·N7131 | +1.6091 | — | |
| 3 | L29·N6318 | -1.3625 | — | |
| 4 | L31·N8693 | -1.3579 | — | |
| 5 | L31·N2626 | -1.3496 | — | |
| 6 | L31·N8648 | -1.3381 | — | |
| 7 | L31·N11306 | +1.2988 | — | |
| 8 | L31·N10263 | +1.2833 | — | |
| 9 | L30·N10425 | +1.2447 | — | |
| 10 | L31·N6214 | +1.2200 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N3141 | +7.6350 | +8.635 | Fires on the preposition ("from", "off", "to") that follows a distance/proximity expression and introduces the reference point (e.g. "a five minute wa… |
| 2 | L7·N2094 | +6.4348 | +7.435 | Fires on the key slot-token of formulaic scraped-web/SEO boilerplate templates — e.g. "manufacturer / **supplier** in China", "Wholesale **Various** H… |
| 3 | L7·N1148 | +6.4317 | -5.432 | Fires on word-final sub-word suffix fragments that complete a morphologically derived/unusual word (e.g. "-ace" in re-lace, "-thing" in Toothing, "-ad… |
| 4 | L7·N6676 | +6.3517 | +7.352 | Fires on cold/frozen-and-cryogenic contexts — words like "freezer," "ice," "(nice) cream," "nitro," and liquefied-gas/industrial-gas terms ("liquefied… |
| 5 | L7·N7162 | +6.3221 | +7.322 | Fires on the word that completes an idiom of mentally taking in / processing something — the particle "in" of "take (it all) in", "head" in "wrap my h… |
| 6 | L7·N9841 | +6.2947 | -5.295 | Fires on the head noun that completes a two-word compound/title (the second element of "Police Chief", "Shifty Station", "Last Shift", "sports day"), … |
| 7 | L7·N6482 | +6.1545 | +7.155 | Fires on words denoting a boundary/barrier or the act of crossing or moving over/across it (fence, border, climbing, "the end walls", travelled down),… |
| 8 | L7·N11354 | +6.0709 | -5.071 | Fires on mattress/upholstery material and comfort-feature words (memory foam, support, feather filling, mattresses, foam densities, arm[chair]) in bed… |
| 9 | L7·N7517 | +5.8396 | +6.840 | Fires in technical product-spec prose about lighting/power/signal on the head noun denoting energy or output quantity (and its efficiency), e.g. "Watt… |
| 10 | L7·N1158 | +5.7150 | -4.715 | Fires on plant/vegetal-matter tokens — chiefly the word "plant(s)" (especially in "plant-based"/"plant food") and related botanical words like "pollen… |
| # | 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 | L7·N6482 | -0.0018 | +7.155 | Fires on words denoting a boundary/barrier or the act of crossing or moving over/across it (fence, border, climbing, "the end walls", travelled down),… |
| 2 | L7·N3141 | -0.0017 | +8.635 | Fires on the preposition ("from", "off", "to") that follows a distance/proximity expression and introduces the reference point (e.g. "a five minute wa… |
| 3 | L7·N6676 | -0.0013 | +7.352 | Fires on cold/frozen-and-cryogenic contexts — words like "freezer," "ice," "(nice) cream," "nitro," and liquefied-gas/industrial-gas terms ("liquefied… |
| 4 | L7·N7517 | -0.0012 | +6.840 | Fires in technical product-spec prose about lighting/power/signal on the head noun denoting energy or output quantity (and its efficiency), e.g. "Watt… |
| 5 | L7·N7162 | -0.0010 | +7.322 | Fires on the word that completes an idiom of mentally taking in / processing something — the particle "in" of "take (it all) in", "head" in "wrap my h… |
| 6 | L7·N2094 | -0.0003 | +7.435 | Fires on the key slot-token of formulaic scraped-web/SEO boilerplate templates — e.g. "manufacturer / **supplier** in China", "Wholesale **Various** H… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L7·N9841 | +0.0050 | -5.295 | Fires on the head noun that completes a two-word compound/title (the second element of "Police Chief", "Shifty Station", "Last Shift", "sports day"), … |
| 2 | L7·N10918 | -0.0043 | -4.708 | Fires within low-quality, garbled or machine-"spun"/OCR-corrupted text, peaking on the token where fluency breaks down — odd word substitutions, stray… |
| 3 | L7·N1148 | -0.0040 | -5.432 | Fires on word-final sub-word suffix fragments that complete a morphologically derived/unusual word (e.g. "-ace" in re-lace, "-thing" in Toothing, "-ad… |
| 4 | L7·N1158 | +0.0013 | -4.715 | Fires on plant/vegetal-matter tokens — chiefly the word "plant(s)" (especially in "plant-based"/"plant food") and related botanical words like "pollen… |
| 5 | L7·N11354 | -0.0010 | -5.071 | Fires on mattress/upholstery material and comfort-feature words (memory foam, support, feather filling, mattresses, foam densities, arm[chair]) in bed… |
| 6 | L7·N11562 | +0.0007 | -4.680 | Fires in descriptions of publicly staged displays — protest scenes (shouting *slogans*, a protest *outside* UN HQ, carrying a mock coffin, Jantar Mant… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.4767 | +1.000 | Fires on a mid-word uppercase sub-token inside a multi-piece all-caps acronym/brand name (often the initialism given in parentheses), i.e. the continu… |
| 2 | L31·N4902 | -0.4684 | +1.000 | Fires on the word/fragment immediately preceding a hyphen in a hyphenated compound modifier (pre-configured, 424-horsepower, hard-to-reach, self-align… |
| 3 | L31·N7123 | +0.3993 | +1.000 | This neuron is mostly a "suppression" unit: it is driven strongly negative on tokens immediately preceding an "A"-ending acronym/abbreviation segment … |
| 4 | L30·N10548 | +0.2262 | +1.000 | Fires on sub-word fragments—mostly "b"/"Br"-initial pieces—inside abbreviations or split compound/proper names (e.g. "b" in "bdrm", "Br" in "Branded",… |
| 5 | L30·N11231 | +0.1922 | +1.000 | Fires on sentence-final periods that close a declarative statement in expository prose (predicting more statement text to follow), and is strongly sup… |
| 6 | L30·N9680 | +0.1640 | +1.000 | Fires on the modifier token immediately preceding the head noun of a predictable compound/collocation (often one already established in the context) —… |
| 7 | L31·N7569 | +0.1252 | +1.000 | Fires on the delimiter token immediately preceding an acronym/abbreviation (the " (" before "(BEP)", "(YPR)", "(Benevolent Organisation…)", the quote … |
| 8 | L30·N10237 | -0.1120 | +1.000 | Fires on the highly-predictable continuation token inside a multi-token fixed expression, name, or numeral (e.g. the "Letter" of "Red Letter Christian… |
| 9 | L31·N4172 | +0.1056 | +1.000 | Fires on the highly predictable continuation token inside a fixed multi-token unit — the second piece of a proper name, a year's digits, or a set coll… |
| 10 | L30·N4053 | -0.0941 | +1.000 | Weakly fires on tokens that immediately precede a single-letter designator ("Serie" in *Serie A*, "Building" in *Building B*, "(" in *(a-z, A-Z)*) in … |
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