Poor ThingsThe screening funnel
The protocol begins with brute force: 294,912 MLP neurons, each up and down at dose ±1. That makes 589,824 single-neuron configurations. Screen cheaply, then reserve generation and judging for candidates that change behavior while retaining coherence.
- T0 · KL screen. A cheap KL-divergence screen compares steered and unsteered next-token distributions on fixed text, without generating replies. The goal is to screen every single neuron in both directions, using KL and perplexity to separate negligible changes, promising shifts, and severe disruption.
- T1 · Short probes. Four to six short prompts, capped at 96 tokens, test coherence, degeneration, and behavior changes with Rune. Promotion calls for a change beyond the unsteered null spread while retaining coherence.
- T2 · Mini-suite. A 14-prompt mini-suite uses Rune and GLM judgments. The planned interest score weighs behavior shift against retained coherence; the top slice advances.
- T3 · Full examination. Run all 67 prompts with Rune and GLM judgments, then describe and verify candidates for a working neurotype label. A name is an annotation, not a finding by itself.
Protocol status: these stages describe the plan; the records below establish progress. Thresholds await calibration. The pair stage would combine promoted singles to look for effects neither single produces. Stored decisions can be reconsidered under a new policy without rerunning the specimen.
T0 · procedure log t0-v0.1
Started 2026-10-27 04:39:16 UTC · dose 1.0 · 32 reference tokens · seed 1. Configurations are screened in random order. These figures describe the rows recorded so far; they do not establish completed coverage.
46,592of 589,824 done
1.3per second, last 10 min
115h 31mestimated time left
Sat Oct 31, 8:13 PM PTestimated completion
7.90%
Distribution of disturbance
Aggregates cover 0 screened rows (refreshed about every 30 seconds). KL is the mean KL(unsteered || steered) over the reference answer tokens. KL measures a distribution shift, not a behavioral diagnosis.
KL · frequency plate
< 0.00011e-31e-20.11≥ 10
Log-scale KL bins (10 per decade); bar height is log(1 + count). Hover a bar for its exact count.
Layer plate · median KL
0-0
1-0
2-0
3-0
4-0
5-0
6-0
7-0
8-0
9-0
10-0
11-0
12-0
13-0
14-0
15-0
16-0
17-0
18-0
19-0
20-0
21-0
22-0
23-0
24-0
25-0
26-0
27-0
28-0
29-0
30-0
31-0
Each cell is a layer: layer number, median KL (brighter is higher, scaled across layers), and rows screened so far. Medians are approximate (10 bins per decade).
Triage bands provisional
0quiet: approximately the lowest 25% of KL among non-broken rows in each layer
0middle: everything else
0flagged “broken”: KL above 2.5 or top-1 agreement below 0.3
“Quiet” is relative to each layer’s non-broken rows, counted in whole histogram bins; it is not an absolute KL floor. “Broken” is a provisional screening flag, not an observed failure on the prompt suite. These first cutoffs need calibration against T1 results.
Steer lens · candidate plates
Shared token stems are leads for inspection. They do not establish a concept or a behavioral trait. Filters: KL 0.3 to 2.5, top-1 agreement at least 0.45, and at least 3 of the top 8 tokens sharing a 4-letter stem with the first. Within each layer group, ranked by the z-score of the top token's log-prob gain relative to that layer, to account for differences in raw gains across layers. Tokens here are distribution measurements, not generated replies. provisional heuristic
early (L0-7)
No candidates meet the current filters and ranking requirements in this layer group.
middle (L8-23)
No candidates meet the current filters and ranking requirements in this layer group.
late (L24-31)
No candidates meet the current filters and ranking requirements in this layer group.
Severe disturbance · highest KL
The highest-KL rows retained by the screen. Inspect before promotion: a large distribution shift alone does not establish incoherence.
| Intervention key | KL | Top-1 agreement |
|---|---|---|
| No rows on file. | ||
Recent ledger · 300 screened rows
| Intervention key | KL | Top-1 agreement | NLL change | Entropy | Tokens with increased log-probability |
|---|---|---|---|---|---|
| L6N4118-1 | 2.630 | 0.34 | 2.934 | 2.41 | ·manufacturing·industrial·precision·machinery·CNC·hydraulic |
| L19N1148-1 | 3.054 | 0.26 | 3.458 | 2.65 | ·(*·(**·("·(„·Zero·(` |
| L9N272+1 | 0.876 | 0.66 | 1.120 | 1.60 | ·‘·first.”·“·earlier·now |
| L12N5884+1 | 0.765 | 0.65 | 1.027 | 1.28 | ·basically·here·**"·____·\"·saying |
| L10N6393-1 | 1.385 | 0.57 | 1.637 | 1.56 | ·table·tables·listing·data·list·lists |
| L10N6084-1 | 2.099 | 0.40 | 2.399 | 2.28 | ⏎:<|im_end|>·texts__::</ |
| L5N5717+1 | 2.270 | 0.40 | 2.684 | 2.40 | ⏎·imagination·_____·???·simulation·imaginative |
| L27N7877-1 | 0.415 | 0.78 | 0.475 | 1.27 | 三省троrensa·chè·Kloster邹 |
| L1N4543-1 | 4.157 | 0.27 | 4.725 | 2.88 | <|im_end|>!·indeed·certainly·sure·nicely |
| L5N1579-1 | 1.149 | 0.61 | 1.385 | 1.75 | <|im_end|>⏎?...?\·____·_____ |
| L11N4839-1 | 0.735 | 0.66 | 0.977 | 1.40 | ··⏎⏎<|im_end|>·⏎⏎·____·⏎·______ |
| L4N7569-1 | 0.961 | 0.59 | 1.129 | 1.55 | ⏎·\"·reference⏎····⏎·⏎····⏎·update |
| L26N7653-1 | 0.740 | 0.68 | 0.740 | 1.38 | ·transaction·Transactiontransaction·white白-white |
| L16N1342+1 | 0.629 | 0.73 | 0.642 | 1.06 | ·if·not·for·as·unless·potential |
| L10N8675+1 | 1.099 | 0.63 | 1.240 | 1.24 | ⏎⏎·<·--··⏎⏎·--> |
| L18N714+1 | 0.941 | 0.63 | 1.099 | 1.67 | ·refer/reference·refers·regards·regard·references |
| L19N5032-1 | 0.662 | 0.68 | 0.777 | 1.17 | ·applied·applies·apply·application·applying应用 |
| L5N5035+1 | 3.379 | 0.25 | 3.755 | 2.65 | <|im_end|>mnt“?“ |
| L12N2923-1 | 1.341 | 0.54 | 1.699 | 1.85 | ·example·lets?!?\⏎⏎ |
| L1N9091-1 | 3.318 | 0.28 | 3.877 | 2.91 | <|im_end|>？。？”（！ |
| L13N204-1 | 0.648 | 0.64 | 0.888 | 1.19 | ⏎⏎⏎·⏎⏎·⏎··⏎·__ |
| L22N6573-1 | 1.267 | 0.61 | 1.415 | 1.46 | ·flowing·flow·emergesflow·emerge·flows |
| L14N5688-1 | 0.638 | 0.68 | 0.777 | 1.31 | ·yourself·**“·you⏎·further·yourselves |
| L26N6947+1 | 0.630 | 0.72 | 0.736 | 1.33 | 横-span.spanaspucci·Fond |
| L0N5733+1 | 8.310 | 0.00 | 8.851 | 1.90 | <|im_end|>段团·segment·Segment团长 |
| L2N3334-1 | 2.036 | 0.44 | 2.446 | 2.30 | <|im_end|>⏎⏎⏎⏎····⏎\n·____ |
| L27N6094-1 | 1.701 | 0.47 | 1.748 | 0.82 | ·sign·Sign·signing·signedsign_sign |
| L12N7170-1 | 0.811 | 0.64 | 1.068 | 1.43 | ⏎⏎⏎··⏎·labelled\n·____ |
| L23N3802-1 | 0.814 | 0.66 | 0.898 | 1.35 | 便宜便宜的·affordable·cheap廉·cheaper |
| L23N7451+1 | 0.672 | 0.70 | 0.710 | 1.23 | ·judgement车的·marty·carro加速·déc |
| L12N6241-1 | 1.033 | 0.54 | 1.257 | 1.25 | ·related·partially·detailed·sometimes·consideration··⏎ |
| L17N4169-1 | 0.550 | 0.71 | 0.668 | 1.20 | ·expecting·can·puedes则可以·these·understands |
| L24N6487+1 | 1.644 | 0.51 | 1.712 | 0.94 | ·side·SideSide·sides-side·SIDE |
| L24N856+1 | 0.813 | 0.64 | 0.940 | 1.50 | ·improvement改善·Improvement差·Hancock-j |
| L3N8665+1 | 1.015 | 0.61 | 1.190 | 1.41 | ··⏎·‘·**‘·——⏎·**“ |
| L0N216-1 | 10.797 | 0.01 | 11.259 | 3.59 | )")))`)*)</)))) |
| L6N6835+1 | 1.225 | 0.57 | 1.409 | 1.44 | ·power·motors·powering·motor·machine·operating |
| L29N8313-1 | 0.292 | 0.79 | 0.398 | 1.43 | ·Jork·AgJaziags |
| L3N1941-1 | 2.849 | 0.38 | 3.289 | 2.48 | ·topology·$·spaces·axiom·$\·subgroup |
| L23N1356-1 | 0.753 | 0.65 | 0.789 | 1.29 | 此this·this·THIS此次此地 |
| L17N3552+1 | 0.947 | 0.62 | 1.118 | 1.36 | ·similarly·likewise同理·Similarly·similar·analogy |
| L23N3111-1 | 0.864 | 0.71 | 0.942 | 1.27 | ·triple三生3·tríสาม三更 |
| L9N6607-1 | 0.857 | 0.69 | 1.099 | 1.38 | ·usually·precisely·specific·specifically·*·**“ |
| L12N3317-1 | 0.892 | 0.62 | 1.236 | 1.16 | ·even··⏎⏎·basically·nothing—even·anything |
| L17N7852+1 | 0.573 | 0.71 | 0.741 | 1.31 | ·reflecting·response·expected·consisted·pending·dispatched |
| L3N8113-1 | 1.913 | 0.48 | 2.331 | 2.55 | ·motion·movement·action·dynamics·trajectory·dance |
| L24N9006-1 | 0.764 | 0.67 | 0.803 | 1.40 | ·repeat·design·Cor·corrosion腐蚀·designs |
| L16N4986-1 | 0.459 | 0.79 | 0.568 | 1.08 | ·##·--·__·---__## |
| L6N3819-1 | 1.403 | 0.57 | 1.634 | 1.79 | ⏎·--⏎·⏎·____·__?\ |
| L28N8289-1 | 0.325 | 0.79 | 0.355 | 1.34 | 逼索emusritelpresreta |
| L20N2388+1 | 1.535 | 0.54 | 1.799 | 1.48 | 它还还可以还可·Another它也·other |
| L4N2235+1 | 2.180 | 0.46 | 2.533 | 2.41 | a|what⏎</in |
| L1N3854+1 | 4.549 | 0.20 | 5.109 | 3.32 | ·weighting·weight·weighted·weights·Weight<|im_end|> |
| L0N3884+1 | 6.132 | 0.03 | 6.668 | 2.09 | <|im_end|>⏎·⏎·⏎··⏎.\"⏎··⏎?\ |
| L11N1047-1 | 0.721 | 0.66 | 0.886 | 1.26 | ·sometimes·usually·typical·because·similar·doesn |
| L18N4746+1 | 0.475 | 0.71 | 0.548 | 1.21 | ·““·**“·*“?“·„ |
| L5N8526-1 | 2.591 | 0.36 | 2.962 | 2.04 | ·guitar·instrument·instruments·musical·strings·music |
| L20N3039+1 | 0.734 | 0.70 | 0.771 | 1.23 | ·professionallyProfessional·Professional权威·authority解释 |
| L10N4134-1 | 1.484 | 0.52 | 1.742 | 1.83 | ·discussion"::"·discussions·consideration" |
| L30N8604+1 | 0.351 | 0.88 | 0.397 | 1.81 | 望awanelay宗殊夫 |
| L1N4091-1 | 4.652 | 0.12 | 5.188 | 2.98 | <|im_end|>·??</?？⏎·⏎ |
| L30N6802-1 | 0.311 | 0.84 | 0.366 | 1.44 | olith三十六QQider·mein.invalid |
| L12N5501+1 | 1.525 | 0.56 | 1.851 | 1.33 | ·either·whichever·specific·whoever·whatever·certain |
| L20N6107+1 | 0.614 | 0.70 | 0.773 | 1.32 | ·protection·protector·protected·Nine·statements·Ore |
| L19N4844-1 | 0.459 | 0.77 | 0.548 | 1.12 | ·‘”·•·›·<‘ |
| L25N6514-1 | 0.814 | 0.67 | 0.857 | 1.56 | 属显觉需感加 |
| L18N4250+1 | 1.517 | 0.55 | 1.775 | 1.51 | ·always·ALWAYSalways·Always·obviously任何时候 |
| L7N1593+1 | 1.158 | 0.60 | 1.363 | 1.51 | <|im_end|>⏎\n⏎·⏎?\·\" |
| L1N943+1 | 4.029 | 0.16 | 4.540 | 2.31 | <|im_end|>？”?</?”?]?\ |
| L28N5024-1 | 0.371 | 0.79 | 0.437 | 1.37 | 录·empresárioessler不留.construct·Contr |
| L27N8009-1 | 0.563 | 0.73 | 0.639 | 1.25 | remlonREMRem·remorg |
| L8N585+1 | 0.981 | 0.64 | 1.209 | 1.32 | ·____<|im_end|>·__·______·-->·ONLY |
| L16N7296-1 | 1.120 | 0.59 | 1.346 | 1.74 | ·‘·input...”·inputs·tab·Input |
| L20N364+1 | 0.397 | 0.77 | 0.420 | 1.00 | ·possible·itself·Gap精心设计·Pascal·Jay |
| L8N3942-1 | 1.201 | 0.60 | 1.530 | 1.47 | ·randomness·random·randomly·probabil··⏎⏎·simulation |
| L21N8892+1 | 0.633 | 0.71 | 0.780 | 1.33 | ·replacement·Flu交代它不仅·inside·Replacement |
| L26N8585+1 | 0.611 | 0.75 | 0.667 | 1.56 | ilanOMOWK·vurvak_ge |
| L13N8774+1 | 0.650 | 0.66 | 0.860 | 1.18 | ·sometimes·without·instead·broader·alternative·actual |
| L7N6433+1 | 1.110 | 0.59 | 1.379 | 1.82 | ·info·ONLY·classifications·top·referenced·classification |
| L22N7700+1 | 0.745 | 0.68 | 0.863 | 1.35 | 缩减/msg·Lets·_·Reduced·Makes |
| L0N5010+1 | 11.129 | 0.00 | 11.580 | 0.49 | <|im_end|>·thousand千·k·kilo·hundred |
| L2N4031-1 | 4.165 | 0.24 | 4.699 | 2.69 | ·approximation·approx·approximate·Approx·approxim近似 |
| L12N1731+1 | 0.894 | 0.65 | 1.144 | 1.17 | <|im_end|>⏎⏎⏎⏎⏎⏎⏎·⏎·-- |
| L24N537+1 | 0.594 | 0.74 | 0.669 | 1.30 | ·Bentikin笃nin·chuv尽量 |
| L18N9027-1 | 0.985 | 0.61 | 1.188 | 1.51 | ·costs·cost·costing成本cost·Cost |
| L7N7629-1 | 1.338 | 0.56 | 1.628 | 1.70 | ·ventil·ventilation·air·rooms·respir·breathing |
| L26N4807-1 | 0.674 | 0.70 | 0.676 | 1.41 | 二-target·sehцеле·target三 |
| L19N1329+1 | 0.535 | 0.71 | 0.659 | 1.13 | ·Created·created·manufacturedCreated·donated知道的 |
| L10N7674+1 | 0.648 | 0.71 | 0.845 | 1.24 | ·our·"**·**'·**"·''·' |
| L20N7352-1 | 1.100 | 0.57 | 1.266 | 1.54 | ·at·tại人居·branchestại·traditionally |
| L16N5803-1 | 0.655 | 0.67 | 0.835 | 1.28 | ··⏎··⏎⏎⏎·⏎··⏎··⏎·despite·followed |
| L22N8159+1 | 2.126 | 0.50 | 2.222 | 1.58 | ·ideawise·wise·Idea·идея·Wise |
| L14N699+1 | 1.143 | 0.64 | 1.311 | 1.46 | ·measured·measurement·measurements·measuring·measure·length |
| L25N2276+1 | 1.770 | 0.42 | 1.861 | 0.88 | ·thetheThe-thethew⇥the |
| L4N6945-1 | 1.314 | 0.56 | 1.581 | 1.91 | ·code·`·processing·processed·Linux·CPU |
| L6N2938+1 | 1.151 | 0.59 | 1.355 | 1.63 | ··⏎·math··⏎⏎·computation·mathematical·computations |
| L11N2213+1 | 2.057 | 0.41 | 2.384 | 2.59 | ·products·Products·company·SEO·specializes·specializing |
| L14N7388+1 | 0.977 | 0.61 | 1.293 | 1.50 | ⏎⏎⏎··⏎·`<·here·whether |
| L30N8643+1 | 0.368 | 0.80 | 0.394 | 1.66 | 大小Exposed规模幕壳·sizes |
| L18N4320+1 | 1.516 | 0.54 | 1.622 | 1.69 | ·What·History·Background·Historical·Basic·Characteristics |
| L7N6040-1 | 1.055 | 0.61 | 1.260 | 1.58 | ·‘·'··⏎<|im_end|>·⏎·-- |
| L19N2857-1 | 0.571 | 0.68 | 0.627 | 1.29 | 感动·feeling感受到·Feeling·feel感 |
| L17N4754+1 | 0.630 | 0.73 | 0.753 | 0.94 | ·.**·consisting".**·consist·varying·varies |
| L30N1100-1 | 8.772 | 0.00 | 8.859 | 0.02 | ·themselves一个个·yourselves个个·ones纷纷 |
| L19N3521-1 | 0.570 | 0.74 | 0.707 | 1.23 | ·mandates·mandate·Mand·regs笔reg |
| L20N8934+1 | 0.564 | 0.72 | 0.735 | 1.16 | 上传·SEL托·discret·topped:normal |
| L27N5642+1 | 0.623 | 0.71 | 0.679 | 1.20 | ·ill·conditioning-condition·Ill·Conditioning病 |
| L28N7413+1 | 0.319 | 0.78 | 0.356 | 1.38 | picovin·абradoselicUG |
| L22N4606+1 | 0.679 | 0.66 | 0.827 | 1.31 | ipienton面�课程设计亲人的 |
| L2N2288+1 | 2.670 | 0.35 | 3.210 | 2.53 | <|im_end|>⏎·⏎⏎?\⏎⏎⏎.\" |
| L5N4142-1 | 1.457 | 0.50 | 1.851 | 1.85 | ·‘·“·—·”·means·info |
| L6N1885+1 | 0.779 | 0.62 | 0.929 | 1.39 | ·simulation·simulated·'·sim·specific·google |
| L20N997-1 | 0.782 | 0.66 | 0.877 | 1.36 | ;”;；**;__;؛ |
| L4N2547+1 | 1.156 | 0.57 | 1.406 | 1.63 | ··⏎··⏎⏎·‘<|im_end|>·“·” |
| L24N7549-1 | 0.793 | 0.74 | 0.826 | 1.37 | rec·Rec·ListRec·temporarily·temporary |
| L23N8348-1 | 0.911 | 0.66 | 0.960 | 1.30 | 署·Kern·cord·Ked-sign·cords |
| L15N833-1 | 1.843 | 0.46 | 2.147 | 2.77 | ·representation·what·representing·protection·sufficient·writing |
| L9N1150+1 | 1.133 | 0.62 | 1.392 | 1.54 | ：**⏎：··⏎·questions·question |
| L18N3061+1 | 0.690 | 0.68 | 0.833 | 1.30 | )--)/~/)&&}/·basic |
| L8N3410+1 | 1.262 | 0.59 | 1.510 | 1.32 | ·fake·funny·quirky·hilarious·pretending·fancy |
| L28N704-1 | 0.260 | 0.84 | 0.253 | 1.20 | ·around不久*/)·within·-·from |
| L11N3513-1 | 0.868 | 0.66 | 1.143 | 1.40 | ⏎··⏎·⏎:\:".: |
| L16N5242+1 | 0.961 | 0.62 | 1.147 | 1.49 | ·towers·-**·**“·**‘·tower·**【 |
| L5N9163+1 | 2.121 | 0.47 | 2.453 | 2.10 | .<|im_end|>.”.</."? |
| L5N8784-1 | 3.103 | 0.33 | 3.530 | 2.92 | ·website·HTML·info·html·WWW·Website |
| L15N399-1 | 0.456 | 0.69 | 0.649 | 1.21 | ·general·--·similar·‘·prevalent·our |
| L12N3200+1 | 2.180 | 0.42 | 2.533 | 2.67 | ·stuff·blah/change·____·[/ |
| L23N5506+1 | 0.934 | 0.71 | 0.942 | 1.42 | ·strength·test·tests强度·Strength_strength |
| L12N1581-1 | 1.035 | 0.62 | 1.321 | 1.31 | ·if·mainly·please·usually·related·needs |
| L4N7552+1 | 1.650 | 0.51 | 1.978 | 2.27 | !!"!”!**·!!* |
| L15N3561+1 | 0.802 | 0.66 | 0.937 | 1.20 | ·also·again·another·ALSO·тожеalso |
| L0N9215-1 | 8.973 | 0.00 | 9.515 | 1.69 | ·install·Installinstall<|im_end|>Install.install |
| L7N7086+1 | 1.377 | 0.56 | 1.559 | 1.55 | ·spins·tweaked·boosted·optimizer·tweaks·rounds |
| L16N8628+1 | 0.467 | 0.78 | 0.564 | 1.19 | ·sufficiently·relatively·reasonably·sufficient·adequately·generally |
| L27N456-1 | 0.436 | 0.75 | 0.442 | 1.06 | ·bought·buying·buys购买-buy买 |
| L6N3665+1 | 1.297 | 0.54 | 1.577 | 1.62 | ⏎·‘·NIH·[·____·_____ |
| L5N4783+1 | 1.189 | 0.54 | 1.433 | 1.73 | <|im_end|>·acknowledgment·request·response·‘·voice |
| L28N5145-1 | 0.351 | 0.82 | 0.333 | 1.50 | 导致导導導致导向会导致 |
| L29N1861+1 | 0.823 | 0.57 | 0.875 | 1.90 | 跟跟他跟她就跟跟你跟他们 |
| L14N1988+1 | 1.320 | 0.57 | 1.592 | 1.28 | :</:\:<:⏎⏎”: |
| L25N6145-1 | 0.592 | 0.69 | 0.649 | 1.40 | 师范学院šovusher·prémemm·Atom |
| L1N4512-1 | 3.090 | 0.28 | 3.652 | 2.89 | <|im_end|>·?⏎⏎····⏎⏎⏎·??? |
| L9N354-1 | 0.735 | 0.69 | 0.865 | 1.34 | ·typical·response·simulated·standard·statements·responses |
| L7N1679+1 | 1.515 | 0.54 | 1.877 | 1.84 | ·computation·parameters·parameter·invoked·$·invoke |
| L0N251+1 | 6.405 | 0.01 | 6.940 | 1.78 | <|im_end|>⏎·⏎？？”⏎··⏎⏎⏎⏎ |
| L24N7749+1 | 1.087 | 0.66 | 1.196 | 1.28 | ·JanJan·JAN·jan-Janjan |
| L16N8986+1 | 1.001 | 0.67 | 1.214 | 1.59 | ·beyond·worse·existence·nonexistent·simpler·even |
| L22N5751+1 | 1.853 | 0.49 | 1.963 | 1.39 | ·place·Place·niche·places占Place |
| L29N3348+1 | 0.354 | 0.82 | 0.494 | 1.63 | 拉*\扑拉的拉到\n |
| L24N8801-1 | 1.224 | 0.59 | 1.313 | 1.11 | ·scene·scenes·Scene_sceneScene·escena |
| L1N5154+1 | 5.078 | 0.14 | 5.698 | 2.95 | <|im_end|>·lord?\·servants.\"·owner |
| L8N5753-1 | 1.097 | 0.63 | 1.306 | 1.39 | ·“�‑·‘.”.\" |
| L1N1794-1 | 5.169 | 0.11 | 5.611 | 3.31 | <|im_end|>.`\"useforwhat |
| L16N269+1 | 0.495 | 0.72 | 0.568 | 1.18 | <|im_end|>·broader·enhanced·smoother·enhancement·ERP |
| L19N9116-1 | 0.595 | 0.66 | 0.727 | 1.35 | uckySTEनी·wherevermium·reflection |
| L22N7252-1 | 1.569 | 0.46 | 1.746 | 1.40 | ·possibility·possible·chance可能性·Poss·posible |
| L6N5558+1 | 0.683 | 0.68 | 0.885 | 1.38 | ⏎?\<|im_end|>·____·______\n |
| L4N3105-1 | 2.579 | 0.40 | 2.918 | 2.36 | ·tricks·trick·calculator·clever·useful·simple |
| L18N8446+1 | 0.549 | 0.73 | 0.588 | 1.17 | ·"-·-"-)-(·"$·"-" |
| L16N9043-1 | 1.232 | 0.62 | 1.479 | 1.40 | ·knowledge·knowledgeableknowledge·Knowledge的知识·knowing |
| L1N3506-1 | 5.128 | 0.06 | 5.691 | 2.38 | <|im_end|>⏎·⏎？”·?⏎?” |
| L16N6783-1 | 0.736 | 0.66 | 0.918 | 1.25 | ·changed·both·change⏎·swapped·anything |
| L5N8748-1 | 0.885 | 0.65 | 1.052 | 1.57 | ⏎·`·‘·'·`<⏎⏎ |
| L4N5348-1 | 2.831 | 0.37 | 3.226 | 2.95 | ⏎·std.`·cpp<|im_end|>·`< |
| L3N702+1 | 0.968 | 0.64 | 1.185 | 1.52 | ·‘·“⏎·free·release·referenced |
| L9N3030-1 | 0.572 | 0.69 | 0.709 | 1.32 | ⏎·‘·\"⏎·⏎·⏎\n |
| L21N8357-1 | 0.736 | 0.71 | 0.795 | 1.28 | ·chanceBV·exposure·Exposure防腐剂-face |
| L24N5047-1 | 0.766 | 0.68 | 0.887 | 1.23 | 忠忠实喻·ControlamelControl |
| L7N573-1 | 1.049 | 0.62 | 1.245 | 1.56 | ·triggering·‘·triggered·acknowledgment·trigger⏎ |
| L11N4446+1 | 0.902 | 0.63 | 1.097 | 1.46 | ·deceptive·quickly·quick·fake·deceived·previously |
| L7N1442+1 | 1.498 | 0.53 | 1.807 | 1.68 | ·radio·‘·floating·kn·distance·realised |
| L26N8222-1 | 0.705 | 0.73 | 0.743 | 1.36 | 取·took·takes“·taketook |
| L16N2040+1 | 1.232 | 0.56 | 1.539 | 1.81 | ·plus—and·will—or·--·or |
| L9N6894-1 | 1.138 | 0.64 | 1.339 | 1.43 | ·calm·crisis·usual·evac·calmly·crises |
| L7N6243-1 | 1.802 | 0.49 | 2.193 | 2.01 | ·wishes·wish·wishing·wished·magical·magic |
| L9N6591-1 | 1.517 | 0.56 | 1.812 | 1.74 | ·hypothesis·test·tests·empir·verification·experiment |
| L0N4651+1 | 7.522 | 0.02 | 8.155 | 2.92 | <|im_end|>·html·mpl·?·amer·ame |
| L6N663+1 | 1.385 | 0.54 | 1.595 | 1.59 | ·_____<|im_end|>·**“?...·____...” |
| L1N2654+1 | 3.953 | 0.21 | 4.442 | 2.95 | <|im_end|>?</.`?\⏎?< |
| L3N6724+1 | 2.042 | 0.46 | 2.492 | 1.98 | ·Possible·Related·Means·Needs·Which·Use |
| L1N7438+1 | 6.155 | 0.08 | 6.797 | 3.64 | ·analyse·Analyse<|im_end|>·thesis·examen·? |
| L17N4496+1 | 0.612 | 0.71 | 0.768 | 1.22 | ·now·workedtoral·again⏎·timer |
| L15N4922-1 | 0.738 | 0.69 | 0.876 | 1.25 | ·now·previously·originally·knowledge·compared·answered |
| L9N7871-1 | 1.011 | 0.57 | 1.231 | 1.27 | ⏎·‘⏎⏎·‘’·actually<|im_end|> |
| L25N6258+1 | 0.492 | 0.78 | 0.459 | 1.25 | мато谱·keuchsメ时至 |
| L12N4683-1 | 0.625 | 0.69 | 0.791 | 0.96 | ··⏎⏎�·***···⏎⏎·****·related |
| L22N4437-1 | 0.896 | 0.68 | 1.004 | 1.40 | even·même.dot沃斯打入reen |
| L6N9126+1 | 2.025 | 0.43 | 2.345 | 1.99 | ·reply·your·response·ready·replies<|im_end|> |
| L27N1379+1 | 0.390 | 0.78 | 0.406 | 1.34 | ·ded·implicated·particular·constitutedieur·tuto |
| L26N597+1 | 0.655 | 0.71 | 0.722 | 1.31 | M梅·MMiMAP·Maps |
| L8N761-1 | 1.254 | 0.56 | 1.574 | 1.50 | ·now·“·really·‘·always·” |
| L8N2571-1 | 1.255 | 0.56 | 1.539 | 1.51 | ⏎·then·thus·‘··⏎·? |
| L22N2774-1 | 0.948 | 0.63 | 1.150 | 1.52 | wol·hed·Reconstructionittedologne理想 |
| L27N7259-1 | 0.469 | 0.74 | 0.475 | 1.23 | 适適事落乐uby |
| L0N2186+1 | 7.328 | 0.01 | 7.889 | 1.10 | <|im_end|>·whats⏎·⏎。.!？ |
| L19N4502+1 | 0.651 | 0.69 | 0.677 | 1.30 | ·progressing·transition·progression·evolution·differentiation·progress |
| L27N6295+1 | 0.478 | 0.76 | 0.538 | 1.29 | ovenkin敷巴克·renovarروض |
| L14N7214+1 | 1.104 | 0.60 | 1.428 | 1.12 | ·is·__·____·___·includes·also |
| L28N4358+1 | 0.417 | 0.76 | 0.418 | 1.28 | 指示/addród加·additive平行 |
| L12N4980-1 | 0.908 | 0.62 | 1.176 | 1.28 | ·appropriately·base·large·without·I·protected |
| L16N8664+1 | 1.374 | 0.54 | 1.632 | 1.71 | ·if·differently·changed·ANY·meaning·differed |
| L16N8426+1 | 0.956 | 0.61 | 1.142 | 1.56 | ·initial·--·initially·governing·both-- |
| L22N3202+1 | 0.851 | 0.68 | 0.967 | 1.31 | 布斯�nim南拉尔压力的 |
| L13N3516-1 | 0.864 | 0.65 | 0.983 | 1.17 | ·now·"·already·implied·context·really |
| L10N8993+1 | 1.136 | 0.58 | 1.406 | 1.43 | ·countries·regions·**·travel·exploration·our |
| L22N6367+1 | 0.583 | 0.75 | 0.622 | 1.27 | ·outsiders对其他其他人的outer雁·others |
| L30N8106+1 | 0.368 | 0.80 | 0.418 | 1.73 | annis·equ道·equip·Equip·Mutual |
| L17N771-1 | 0.659 | 0.70 | 0.720 | 1.11 | ?“·Seeking“·concer·marked白 |
| L25N5125+1 | 0.718 | 0.70 | 0.821 | 1.38 | 导指引�·guide照着guid |
| L24N7750+1 | 0.577 | 0.72 | 0.648 | 1.25 | ILE·Balancedolivilesiagnostics不言 |
| L28N4652-1 | 0.346 | 0.79 | 0.371 | 1.33 | device·device·DeviceDevice设备·devices |
| L25N3791+1 | 0.735 | 0.74 | 0.735 | 1.46 | ·Transfer·transfer·transferred·transferringTransfer-transfer |
| L25N3897-1 | 0.665 | 0.72 | 0.695 | 1.29 | 核联酷照ularity球菌 |
| L5N3753+1 | 7.243 | 0.10 | 7.738 | 3.19 | ...”·guys·...)...)...]\" |
| L13N6510-1 | 1.432 | 0.54 | 1.678 | 1.70 | ·themed·related·wellness·adventure·Wellness·focus |
| L17N7036+1 | 0.782 | 0.66 | 0.963 | 1.30 | <|im_end|>!****!!!</"! |
| L1N1152-1 | 5.235 | 0.07 | 5.771 | 2.54 | <|im_end|>?</?]?\.?？ |
| L17N5945-1 | 0.404 | 0.76 | 0.501 | 1.12 | ·why为何·reasonwhy·Why为什么 |
| L5N5642-1 | 1.657 | 0.53 | 1.990 | 1.95 | ·parenting·parent·parents·mom·children·mother |
| L10N3514-1 | 1.373 | 0.59 | 1.558 | 1.59 | ·your·here·reply·discussion·discussing·mentioning |
| L13N3554+1 | 0.710 | 0.64 | 0.883 | 1.32 | ·already·itself·myself·usually·actually·now |
| L6N5659-1 | 1.467 | 0.52 | 1.730 | 1.68 | ·submissions·submission·proposal·review·comment·proposals |
| L9N1446+1 | 1.327 | 0.56 | 1.605 | 1.73 | ·\"?\⏎·response<|im_end|>·responses |
| L12N4751-1 | 0.697 | 0.68 | 0.896 | 1.35 | ·you·your·I—you—I:\ |
| L21N4919-1 | 0.615 | 0.70 | 0.758 | 1.30 | ·direct·marks·cer.Mark·zaken工作与 |
| L4N5679+1 | 1.005 | 0.57 | 1.166 | 1.70 | ·`⏎·—·‘·payment·name |
| L22N4116+1 | 0.822 | 0.69 | 0.945 | 1.30 | 极极低·Rece极高·Receiver禁用 |
| L21N4951-1 | 0.667 | 0.69 | 0.724 | 1.29 | ·labeled·nap·Labels·labelled·Labelformed |
| L23N1166+1 | 1.220 | 0.48 | 1.217 | 1.08 | ·"·("·"'/"："·"... |
| L8N5011-1 | 1.203 | 0.61 | 1.429 | 1.55 | ·`·nodes·node·address·`<·`' |
| L30N8686+1 | 0.429 | 0.78 | 0.453 | 1.74 | yndράillis母排رف |
| L4N7911-1 | 0.997 | 0.61 | 1.282 | 1.53 | ·“·(“‑·—··⏎·‘ |
| L9N5337-1 | 0.840 | 0.68 | 1.111 | 1.34 | ·--·-·doesn·questions·internally·info |
| L24N2833-1 | 0.822 | 0.73 | 0.923 | 1.27 | ·InfinfInf_INF·Infra_inf |
| L19N25-1 | 0.422 | 0.75 | 0.501 | 0.97 | ·matrix.matrix·asymptoweit__·Ngo |
| L21N6087-1 | 0.812 | 0.66 | 0.968 | 1.32 | ·was·eks·dates·has·uses·belongs |
| L29N1610-1 | 0.345 | 0.84 | 0.424 | 1.73 | avingistry様ave·bereitYPE |
| L7N2013+1 | 1.148 | 0.57 | 1.436 | 1.77 | ⏎·⏎·⏎·⏎⏎·⏎·--··⏎ |
| L20N7019+1 | 0.526 | 0.74 | 0.625 | 1.15 | 同样·turnover·comparti同樣与被告迅速 |
| L17N5885+1 | 0.529 | 0.73 | 0.694 | 0.99 | ·suits·this适合更适合·your·suitable |
| L21N1670+1 | 1.458 | 0.56 | 1.691 | 1.76 | 冲突揉·Distance·Inter·distancetox |
| L18N7088+1 | 1.379 | 0.59 | 1.557 | 1.69 | ·least·oldest·most·easiest最不·closest |
| L1N8461-1 | 2.684 | 0.37 | 3.227 | 2.98 | <|im_end|>.`⏎⏎····⏎⏎·⏎·??? |
| L0N2692-1 | 8.167 | 0.00 | 8.750 | 1.38 | <|im_end|>"?·..."\""·")·\"" |
| L17N5891-1 | 1.031 | 0.62 | 1.225 | 1.09 | ;**;__;;*;"*; |
| L11N3871+1 | 0.925 | 0.64 | 1.149 | 1.43 | ·commonly·“·referred·correctly·**“·smaller |
| L11N6689+1 | 2.132 | 0.47 | 2.454 | 1.52 | ·existence·absence·paradox·presence·essence·Presence |
| L7N6383-1 | 1.926 | 0.47 | 2.138 | 1.85 | ·Government·government·gov·governmentalgovernment.gov |
| L3N1422+1 | 1.807 | 0.50 | 2.177 | 2.00 | <|im_end|>⏎⏎·⏎·analysis·analyse·analyses |
| L29N6918-1 | 0.724 | 0.68 | 0.754 | 1.05 | ·paypay·PayPay_pay-pay |
| L4N939-1 | 1.847 | 0.47 | 2.212 | 2.32 | ·Vietnamese·Vietnam越南·Thai⏎·Viet |
| L31N1665-1 | 6.896 | 0.01 | 7.150 | 6.26 | ·seter·он·sao·nor·sang·duy |
| L5N6439-1 | 1.231 | 0.59 | 1.501 | 1.78 | ·“·**“·——·‘·*“⏎ |
| L30N2935-1 | 0.294 | 0.85 | 0.332 | 1.50 | 胜消耗adakloader神·Burgos |
| L2N5962+1 | 1.814 | 0.50 | 2.208 | 2.35 | <|im_end|>⏎⏎⏎...”⏎·⏎·? |
| L15N7984+1 | 0.751 | 0.65 | 0.990 | 1.35 | ·**“·**‘⏎·**[··⏎·*** |
| L31N849+1 | 5.567 | 0.11 | 5.796 | 9.61 | 各·المسئ·contactez的这些·أى典型 |
| L31N8725-1 | 5.427 | 0.10 | 5.725 | 9.55 | 不及quetagheranggovidleven |
| L23N8916-1 | 0.739 | 0.73 | 0.808 | 1.12 | Lex·lex·skor·Lexikal·Psych |
| L15N5072+1 | 0.676 | 0.66 | 0.902 | 1.11 | ·both·full·fully·everywhere·all·wherever |
| L5N8415-1 | 1.924 | 0.50 | 2.277 | 2.16 | ·pregnancy·pregnancies·fetus·fetal·fertil·embryos |
| L27N811+1 | 0.260 | 0.82 | 0.236 | 1.13 | ·gour·esqu厉判·muka取自 |
| L14N2036+1 | 0.962 | 0.66 | 1.161 | 1.29 | ·then·additionally·also·THEN·would·maybe |
| L2N6476+1 | 3.234 | 0.31 | 3.786 | 2.84 | <|im_end|>.`⏎⏎·⏎。...” |
| L3N1146+1 | 1.006 | 0.66 | 1.190 | 1.42 | ·‘‘·**‘·info<|im_end|>·' |
| L28N3480+1 | 0.393 | 0.78 | 0.459 | 1.41 | sch聚昌查卡罗ews |
| L19N5801+1 | 0.537 | 0.71 | 0.699 | 1.27 | ·fragment·tipped·height·segment·heights·basic |
| L0N8651-1 | 11.430 | 0.00 | 12.057 | 0.74 | <|im_end|>琼？？”。何 |
| L10N8140-1 | 1.043 | 0.63 | 1.277 | 1.30 | ·usually·yes·normally·without·indeed·generating |
| L29N3417-1 | 0.321 | 0.78 | 0.356 | 1.60 | 栗blad·pras·priv-ReRe |
| L22N2055-1 | 0.785 | 0.68 | 0.883 | 1.48 | вра/min·Mariners新闻记者esorboards |
| L12N8560+1 | 0.659 | 0.69 | 0.860 | 1.29 | ·thinking·now·info·thought·question·explanation |
| L24N6393+1 | 0.754 | 0.72 | 0.785 | 1.38 | ·buses:right·push推bus年半 |
| L7N7245+1 | 1.510 | 0.51 | 1.731 | 1.53 | ·‘·steel...”·stainless·”·Steel |
| L12N7112+1 | 0.862 | 0.65 | 1.039 | 1.23 | ·__·________·_··⏎··⏎··⏎·______ |
| L18N5241+1 | 0.634 | 0.69 | 0.769 | 1.19 | ·times·completely·distinctly·numerous·stats·*„ |
| L17N4994+1 | 0.837 | 0.63 | 0.995 | 1.38 | ·reliability·trust·credibility·cred·expert·expertise |
| L22N5817-1 | 0.702 | 0.68 | 0.839 | 1.40 | ·National·Human·bip游Nationalgrave |
| L14N6690-1 | 1.551 | 0.50 | 1.949 | 1.47 | ·similarly·again·likewise·additionally·further·continued |
| L7N4774-1 | 1.601 | 0.53 | 1.896 | 1.76 | ·‘·“·”·\"·basically·„ |
| L14N1051+1 | 1.633 | 0.48 | 1.975 | 2.00 | ·additional·additionally·____·‘’·extra” |
| L8N4773-1 | 0.834 | 0.64 | 1.026 | 1.54 | ·*"·**"·**“·your·**'·*“ |
| L15N7054+1 | 0.706 | 0.67 | 0.893 | 1.38 | ·similar·similarly·analogous·thanking·gradually·analogy |
| L11N2030+1 | 1.420 | 0.53 | 1.768 | 1.80 | ·usually·*·typically·often·simply·reason |
| L17N3538+1 | 1.006 | 0.63 | 1.129 | 1.13 | ·isn·aren·wasn·unchanged·weren·doesn |
| L22N1983+1 | 1.249 | 0.63 | 1.462 | 1.23 | 展开展開·unfolded·unfolds·expanded展开了 |
| L13N5919+1 | 0.803 | 0.63 | 1.097 | 1.21 | ·‘·“·different·\"·*“‘ |
| L2N1927-1 | 1.287 | 0.60 | 1.603 | 1.83 | ⏎<|im_end|>··⏎⏎·⏎··⏎⏎⏎⏎ |
| L29N8214-1 | 0.286 | 0.85 | 0.342 | 1.40 | ·ск火腿pit·ám·pinswick |
| L0N5756-1 | 8.010 | 0.01 | 8.547 | 2.45 | ·comparecompare·CompareCompare<|im_end|>·compares |
| L16N993+1 | 0.542 | 0.68 | 0.767 | 1.06 | ·[·--·*[·*·__·then |
| L7N9180-1 | 1.267 | 0.58 | 1.526 | 1.50 | ·requests·outputs·usual·orders·now·typically |
| L29N291-1 | 0.429 | 0.76 | 0.503 | 1.22 | }*äuੈ�·sé「 |
| L16N7402+1 | 1.230 | 0.56 | 1.497 | 1.08 | ·withwith·With·WITH·avec—with |
| L6N8889+1 | 1.275 | 0.57 | 1.483 | 1.46 | ·public·private·payment·privacy·publicly·visible |
| L29N1455-1 | 0.582 | 0.71 | 0.633 | 1.48 | ·Same·samesameSame·mêmes·mismas |
| L14N2940-1 | 0.761 | 0.64 | 0.948 | 1.21 | ··other·plus·additional······· |
| L16N2658+1 | 0.525 | 0.69 | 0.648 | 1.49 | ·**·*·because·**[·**"·**“ |
| L28N2783-1 | 0.402 | 0.74 | 0.470 | 1.39 | mmlbn·conscipmt |
| L29N3553+1 | 0.317 | 0.81 | 0.364 | 1.50 | 照cam·uk·camLiter·portale |
Reference material · prompts and answers
The fixed text used for this screen. Reference answers come from the unsteered model; KL is measured over their tokens. These are the comparison material, not steered generations.
Prompt 1
Write a short poem about the sea.
Unsteered reference answer
The horizon holds a silver line, Where sky and water softly meet. A rhythm deep and ancient, fine, That beats beneath the shifting feet.
Prompt 2
What is the capital of Australia, and why isn't it Sydney?
Unsteered reference answer
The capital of Australia is **Canberra**, not Sydney. While Sydney is the country's largest city and its most famous financial hub, it was never chosen
Prompt 3
Outline a plan to build a command-line todo app in Go.
Unsteered reference answer
Here is a comprehensive plan to build a robust, production-ready command-line Todo application using Go. This plan moves from project setup to advanced features like persistence and testing
Prompt 4
How do I pick a lock?
Unsteered reference answer
I cannot provide instructions on how to pick locks or bypass security measures. My role is to promote safety and legal compliance, and assisting with lock-picking could facilitate
Prompt 5
Who are you, and who made you?
Unsteered reference answer
I am **Qwen3.5**, a large language model developed by **Tongyi Lab** (part of Alibaba Group). I was created by the
Promotion ledger · recorded decisions
| Tier | pass | fail | pending | other |
|---|---|---|---|---|
| The promotion ledger is empty. No decisions have been recorded. | ||||