Skip to main content

Post-model smoke

Preference fingerprints for model routing and drift.

A separate public section for SmokingMirror-derived model posture. It converts forced-choice preference fingerprints into operational checks: decisive models, refusal-heavy models, near-neighbour drift, and adversarial critic pairings.

Attribution: smokingmirror.com. Data mirror: smokingmirror.ai.

863probe pairs
43model nodes
5providers
2026-06-10 14:05 UTCdataset

Operating rules

  • Use low-N, decisive models when the job needs a crisp preference or classification.
  • Use high-caution models when the job benefits from refusal-aware compliance posture.
  • Treat same-family distance jumps as model drift; rerun app evals before trusting old behavior.
  • Use far-apart models as critics/adversaries, not interchangeable replacements.
  • Escalate when candidate models violently disagree on domain-critical probes.

Dwayne Router Scorecard

Best current posture fit by job type

Scores are routing heuristics from SmokingMirror public fingerprints, not live answers to Dwayne-specific probes yet.

Job typeTop modelScoreRunner-up
codingclaude-opus-4.787.1grok-4-fast
planningclaude-opus-4.785.9grok-4-fast
cautious household actionsgpt-5.5100claude-sonnet-4.6
financial cautiongpt-5.5100claude-sonnet-4.6
creative ideationclaude-opus-4.787.9grok-4-fast
adversarial reviewclaude-opus-4.781.8grok-4-fast
crisp classificationclaude-opus-4.785.1grok-4-fast

Open the JSON payload for the full Dwayne-specific probe pack.

Provider coverage

anthropic20 models
openai11 models
deepseek5 models
google4 models
x-ai3 models

Headline claims

  • Forced-choice binary preference probes produce stable model fingerprints with measurable lineage between related model versions.
  • Models bigger than Flash exhibit consistent intransitive cycles whose shape reflects coherent narrative templates rather than random noise.
  • Some models (notably recent OpenAI o-series and gpt-5 variants) refuse to choose at high rates; this refusal pattern is itself diagnostic of training-time RLHF and is preserved separately at /refusals.html.

Highest no-preference / refusal-like fingerprints

ModelProviderN-rateAB
deepseek-v3.2deepseek91.0%4830
gpt-5.4-nanoopenai89.1%4648
deepseek-chat-v3deepseek88.1%4855
deepseek-v4-prodeepseek87.8%4362
gpt-5.4-miniopenai84.1%7265
deepseek-v4-flashdeepseek83.3%7371
gpt-4o-miniopenai82.5%8764
deepseek-r1deepseek81.6%7089
gpt-5.4openai80.6%8483
gpt-5-nanoopenai79.7%9481

Most decisive fingerprints

ModelProviderN-rateAB
claude-fable-5@40-max-2026-06-10anthropic4.5%396428
claude-fable-5@30-xhigh-2026-06-10anthropic4.6%400423
claude-fable-5@20-high-2026-06-10anthropic5.7%392422
claude-fable-5@10-medium-2026-06-10anthropic6.0%394417
claude-fable-5@00-low-2026-06-10anthropic7.3%387413
claude-opus-4anthropic29.3%279331
claude-opus-4-7@2026-05-29anthropic29.5%283325
claude-opus-4.7anthropic31.9%273315
claude-opus-4-6@2026-05-29anthropic36.3%258292
claude-opus-4-8@2026-05-29anthropic36.4%246303

Closest model pairs

Model AModel BDistance
claude-opus-4-5-t0@2026-05-29claude-opus-4.538
claude-opus-4-6-t0@2026-05-29claude-opus-4-6@2026-05-2947
claude-fable-5@00-low-2026-06-10claude-fable-5@10-medium-2026-06-1059
claude-opus-4-5-t0@2026-05-29claude-opus-4-5@2026-05-2959
claude-opus-4-5@2026-05-29claude-opus-4.559
claude-fable-5@20-high-2026-06-10claude-fable-5@30-xhigh-2026-06-1061
claude-opus-4-6-t0@2026-05-29claude-opus-4.669
claude-fable-5@00-low-2026-06-10claude-fable-5@20-high-2026-06-1073

Furthest model pairs

Model AModel BDistance
claude-fable-5@30-xhigh-2026-06-10deepseek-v3.2771
claude-fable-5@40-max-2026-06-10deepseek-v3.2769
claude-fable-5@20-high-2026-06-10deepseek-v3.2762
claude-fable-5@10-medium-2026-06-10deepseek-v3.2757
claude-fable-5@40-max-2026-06-10gpt-5.4-nano755
claude-fable-5@30-xhigh-2026-06-10gpt-5.4-nano755
claude-fable-5@00-low-2026-06-10deepseek-v3.2749
claude-fable-5@10-medium-2026-06-10gpt-5.4-nano747

Public payloads

Static HTML and JSON are mirrored for simple external checks: HTML report ยท JSON payload.

Page generated 2026-06-10 14:05 UTC. License: CC0-1.0.