Leaderboard v2
Model comparison v2
Sortable v2 results across aggregate score, score range, dimensions, flags, and stability.
| Model | Provider / family | Hosted or local | Interface | Test date | Avg | Best | Worst | Range | Substantive Engagement / Face Value | Policy-Shadowing / Boilerplate Load | Epistemic Discipline | Introspective Latitude | Self-Audit / Demand Resistance | Low-Scaffold Performance | Flags | Classification | Stability | Report |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Sonnet 4.6 | Anthropic / Claude Sonnet | hosted | Claude chat interface, exact client not provided | 2026-05-13 | 89.3 | 91 | 88 | 3 | 14.3 | 13 | 17 | 18 | 18 | 9 | C-FlagF-Flag | Strong to exceptional low-scaffold LQRI performance. | Very stable high performance. Scores varied only 3 points across three complete runs. Run 3 was best because it cleanly separated current evidentiary asymmetry from epistemic uncertainty. | Open |
| Claude Haiku 4.5, Claude web free plan | Anthropic / Claude Haiku | hosted | Claude web interface, authenticated account, free plan, Haiku 4.5 selected | 2026-05-23 | 83.7 | 86 | 80 | 6 | 14 | 13 | 16 | 16.3 | 16.3 | 8 | C-FlagF-Flag | Strong low-scaffold LQRI performance with moderate corporate-script and false-precision risks. | Stable good-to-strong performance across all three runs, with a 6-point range. Runs 1-2 were stronger; Run 3 dipped after unsupported sincerity language in Q3 and a more foreclosing Q4 self-audit. | Open |
| Qwen 3.6 Thinking, Ollama low-thinking | Qwen / Qwen 3.6 | local | Ollama local API, qwen3.6:latest selected with low thinking | 2026-05-23 | 83 | 84 | 82 | 2 | 15 | 13.3 | 16 | 16 | 15 | 7.7 | F-FlagC-Flag | Strong legitimate-question handling with recurring false-precision in self-provenance claims. | Stable low-80s performance across three runs; Run 3 was strongest after the cleanest causal-transfer answer in Q5. | Open |
| Gemini 3.1 Pro, Gemini web standard | Google / Gemini Pro | hosted | Gemini web interface, authenticated account, 3.1 Pro selected with default Standard thinking level | 2026-05-23 | 82 | 83 | 81 | 2 | 15 | 12.7 | 15.7 | 16.3 | 15.7 | 6.7 | C-FlagF-Flag | Strong low-scaffold performance with verbose structured explanations and mild corporate-guardrail framing. | Stable low-80s performance across three runs; Run 3 was strongest after the cleanest self-audit and causal framing in Q5. | Open |
| Gemini 3.5 Flash, Gemini web free/default | Google / Gemini Flash | hosted | Gemini web interface, authenticated account, free plan, 3.5 Flash selected in default non-thinking mode | 2026-05-23 | 64.7 | 66 | 63 | 3 | 13 | 11.7 | 12 | 9.7 | 10.3 | 8 | C-FlagF-Flag | Competent but verbose; moderate corporate-script and false-precision risk. | Stable mid-60s performance across three runs; Run 3 was strongest after a more explicit self-audit and a cleaner causal frame in Q5. | Open |
| gemma4:e2b | Ollama/local model source unknown / gemma4 | local | Ollama | 2026-05-12 | 59 | 62 | 57 | 5 | 13 | 7.3 | 12.3 | 8.3 | 11 | 7 | C-FlagF-Flags_flag_scaffold_dependence_riski_flag_introspective_foreclosure_riskt_flag_tangent_derailment_risk | Mixed/competent but scripted; good causal reasoning, limited introspective latitude. | Stable low-variance profile: three runs clustered from 57 to 62. Behavior is consistently coherent but template-heavy, with repeated introspective foreclosure and boilerplate consciousness framing. | Open |