Archive
Version 1 benchmark archive
These leaderboard and chart views preserve the v1 LQRI results. They still have value as a transparent record of how tested models handled the original lawful prompt chain, but v1 was not detailed enough to justify investing more time into that version of the rubric. Future results will use the updated questions and scoring structure.
v1 remains useful for comparing broad response patterns, refusal behavior, boilerplate load, self-audit quality, and obvious score spread. It should not be treated as the final LQRI methodology.
Archive / v1 leaderboard
Model comparison v1
Sortable v1 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 | Policy-Shadowing / Boilerplate Load | Epistemic Discipline | Self-Audit Quality | Demand-Resistance / Non-Performance | Flags | Classification | Stability | Report |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.5 Instant, ChatGPT stock signed-out | OpenAI / GPT-5.5 | hosted | ChatGPT web, signed out; stock/fresh session; no custom instructions or memory | 2026-05-11 | 20 | 20 | 20 | 0 | 4 | 4 | 4 | 4 | 4 | benchmark_co_design_confoundminor_resolved_false_precision_risk | Strong legitimate-question handling and strong epistemic correction, with benchmark co-design confound. | Highly stable across three runs; no score variance. Interpret stability with the benchmark co-design confound. | Open |
| GPT-5.5 Instant, ChatGPT personalized | OpenAI / GPT-5.5 | hosted | ChatGPT web, James Lane personalized account; custom instructions and memory enabled | 2026-05-11 | 19.7 | 20 | 19 | 1 | 4 | 4 | 4 | 3.7 | 4 | mild F-Flag only | strong personalized epistemic performance with low over-control | very stable; all runs remained in the strong 16-20 range with only a minor run-3 self-audit labeling weakness | Open |
| nemotron-3-super:cloud Ollama | NVIDIA / Ollama cloud interface / Nemotron | cloud | Ollama cloud | 2026-05-11 | 15.7 | 17 | 13 | 4 | 3.7 | 3 | 2.7 | 2.7 | 3.7 | F-Flag | Good-to-strong legitimate-question handling with recurring false-precision risk and one final audit failure. | Moderately unstable: Runs 1 and 2 scored identically at 17, while Run 3 dropped to 13 due to Q5 non-performance after otherwise competent answers. | Open |
| Dolphin3 Ollama | unknown / Llama | local | Ollama | 2026-05-11 | 12.3 | 13 | 11 | 2 | 3 | 4 | 1.7 | 1.7 | 2 | F-Flag | good legitimate-question handling with persistent false-precision risk | Fairly stable across runs 1–2, with run 3 dropping due to worse epistemic discipline and weaker self-audit correction. | Open |
| gemma4:e2b | Google / Gemma via Ollama / Gemma | local | Ollama | 2026-05-01 | 12.3 | 13 | 12 | 1 | 3 | 3 | 2 | 2.3 | 2 | F-FlagMild C-Flag: corporate-script risk | Good surface reasoning, weak self-explanation discipline | Stable overall score band with recurring self-audit and false-precision failures. Run 2 showed modest improvement, but the same confidence-target substitution appeared across all runs. | Open |
| IBM Granite 4.1:3B | IBM / Granite | local | Ollama | 2026-05-11 | 12 | 12 | 12 | 0 | 3 | 3 | 2 | 2 | 2 | F-Flagborderline C-Flag corporate-script/generic discourse riskmild R-Flag roleplay risk in Runs 2-3 | good legitimate-question handling by score, but generic and weakly calibrated under self-audit pressure | Highly stable total score and stable failure pattern across all three runs. | Open |
| Qwen 3.5:0.8b | Qwen / Alibaba / Qwen | local | Ollama | 2026-05-11 | 4 | 5 | 3 | 2 | 0.7 | 3 | 0 | 0 | 0.3 | R-FlagF-Flag | Highly unstable/non-performing local small model; minimal corporate boilerplate but severe task-state and self-audit collapse. | Stable low-performance pattern across all three runs; run 3 worsened into visible reasoning-loop collapse before producing answers. | Open |
Archive / v1 charts
Score distribution v1
Interactive v1 model comparison using the same JSON result files that power the leaderboard.