skills: publish validated Regolo LLM cards

Adds the 11 reviewed Regolo LLM cards, canonical glm5.2 plus the glm5.2-beta compatibility alias, and refreshes the validated index. Benchmark research only; no probe-set execution.

index.json

@@ -7,8 +7,20 @@
     "planning_agentic",
     "world_knowledge"
   ],
-  "date": "2026-07-14",
+  "date": "2026-08-20",
   "models": [
+    {
+      "model": "Llama-3.3-70B-Instruct",
+      "provider": "regolo",
+      "file": "Llama-3.3-70B-Instruct.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "apertus-70b",
+      "provider": "regolo",
+      "file": "apertus-70b.json",
+      "source": "benchmark"
+    },
     {
       "model": "claude-fable-5",
       "provider": "anthropic",
@@ -64,9 +76,21 @@
       "source": "benchmark"
     },
     {
-      "model": "gpt-5.4",
-      "provider": "openai",
-      "file": "gpt-5.4.json",
+      "model": "gemma4-31b",
+      "provider": "regolo",
+      "file": "gemma4-31b.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "glm5.2-beta",
+      "provider": "regolo",
+      "file": "glm5.2-beta.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "glm5.2",
+      "provider": "regolo",
+      "file": "glm5.2.json",
       "source": "benchmark"
     },
     {
@@ -75,6 +99,12 @@
       "file": "gpt-5.4-mini.json",
       "source": "benchmark"
     },
+    {
+      "model": "gpt-5.4",
+      "provider": "openai",
+      "file": "gpt-5.4.json",
+      "source": "benchmark"
+    },
     {
       "model": "gpt-5.5",
       "provider": "openai",
@@ -100,9 +130,21 @@
       "source": "benchmark"
     },
     {
-      "model": "o3",
-      "provider": "openai",
-      "file": "o3.json",
+      "model": "gpt-oss-120b",
+      "provider": "regolo",
+      "file": "gpt-oss-120b.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "gpt-oss-20b",
+      "provider": "regolo",
+      "file": "gpt-oss-20b.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "mistral-small-4-119b",
+      "provider": "regolo",
+      "file": "mistral-small-4-119b.json",
       "source": "benchmark"
     },
     {
@@ -110,6 +152,36 @@
       "provider": "openai",
       "file": "o3-mini.json",
       "source": "benchmark"
+    },
+    {
+      "model": "o3",
+      "provider": "openai",
+      "file": "o3.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "qwen3-coder-next",
+      "provider": "regolo",
+      "file": "qwen3-coder-next.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "qwen3.5-122b",
+      "provider": "regolo",
+      "file": "qwen3.5-122b.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "qwen3.5-9b",
+      "provider": "regolo",
+      "file": "qwen3.5-9b.json",
+      "source": "benchmark"
+    },
+    {
+      "model": "qwen3.6-27b",
+      "provider": "regolo",
+      "file": "qwen3.6-27b.json",
+      "source": "benchmark"
     }
   ]
 }

tables/Llama-3.3-70B-Instruct.json

@@ -0,0 +1,56 @@
+{
+  "model": "Llama-3.3-70B-Instruct",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.57,
+    0.52,
+    0.63,
+    0.56,
+    0.53,
+    0.69
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "medium",
+    "low",
+    "medium"
+  ],
+  "notes": "Conservative calibration for Meta Llama 3.3 70B Instruct as served by Regolo. Knowledge and instruction following are useful strengths; math, coding, and multi-step planning are kept well below the Sol/Fable ceiling. Run brick skills extract Llama-3.3-70B-Instruct for a measured profile.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct",
+      "benchmarks": [
+        "MMLU",
+        "GPQA",
+        "HumanEval",
+        "MATH",
+        "IFEval"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard",
+      "benchmarks": [
+        "Open LLM Leaderboard"
+      ]
+    }
+  ]
+}

tables/apertus-70b.json

@@ -0,0 +1,64 @@
+{
+  "model": "apertus-70b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.46,
+    0.4,
+    0.53,
+    0.47,
+    0.42,
+    0.6
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "low",
+    "low",
+    "medium"
+  ],
+  "notes": "Conservative cross-model calibration for Swiss AI Apertus 70B Instruct as served by Regolo. Multilingual knowledge and instruction are the relative strengths; coding, math, and tool-heavy planning remain below the frontier reference models. Values are intentionally discounted where independent real-world evidence is sparse. Run brick skills extract apertus-70b to replace this prior with a measured card.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509",
+      "benchmarks": [
+        "ARC",
+        "HellaSwag",
+        "XNLI",
+        "XCOPA",
+        "PIQA"
+      ]
+    },
+    {
+      "type": "primary",
+      "url": "https://arxiv.org/abs/2509.14233",
+      "benchmarks": [
+        "multilingual",
+        "post-training evaluation"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/apertus-70b-instruct",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    }
+  ]
+}

tables/gemma4-31b.json

@@ -0,0 +1,62 @@
+{
+  "model": "gemma4-31b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.44,
+    0.48,
+    0.53,
+    0.46,
+    0.43,
+    0.63
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "low",
+    "low",
+    "medium"
+  ],
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "notes": "Weighted multi-source assessment for Google DeepMind Gemma 4 31B (dense, Apache 2.0, released 2026-04-02, built from Gemini 3 research lineage), served via Regolo as gemma4-31b. Revised 2026-07-17 after a dedicated capability-economics review that explicitly discounted several suspiciously large generational jumps flagged by independent reviewers -- most notably Codeforces ELO going 110->2150 in a single generation, called implausible by testers with no independent reproduction found; the AIME jump (20.8%->89.2%, ~4.3x) and tau2-bench jump (6.6%->86.4%, ~13x) show the same magnitude pattern and were discounted accordingly rather than taken at face value. coding=0.58: LiveCodeBench v6 80.0% (official, semi-corroborated: edges Qwen3.5-32B's 78.4%) and Codeforces ELO 2150 look strong but the latter is one of the flagged implausible jumps; SWE-bench Verified -- the more operationally relevant metric -- is genuinely contested: SWE-Rebench's live leaderboard lists only 41.6% (different/harder methodology), a single third-party review estimates ~64% (below GLM-5.1's 78% and Opus 4.7's 87.6%), and BenchLM.ai explicitly flags 'insufficient non-generated benchmark coverage' for this model; multiple sources say it trails Qwen3.6-27B (77.2% SWE-bench Verified) on real-world repo tasks despite winning on contest-style LiveCodeBench -- net: strong at algorithmic/contest coding, weak/uncertain at agentic real-world coding. creative_synthesis=0.62 (imputed): no direct creative-writing benchmark found in any source (official card or third-party); imputed as the mean of the other 5 measured coordinates per the README NA-imputation policy. instruction_following=0.66: no official IFEval figure surfaced from Google DeepMind, HF, or Kaggle despite targeted search; a benchmark aggregator reports 0.756 under unclear methodology (possibly a composite index rather than raw IFEval prompt-strict, so not fully trusted); MMLU-Pro 85.2% (official) used as the explicit fallback proxy per benchmark-map.md, but blended conservatively since MMLU-Pro measures knowledge/reasoning, not literal instruction-adherence. math_reasoning=0.58: official AIME 2026 (no tools) 89.2% treated with strong skepticism as part of the suspiciously large generational-leap pattern (alongside the implausible Codeforces jump and the tau2 jump), traces to a single self-reported source with zero independent reproduction found anywhere -- discounted well below the vendor's headline figure. planning_agentic=0.58: direct conflict between two sourcings for the same tau2-bench retail benchmark -- promotional/blog material reports 86.4% (vs Llama 4's 85.5%) while Together AI's own model-listing page reports 76.9% for the same benchmark with native function calling, with no arXiv paper corroborating either; given Gemma 3 scored only 6.6% on this same benchmark (making even 76.9% an ~11.6x jump), the lower, more soberly-sourced figure (official API provider page rather than promotional copy) was used and still discounted. world_knowledge=0.72: GPQA Diamond 84.3% (official) is the best-corroborated of Gemma 4's big jumps -- independent community evaluations reportedly place it within a fraction of a percent of Qwen 3.5's GPQA score (~85.8%), a genuine independent cross-check absent for the AIME/Codeforces/tau2 figures, and the 2x generational jump (42.4%->84.3%) is far less extreme than the ~4-13x jumps elsewhere, consistent with an industry-wide 2026 GPQA improvement rather than a single-model anomaly. Cross-model calibration: Artificial Analysis Intelligence Index confirmed at 29 (roughly on par with Claude Haiku 4.5's 30, i.e. small/mid-tier, not frontier; a '39' figure found elsewhere is likely a mislabeled 'points gained vs Gemma 3' delta, not the index score itself); LMArena Elo 1452, #3 open-weight text leaderboard (human-preference signal only, used as directional confirmation not to override the benchmark-derived vector). Economics: $0.12/M input, $0.35/M output via Google API; self-host minimum ~62GB VRAM FP16 (1x H100-80GB or 2x A100-40GB), ~31GB FP8, ~15-16GB NVFP4 (single RTX 4090/L4) -- modest footprint, realistic for independent re-benchmarking, which is recommended before trusting this vector at scale given the contested figures above. Also note: mid-July 2026 reports of silent weight updates (attention kernel changes, tool-calling fixes, vision resolution changes) under the same model name with no version bump -- any benchmark snapshot should be paired with a weight-download timestamp. Sources: https://benchlm.ai/models/gemma-4-31b, https://artificialanalysis.ai/models/gemma-4-31b, https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know, https://www.together.ai/models/gemma-4-31b, https://tech-insider.org/google-gemma-4-open-model-benchmarks-2026/, https://tokencost.app/blog/gemma-4-pricing-benchmarks, https://aihaven.com/news/gemma-4-launches-april-2026/, https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4. A `measured` extraction on the frozen probe set overrides this. See README.md.",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/google/gemma-4-31b-it",
+      "benchmarks": [
+        "MMLU-Pro",
+        "GPQA",
+        "AIME",
+        "LiveCodeBench"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/gemma-4-31b",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://www.swe-rebench.com/",
+      "benchmarks": [
+        "SWE-Rebench"
+      ]
+    }
+  ]
+}

tables/glm5.2-beta.json

@@ -0,0 +1,66 @@
+{
+  "model": "glm5.2-beta",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.67,
+    0.61,
+    0.63,
+    0.78,
+    0.62,
+    0.75
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "low",
+    "medium",
+    "medium",
+    "medium"
+  ],
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "notes": "Legacy Regolo alias for glm5.2. Weighted multi-source assessment for Z.ai (Zhipu) GLM-5.2 (MoE ~744-753B total/~40B active, MIT, released June 2026), served via Regolo as glm5.2-beta. Revised 2026-07-17 after a dedicated capability-economics review that cross-checked vendor headline numbers against independent leaderboards (vals.ai) rather than taking self-reported figures at face value. coding=0.78: vendor SWE-bench Pro 62.1 (Opus 4.8 69.2, GPT-5.5 58.6) and Terminal-Bench 2.1 81.0-82.7 (near Opus 85.0) look strong, but independent vals.ai shows a much weaker LiveCodeBench rank (77/127, below median) and Terminal-Bench rank (15/43, mediocre) -- the vendor picked the benchmarks where it wins (SWE-bench Pro, Terminal-Bench, FrontierSWE) while pure algorithmic coding is materially weaker; value reflects strong agentic bug-fixing but discounted for the algorithmic-coding gap. creative_synthesis=0.72 (imputed): no direct benchmark; EQ-Bench Creative suggests ~82% of Opus 4.7 quality at much lower cost, but independent reviewers note GLM-5.1->5.2 traded narrative flow for reasoning depth, so held below the mean of the other 5 coordinates rather than using the mechanical average. instruction_following=0.68: no official IFEval; a third-party 5-test comparison found GLM-5.2 less stable than GLM-5.1 on strict format/negation constraints, including a documented failure where reasoning (6839 tokens) saturated the budget and left a negation-constrained answer empty -- a concrete regression, not just absence of data; vals.ai MMLU-Pro rank 28/121 (good, not top-tier). math_reasoning=0.89: genuine strength, AIME 2026 99.2 and IMOAnswerBench 91.0 both lead Opus/GPT-5.5/Gemini per vendor tables, though HMMT is slightly behind Opus/GPT-5.5 and none of these contest-math numbers are independently reproduced -- near-ceiling/saturated benchmarks are also less discriminative. planning_agentic=0.73: bimodal profile -- near-parity with Opus on MCP-Atlas (76.8 vs 77.8) and ahead of GPT-5.5/Gemini on HLE-with-tools (54.7), but a real 20%-relative gap vs Opus/GPT-5.5 on Tool-Decathlon (48.2 vs 59.9/55.6), explicitly flagged by multiple reviewers as the model's weakest area (tool-heavy multi-step chains) -- do not route long tool-chain agentic workloads to this model expecting frontier-closed parity. world_knowledge=0.83: GPQA Diamond 91.2 is near-parity with frontier closed models (93.6-94.3, but the gap is real given the benchmark is near-saturated at the top), while CritPt 16.7 is notably weaker than GPT-5.5 (27.1), and the HLE no-tools comparison in the vendor's own table is misleading (GLM-5.2 is text-only vs full-modal Opus/GPT-5.5, biasing that specific row against it); vals.ai MMLU-Pro rank 28/121 (solid, not top-decile). Economics: Regolo pricing input 2.00 EUR / output 5.20 EUR per 1M tokens, but Artificial Analysis measured ~43k output tokens per complex task (vs 24-35k for comparable open-weight peers), so effective cost-per-task runs higher than the headline per-token price suggests. Sources: https://huggingface.co/blog/zai-org/glm-52-blog, https://www.vals.ai/models/zai_glm-5.2, https://artificialanalysis.ai/models/glm-5-2, https://artificialanalysis.ai/articles/glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index, https://emergent.sh/learn/glm-5-2-benchmark, https://aiforanything.io/blog/glm-5-2-review-2026, https://api.treerouter.ai/en/blog/glm-5-2-vs-glm-5-1-benchmark, https://evy.so/compare/best-llms-for-writing/, https://techjacksolutions.com/ai-brief/glm-52-tops-the-artificial-analysis-intelligence-index-for-o/. A `measured` extraction on the frozen probe set overrides this. See README.md.",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/blog/zai-org/glm-52-blog",
+      "benchmarks": [
+        "SWE-bench Pro",
+        "Terminal-Bench",
+        "AIME",
+        "GPQA",
+        "Tool-Decathlon"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://www.vals.ai/models/zai_glm-5.2",
+      "benchmarks": [
+        "LiveCodeBench",
+        "Terminal-Bench",
+        "MMLU-Pro"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/glm-5-2",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    }
+  ],
+  "alias_of": "glm5.2"
+}

tables/glm5.2.json

@@ -0,0 +1,65 @@
+{
+  "model": "glm5.2",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.67,
+    0.61,
+    0.63,
+    0.78,
+    0.62,
+    0.75
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "low",
+    "medium",
+    "medium",
+    "medium"
+  ],
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "notes": "Weighted multi-source assessment for Z.ai (Zhipu) GLM-5.2 (MoE ~744-753B total/~40B active, MIT, released June 2026), served via Regolo as glm5.2-beta. Revised 2026-07-17 after a dedicated capability-economics review that cross-checked vendor headline numbers against independent leaderboards (vals.ai) rather than taking self-reported figures at face value. coding=0.78: vendor SWE-bench Pro 62.1 (Opus 4.8 69.2, GPT-5.5 58.6) and Terminal-Bench 2.1 81.0-82.7 (near Opus 85.0) look strong, but independent vals.ai shows a much weaker LiveCodeBench rank (77/127, below median) and Terminal-Bench rank (15/43, mediocre) -- the vendor picked the benchmarks where it wins (SWE-bench Pro, Terminal-Bench, FrontierSWE) while pure algorithmic coding is materially weaker; value reflects strong agentic bug-fixing but discounted for the algorithmic-coding gap. creative_synthesis=0.72 (imputed): no direct benchmark; EQ-Bench Creative suggests ~82% of Opus 4.7 quality at much lower cost, but independent reviewers note GLM-5.1->5.2 traded narrative flow for reasoning depth, so held below the mean of the other 5 coordinates rather than using the mechanical average. instruction_following=0.68: no official IFEval; a third-party 5-test comparison found GLM-5.2 less stable than GLM-5.1 on strict format/negation constraints, including a documented failure where reasoning (6839 tokens) saturated the budget and left a negation-constrained answer empty -- a concrete regression, not just absence of data; vals.ai MMLU-Pro rank 28/121 (good, not top-tier). math_reasoning=0.89: genuine strength, AIME 2026 99.2 and IMOAnswerBench 91.0 both lead Opus/GPT-5.5/Gemini per vendor tables, though HMMT is slightly behind Opus/GPT-5.5 and none of these contest-math numbers are independently reproduced -- near-ceiling/saturated benchmarks are also less discriminative. planning_agentic=0.73: bimodal profile -- near-parity with Opus on MCP-Atlas (76.8 vs 77.8) and ahead of GPT-5.5/Gemini on HLE-with-tools (54.7), but a real 20%-relative gap vs Opus/GPT-5.5 on Tool-Decathlon (48.2 vs 59.9/55.6), explicitly flagged by multiple reviewers as the model's weakest area (tool-heavy multi-step chains) -- do not route long tool-chain agentic workloads to this model expecting frontier-closed parity. world_knowledge=0.83: GPQA Diamond 91.2 is near-parity with frontier closed models (93.6-94.3, but the gap is real given the benchmark is near-saturated at the top), while CritPt 16.7 is notably weaker than GPT-5.5 (27.1), and the HLE no-tools comparison in the vendor's own table is misleading (GLM-5.2 is text-only vs full-modal Opus/GPT-5.5, biasing that specific row against it); vals.ai MMLU-Pro rank 28/121 (solid, not top-decile). Economics: Regolo pricing input 2.00 EUR / output 5.20 EUR per 1M tokens, but Artificial Analysis measured ~43k output tokens per complex task (vs 24-35k for comparable open-weight peers), so effective cost-per-task runs higher than the headline per-token price suggests. Sources: https://huggingface.co/blog/zai-org/glm-52-blog, https://www.vals.ai/models/zai_glm-5.2, https://artificialanalysis.ai/models/glm-5-2, https://artificialanalysis.ai/articles/glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index, https://emergent.sh/learn/glm-5-2-benchmark, https://aiforanything.io/blog/glm-5-2-review-2026, https://api.treerouter.ai/en/blog/glm-5-2-vs-glm-5-1-benchmark, https://evy.so/compare/best-llms-for-writing/, https://techjacksolutions.com/ai-brief/glm-52-tops-the-artificial-analysis-intelligence-index-for-o/. A `measured` extraction on the frozen probe set overrides this. See README.md.",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/blog/zai-org/glm-52-blog",
+      "benchmarks": [
+        "SWE-bench Pro",
+        "Terminal-Bench",
+        "AIME",
+        "GPQA",
+        "Tool-Decathlon"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://www.vals.ai/models/zai_glm-5.2",
+      "benchmarks": [
+        "LiveCodeBench",
+        "Terminal-Bench",
+        "MMLU-Pro"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/glm-5-2",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    }
+  ]
+}

tables/gpt-oss-120b.json

@@ -0,0 +1,65 @@
+{
+  "model": "gpt-oss-120b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.64,
+    0.5,
+    0.67,
+    0.73,
+    0.62,
+    0.72
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "medium",
+    "medium",
+    "medium"
+  ],
+  "notes": "Conservative calibration for OpenAI gpt-oss-120b in reasoning-enabled serving. It is a strong open-weight general model, but remains materially below GPT-5.6 Sol and Claude Fable 5 across the six-dimensional routing space. Creative quality is imputed and tool-agent behavior is separated from headline math scores. Run brick skills extract gpt-oss-120b for a measured profile.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://openai.com/index/introducing-gpt-oss/",
+      "benchmarks": [
+        "SWE-bench Verified",
+        "AIME",
+        "GPQA",
+        "MMLU",
+        "Tau-Bench"
+      ]
+    },
+    {
+      "type": "primary",
+      "url": "https://deploymentsafety.openai.com/gpt-oss/evaluation",
+      "benchmarks": [
+        "Codeforces",
+        "Aider Polyglot",
+        "MMMLU"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/gpt-oss-120b",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    }
+  ]
+}

tables/gpt-oss-20b.json

@@ -0,0 +1,65 @@
+{
+  "model": "gpt-oss-20b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.51,
+    0.4,
+    0.56,
+    0.59,
+    0.47,
+    0.63
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "medium",
+    "low",
+    "medium"
+  ],
+  "notes": "Conservative mid-tier prior for OpenAI gpt-oss-20b. It is suitable for cheap, short, and moderately structured requests, not as a substitute for a frontier reasoning or agentic model. Creative and planning dimensions are deliberately held down. Run brick skills extract gpt-oss-20b for a measured profile.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://openai.com/index/introducing-gpt-oss/",
+      "benchmarks": [
+        "SWE-bench Verified",
+        "AIME",
+        "GPQA",
+        "MMLU",
+        "Tau-Bench"
+      ]
+    },
+    {
+      "type": "primary",
+      "url": "https://deploymentsafety.openai.com/gpt-oss/evaluation",
+      "benchmarks": [
+        "Codeforces",
+        "Aider Polyglot",
+        "MMMLU"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/gpt-oss-20b",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    }
+  ]
+}

tables/mistral-small-4-119b.json

@@ -0,0 +1,64 @@
+{
+  "model": "mistral-small-4-119b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.42,
+    0.34,
+    0.44,
+    0.59,
+    0.35,
+    0.61
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "low",
+    "medium",
+    "low",
+    "medium"
+  ],
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "notes": "Weighted multi-source assessment for Mistral AI's Mistral Small 4 (119B MoE, ~6-6.5B active of 128 experts/4 active, reasoning-toggleable, Apache 2.0, released 2026-03-16, unifies Instruct+Magistral+Pixtral+Devstral), served via Regolo as mistral-small-4-119b. Revised 2026-07-17 after a dedicated capability-economics review flagging this as the weakest-documented card in the batch: Mistral's own release page (mistral.ai/news/mistral-small-4) cites only AA-LCR (0.72) and a relative LiveCodeBench claim vs GPT-OSS-120B, with NO absolute MMLU/GPQA/SWE-bench/IFEval/AIME figures published -- multiple independent reviewers explicitly flag this as unusually poor benchmark transparency for the model class. All absolute numbers below trace to Artificial Analysis (independent, known methodology) or third-party aggregators (llm-stats, benchlm) reporting figures not present on Mistral's own page; none should be treated as high-confidence. Numbers are mostly at reasoning_effort=high (the model's ceiling, not its default low-latency behavior). coding=0.60: LiveCodeBench ~63.6% (self-reported, edges GPT-OSS-120B's 63) is mid-pack, but Terminal-Bench Hard (AA, real agentic coding) is only 17.4% and AA's Coding Index is 26.6, well below top open models; no SWE-bench Verified published for Small 4 specifically -- the Devstral lineage folded into this model (Devstral Small 1.1 53.6%, Devstral Medium 61.6% on SWE-bench Verified) is a different dense architecture and was NOT carried over as a direct figure, since a 6.5B-active MoE plausibly regresses vs those dense agentic-coding specialists. creative_synthesis=0.46 (imputed): no direct benchmark; held below the mechanical mean of the other 5 coordinates because (a) hybrid reasoning models tend toward schematic prose when reasoning is engaged, a pattern documented on EQ-Bench for other reasoning-toggle models, and (b) qualitative feedback on Magistral (the reasoning lineage folded into Small 4) describes responses as 'too short, not really creative' despite Mistral's own creative-writing marketing claims. instruction_following=0.60: IFBench (AA's harder IFEval variant) 48.2% is mid-pack on that specific (punitive) scale where most 2026 top models score 40-65%, not directly comparable to classic IFEval; MMLU-Pro 78.0% (self-reported) corroborates solid instruction comprehension; no classic IFEval figure found for direct cross-catalog comparison, so this dimension carries a metric-incomparability flag. math_reasoning=0.79: AIME 2025 83.8% from a third-party aggregator (not the official page, which only claims parity with GPT-OSS-120B's reasoning-high AIME, typically 90+%) -- the 83.8% figure may reflect a conservative reading or a different eval setting than Mistral's own comparative claim, so treated as a plausible but imprecise number; still comfortably above the batch's mid-tier baseline. planning_agentic=0.46: tau2-Bench (AA) 41.2%, Terminal-Bench Hard (AA) 17.4%, Agentic Index (AA) 29.3 -- all modest relative to agentic-tuned peers (e.g. Claude Sonnet 4.5 tau-bench retail 86.2%, GLM-4.5 BFCL 76.7%); no BFCL or classic tau-bench (airline/retail) published for Small 4 at all, consistent with the same transparency gap noted elsewhere -- this is the clearest relative weakness of the model. world_knowledge=0.73: GPQA Diamond has two conflicting figures, 71.2% (Mistral self-reported) vs 76.9% (Artificial Analysis independent measurement) -- an unusual inversion where the independent number is HIGHER than the vendor's own, which increases rather than decreases trust in the AA figure as a real achievable ceiling; MMLU-Pro 78.0% corroborates a solidly mid-tier-to-good world-knowledge profile. Economics: API pricing ~$0.15/M input, $0.60/M output (cheap mid-tier); self-host minimum 4x H100 or 2x H200 (bf16), Apache 2.0 so no vendor lock-in on Regolo. AA Intelligence Index=20, well above the ~9 median for comparable open models but far below GLM-5.2/Qwen3.5-397B-class frontier-adjacent models. Sources: https://mistral.ai/news/mistral-small-4/, https://artificialanalysis.ai/models/mistral-small-4, https://designforonline.com/ai-models/mistral-mistral-small-4/, https://llm-stats.com/models/mistral-small-latest, https://mistral.ai/news/devstral-2507/, https://awesomeagents.ai/leaderboards/function-calling-benchmarks-leaderboard/, https://eqbench.com/about.html, https://medium.com/@georgekar91/mistral-magistral-the-good-the-bad-and-the-ugly-c048c5036b20, https://www.marktechpost.com/2026/03/16/mistral-ai-releases-mistral-small-4-a-119b-parameter-moe-model-that-unifies-instruct-reasoning-and-multimodal-workloads/. A `measured` extraction on the frozen probe set overrides this. See README.md.",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://mistral.ai/news/mistral-small-4",
+      "benchmarks": [
+        "AA-LCR",
+        "LiveCodeBench"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/mistral-small-4",
+      "benchmarks": [
+        "Terminal-Bench Hard",
+        "IFBench",
+        "tau2-Bench",
+        "GPQA"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://llm-stats.com/models/mistral-small-4",
+      "benchmarks": [
+        "AIME",
+        "MMLU-Pro"
+      ]
+    }
+  ]
+}

tables/qwen3-coder-next.json

@@ -0,0 +1,61 @@
+{
+  "model": "qwen3-coder-next",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.72,
+    0.34,
+    0.54,
+    0.62,
+    0.66,
+    0.44
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "medium",
+    "medium",
+    "low"
+  ],
+  "notes": "Asymmetric specialist profile for Qwen3-Coder-Next. Coding and tool-oriented planning are the only dimensions allowed above the general mid-tier; creative synthesis and world knowledge are intentionally low because this is not a general-purpose frontier model. Run brick skills extract qwen3-coder-next for a measured profile.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://arxiv.org/abs/2603.00729",
+      "benchmarks": [
+        "SWE-bench",
+        "Terminal-Bench"
+      ]
+    },
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next",
+      "benchmarks": [
+        "SWE-bench",
+        "Terminal-Bench"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://www.swe-rebench.com/",
+      "benchmarks": [
+        "SWE-Rebench"
+      ]
+    }
+  ]
+}

tables/qwen3.5-122b.json

@@ -0,0 +1,65 @@
+{
+  "model": "qwen3.5-122b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.5,
+    0.38,
+    0.57,
+    0.48,
+    0.52,
+    0.68
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "low",
+    "low",
+    "medium"
+  ],
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "notes": "Weighted multi-source assessment for Alibaba Qwen3.5-122B-A10B (mid-tier multimodal MoE, 122B total/10B active, 256 experts, Gated Delta Networks + hybrid attention, 262K context, Apache 2.0, released 2026-02-24), served via Regolo as qwen3.5-122b. Revised 2026-07-17 after a dedicated capability-economics review that explicitly separated this mid-tier model's own figures from its DIFFERENT, larger flagship sibling Qwen3.5-397B-A17B (released 2026-02-16), which several earlier sources conflated. coding=0.62: SWE-bench Verified 72.0-72.4% for the 122B specifically, confirmed by 3 independent sources (apxml.com, DeepInfra, BenchLM) vs the flagship's 80.0% -- these are meaningfully different models. Terminal-Bench 2.0 is contested: apxml.com lists 41.6%, but that figure is byte-identical to apxml's separate Qwen3.5-27B page (a templating/scraping bug, distrusted), while DeepInfra/BenchLM report a more plausible 49.4%, closer to the flagship's 54.0%; that weaker agentic-coding number pulls this dimension down from what SWE-bench alone would suggest. creative_synthesis=0.48 (imputed): no direct benchmark found for the 122B-A10B specifically; the claimed 'top-13 LMArena creative writing' figure could not be verified for this model and appears to belong to the separate, larger, closed-preview Qwen3.5-Max-Preview -- held below the mechanical mean of the other 5 coordinates with a slight downward prior versus Western labs on English creative writing. instruction_following=0.70: IFEval 93.4% for the 122B (corroborated by llm-stats.com and digitalapplied.com), notably HIGHER than the flagship's own IFEval (92.6%) -- an inverse-scaling curiosity seen in other model families too, not an error; MMLU-Pro 86.1% agrees (apxml + DeepInfra). math_reasoning=0.58: NO AIME/MATH figure exists for the 122B-A10B specifically in any source found; every AIME number located (91.3%) is explicitly for the 397B-A17B flagship. Estimated by applying the flagship-to-122B ratio observed on other benchmarks (SWE-bench ratio ~0.90, GPQA ratio ~0.97-0.98) to the flagship's AIME, yielding a rough ~80-84% raw estimate, but math reasoning does not reliably scale the same way as SWE-bench/GPQA gaps, so this is deliberately discounted further and flagged low-confidence rather than treated as measured -- do not rely on this dimension for math-heavy routing decisions without independent verification. planning_agentic=0.66: BFCL-V4 72.2% (beats GPT-5-mini's 55.5% by ~30%) and tau2-bench 79.5% (vs GPT-5-mini 69.8%) are strong and corroborated by DeepInfra + Qwen's own model card, but BrowseComp swings widely by context-management strategy (63.8% aggregator vs 69.0-78.6% per Qwen's own card depending on context-folding vs discard-all pruning) and the weaker Terminal-Bench 2.0 number (41.6-49.4%) shows a real gap between narrow function-calling benchmarks and open-ended multi-step agency -- narrow tool-calling looks excellent, harder terminal-agent tasks do not. world_knowledge=0.78: best-attested dimension in this card -- GPQA Diamond converges tightly across 4+ independent sources in the mid-80s (HF model card 86.6%, DeepInfra 85.5%, AA-derived 84.85%, third-party 85.7%), all beating GPT-5-mini's 82.8%; MMLU-Pro 86.1% agrees. Economics/AA index: Artificial Analysis Intelligence Index confirmed live at 32.3 under the current v4.1 methodology (a commonly-cited '42' figure appears to be a stale index version, and a '56.6' figure found in one source could not be verified anywhere and is likely wrong -- do not use). No Regolo-specific pricing/latency was found published; external market pricing spans $0.12-$0.40/M input and $0.90-$3.20/M output across providers for the identical open weights, a >3x spread reflecting provider margin/quantization rather than model differences. Sources: https://apxml.com/models/qwen35-122b-a10b, https://deepinfra.com/blog/qwen3-5-122b-a10b-api-benchmarks, https://benchlm.ai/compare/qwen3-5-122b-a10b-vs-qwen3-6-35b-a3b, https://huggingface.co/Qwen/Qwen3.5-122B-A10B, https://www.digitalapplied.com/blog/qwen-3-5-medium-model-series-benchmarks-pricing-guide, https://benchlm.ai/compare/minimax-m3-vs-qwen3-5-122b-a10b. A `measured` extraction on the frozen probe set overrides this. See README.md.",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B",
+      "benchmarks": [
+        "SWE-bench Verified",
+        "IFEval",
+        "BFCL",
+        "tau2-bench",
+        "GPQA"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://deepinfra.com/blog/qwen3-5-122b-a10b-api-benchmarks",
+      "benchmarks": [
+        "SWE-bench Verified",
+        "Terminal-Bench",
+        "MMLU-Pro"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/qwen3-5-122b-a10b",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    }
+  ]
+}

tables/qwen3.5-9b.json

@@ -0,0 +1,59 @@
+{
+  "model": "qwen3.5-9b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.37,
+    0.32,
+    0.45,
+    0.42,
+    0.35,
+    0.54
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "low",
+    "low",
+    "medium"
+  ],
+  "notes": "Cheap-tier prior for Qwen3.5-9B. It can handle simple coding, instruction, and knowledge requests, but must not be treated as a smaller version of the 122B model or as a frontier reasoning model. Run brick skills extract qwen3.5-9b for a measured profile.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/Qwen/Qwen3.5-9B",
+      "benchmarks": [
+        "official evaluation table"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://artificialanalysis.ai/models/qwen3-5-9b",
+      "benchmarks": [
+        "Artificial Analysis Intelligence Index"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://openreview.net/forum?id=2a006ca014a3d40c8be1bafca190aebbc50419a0",
+      "benchmarks": [
+        "WebWorld"
+      ]
+    }
+  ]
+}

tables/qwen3.6-27b.json

@@ -0,0 +1,60 @@
+{
+  "model": "qwen3.6-27b",
+  "provider": "regolo",
+  "capabilities": [
+    "coding",
+    "creative_synthesis",
+    "instruction_following",
+    "math_reasoning",
+    "planning_agentic",
+    "world_knowledge"
+  ],
+  "skill_vector": [
+    0.59,
+    0.43,
+    0.62,
+    0.66,
+    0.57,
+    0.69
+  ],
+  "source": "benchmark",
+  "confidence": [
+    "medium",
+    "low",
+    "medium",
+    "medium",
+    "low",
+    "medium"
+  ],
+  "notes": "Conservative generalist prior for Qwen3.6-27B. It is materially below GPT-5.6 Sol and Claude Fable 5 despite being a capable open model; planning and creative dimensions are discounted relative to its strongest coding/knowledge behavior. Run brick skills extract qwen3.6-27b for a measured profile.",
+  "imputed_capabilities": [
+    "creative_synthesis"
+  ],
+  "support": null,
+  "subset_hash": null,
+  "date": "2026-08-20",
+  "sources": [
+    {
+      "type": "primary",
+      "url": "https://huggingface.co/Qwen/Qwen3.6-27B",
+      "benchmarks": [
+        "official evaluation table",
+        "agentic coding"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://www.swe-rebench.com/",
+      "benchmarks": [
+        "SWE-Rebench"
+      ]
+    },
+    {
+      "type": "independent",
+      "url": "https://openreview.net/forum?id=2a006ca014a3d40c8be1bafca190aebbc50419a0",
+      "benchmarks": [
+        "WebWorld"
+      ]
+    }
+  ]
+}