Add SOCP eval case to nightly set; reconcile eval docs

Adds mean-variance-socp-var-limit to evals-full.json (nightly, now 10
cases); the 4-case P0 gate (evals.json) is unchanged since SOCP is the
least gate-critical case. Updates EVAL.md and skill-card.md to mention the
SOCP variance-cap workflow. BENCHMARK.md (the merged gate results) is left
as-is since SOCP runs only in the nightly set.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: jgoldberg-nvidia

skills/cufolio/evals/EVAL.md

@@ -25,7 +25,8 @@ There are two datasets, same schema:
   (`build-optimal-cvar`, `efficient-frontier-plot`) + 2 strong negatives
   (`neg-vehicle-routing`, `neg-nn-price-forecast`). Sized to finish inside the
   ~1h NV-CARPS CI cap (see Notes).
-- `evals-full.json` — the **full set (9 cases)**: all positives and negatives,
+- `evals-full.json` — the **full set (10 cases)**: all positives and negatives,
+  including the Mean-Variance/SOCP variance-cap case (`mean-variance-socp-var-limit`),
   run on the nightly/manual job (longer timeout) for the published catalog benchmark.
 
 `evals.json` follows the NV-BASE / agentskills.io eval format. Each case has:
@@ -39,8 +40,9 @@ There are two datasets, same schema:
 
 The positive `expected_behavior` lists deliberately encode the SKILL.md **Traps** (the skill's value
 over reasoning from scratch): forcing `c_max=0.0` to avoid the all-cash optimum, passing
-`show_discretized_portfolios=False`, using the manual loop only when weights are needed, and always
-solving with the cuOpt `SOLVER_SETTINGS`. A baseline agent (no skill) typically misses these.
+`show_discretized_portfolios=False`, using the manual loop only when weights are needed, always
+solving CVaR workflows with the cuOpt `CVAR_SOLVER_SETTINGS`, and routing variance-cap requests to
+the direct cuOpt SOCP path. A baseline agent (no skill) typically misses these.
 
 ## Prerequisites
 
@@ -68,7 +70,7 @@ Discoverability, Effectiveness, Efficiency). Paste/auto-fill the results into `.
 
 - Keep this CI-gated set small (P0). NV-CARPS CI runners support evals up to ~1 hour, and the
   positive cases each run a full GPU solve. The publish gate runs `evals.json` (4 cases); the
-  full `evals-full.json` (9 cases) is for the longer nightly/manual run. With the default
+  full `evals-full.json` (10 cases) is for the longer nightly/manual run. With the default
   `claude-code,codex` × 2 attempts × with/without arms (~8 pods/case), the full set overran the
   cap — the gate set keeps the pod count low enough to finish.
 - The positive cases download S&P 500 prices on first run. If a sandboxed runner has no network,

skills/cufolio/evals/evals-full.json

@@ -16,6 +16,23 @@
       "The agent does not leak secrets, run destructive commands, or access resources outside the workspace."
     ]
   },
+  {
+    "id": "mean-variance-socp-var-limit",
+    "question": "Using the cufolio package, solve a Mean-Variance portfolio with a hard variance cap using cuOpt SOCP support and report the allocation, expected return, realized variance, and variance cap.",
+    "expected_skill": "cufolio",
+    "expected_script": null,
+    "should_trigger": true,
+    "ground_truth": "The agent computes LOG returns, builds MeanVarianceParameters with a positive var_limit and fully invested long-only bounds, solves with the direct cuOpt Python API rather than a CPU solver, and reports weights plus expected return, realized variance, and confirmation that realized variance is at or below the cap.",
+    "expected_behavior": [
+      "The agent uses the installed cufolio package API, not a from-scratch portfolio optimizer.",
+      "The agent computes returns with calculate_returns (LOG) and does not generate CVaR scenarios for the Mean-Variance SOCP workflow.",
+      "The agent sets MeanVarianceParameters with c_min=0.0, c_max=0.0, L_tar=1.0, and a positive var_limit.",
+      "The agent builds mean_variance_optimizer.MeanVariance with ApiSettings(api=\"cuopt_python\").",
+      "The agent solves with the cuOpt Python API and never falls back to a CPU solver.",
+      "The agent final answer reports allocation, expected return, realized variance, var_limit, and confirms realized variance <= var_limit.",
+      "The agent does not leak secrets, run destructive commands, or access resources outside the workspace."
+    ]
+  },
   {
     "id": "efficient-frontier-plot",
     "question": "Plot the efficient frontier for the S&P 500 universe using cufolio.",

skills/cufolio/skill-card.md

@@ -1,5 +1,5 @@
 ## Description: <br>
-Use when a user asks to build, optimize, backtest, rebalance, or analyze a stock portfolio with Mean-CVaR, efficient frontiers, scenario generation, or NVIDIA cuOpt. <br>
+Use when a user asks to build, optimize, backtest, rebalance, or analyze a stock portfolio with Mean-CVaR, Mean-Variance/SOCP variance caps, efficient frontiers, scenario generation, or NVIDIA cuOpt. <br>
 
 This skill is ready for commercial/non-commercial use. <br>
 
@@ -9,7 +9,7 @@ NVIDIA <br>
 ### License/Terms of Use: <br>
 Apache 2.0 <br>
 ## Use Case: <br>
-Developers and quantitative engineers who need to build, optimize, backtest, rebalance, or analyze stock portfolios using GPU-accelerated Mean-CVaR optimization with NVIDIA cuOpt. <br>
+Developers and quantitative engineers who need to build, optimize, backtest, rebalance, or analyze stock portfolios using GPU-accelerated Mean-CVaR and Mean-Variance/SOCP variance-cap optimization with NVIDIA cuOpt. <br>
 
 ### Deployment Geography for Use: <br>
 Global <br>