diff --git a/AGENTS.md b/AGENTS.md
index cd0fc91..83988aa 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -23,7 +23,7 @@ Tests are in `tests/` with shared fixtures in `tests/conftest.py`. Supporting ma
 - `uv run python scripts/smoke_harness.py --budget 1` — run CLI smoke harness.
 - `uv run python scripts/score_check.py` — score regression check.
 - `uv run python scripts/score_check.py --update` — update baselines after improvement.
-- `uv run python scripts/live_integration.py --phase green --artifact FILE --model openai/gpt-4o-mini --budget 15 --objective "..." --evaluator-command bash evaluators/eval.sh` — GREEN phase optimization.
+- `uv run python scripts/live_integration.py --phase green --artifact FILE --model openai/gpt-5.6-sol --budget 15 --objective "..." --evaluator-command bash evaluators/eval.sh` — GREEN phase optimization.
 
 ## Coding Style & Naming Conventions
 Target Python is `>=3.10`. Follow existing style:
diff --git a/CLAUDE.md b/CLAUDE.md
index e641cc1..530512a 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -10,20 +10,20 @@ uv run pytest
 uv run optimize-anything --help
 
 # Optimize
-uv run optimize-anything optimize seed.txt --evaluator-command bash eval.sh --model openai/gpt-4o-mini --objective "Improve quality"
+uv run optimize-anything optimize seed.txt --evaluator-command bash eval.sh --model openai/gpt-5.6-sol --objective "Improve quality"
 
 # Generate evaluator (default type: judge)
 uv run optimize-anything generate-evaluator seed.txt --objective "Score quality" > eval.py
 
 # Score one artifact
-uv run optimize-anything score artifact.txt --judge-model openai/gpt-4o-mini --objective "Score clarity"
+uv run optimize-anything score artifact.txt --judge-model openai/gpt-5.6-luna --objective "Score clarity"
 
 # Analyze for quality dimensions
-uv run optimize-anything analyze artifact.txt --judge-model openai/gpt-4o-mini --objective "Quality"
+uv run optimize-anything analyze artifact.txt --judge-model openai/gpt-5.6-luna --objective "Quality"
 
 # Validate across providers
 uv run optimize-anything validate artifact.txt \
-  --providers openai/gpt-4o-mini anthropic/claude-sonnet-4-5 google/gemini-2.0-flash \
+  --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
   --objective "Score quality" \
   --intake-file intake.json
 ```
@@ -70,6 +70,8 @@ uv run optimize-anything validate artifact.txt \
 Notes:
 - Exactly one evaluator source: `--evaluator-command` OR `--evaluator-url` OR `--judge-model`.
 - Early stop auto-activates when budget > 30.
+- The proposer defaults to `openai/gpt-5.6-sol` after checking `--model` and `OPTIMIZE_ANYTHING_MODEL`; generated judge and composite evaluators default to `openai/gpt-5.6-luna`.
+- LLM judge calls use each provider's sampling defaults unless an explicit temperature is supplied.
 
 ## Validate subcommand
 
@@ -77,11 +79,13 @@ Notes:
 
 ```bash
 uv run optimize-anything validate runs/run-20260303-130000/best_artifact.txt \
-  --providers openai/gpt-4o-mini anthropic/claude-sonnet-4-5 google/gemini-2.0-flash \
+  --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
   --objective "Score clarity, constraints, and robustness" \
   --intake-file intake.json
 ```
 
+Gemini uses LiteLLM's `gemini/` provider prefix.
+
 ## Plugin structure
 
 ```text
diff --git a/EXAMPLES.md b/EXAMPLES.md
index 67696ef..d803835 100644
--- a/EXAMPLES.md
+++ b/EXAMPLES.md
@@ -36,7 +36,7 @@ Optimization output examples should follow this structure:
 ```bash
 optimize-anything optimize parse_duration.py \
   --evaluator-command bash eval_tests.sh \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --objective "Pass all duration parsing tests" \
   --budget 20
 ```
@@ -61,7 +61,7 @@ Sample output excerpt:
 ```bash
 optimize-anything optimize error_template.txt \
   --evaluator-url http://localhost:8080/evaluate \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --objective "Make API error messages clear and actionable" \
   --budget 15
 ```
@@ -72,9 +72,9 @@ optimize-anything optimize error_template.txt \
 
 ```bash
 optimize-anything optimize support_prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Improve clarity, constraints, and tone" \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 30
 ```
 
@@ -84,10 +84,10 @@ optimize-anything optimize support_prompt.txt \
 
 ```bash
 optimize-anything optimize prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Generalize across user intents" \
   --dataset data/train.jsonl \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 120 \
   --workers 6 \
   --cache --run-dir runs
@@ -101,11 +101,11 @@ With validation set:
 
 ```bash
 optimize-anything optimize prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Generalize to unseen examples" \
   --dataset data/train.jsonl \
   --valset data/val.jsonl \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 150 \
   --cache --cache-from runs/run-20260303-120000 \
   --run-dir runs
@@ -117,11 +117,13 @@ optimize-anything optimize prompt.txt \
 
 ```bash
 optimize-anything validate runs/run-20260303-130000/best_artifact.txt \
-  --providers openai/gpt-4o-mini anthropic/claude-sonnet-4-5 google/gemini-2.0-flash \
+  --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
   --objective "Score clarity, correctness, and robustness" \
   --intake-file intake.json
 ```
 
+Gemini uses LiteLLM's `gemini/` provider prefix.
+
 Sample output excerpt:
 
 ```json
@@ -129,9 +131,9 @@ Sample output excerpt:
   "artifact_file": "runs/run-20260303-130000/best_artifact.txt",
   "objective": "Score clarity, correctness, and robustness",
   "providers": [
-    {"provider": "openai/gpt-4o-mini", "score": 0.82, "reasoning": "..."},
-    {"provider": "anthropic/claude-sonnet-4-5", "score": 0.79, "reasoning": "..."},
-    {"provider": "google/gemini-2.0-flash", "score": 0.81, "reasoning": "..."}
+    {"provider": "openai/gpt-5.6-luna", "score": 0.82, "reasoning": "..."},
+    {"provider": "anthropic/claude-sonnet-5", "score": 0.79, "reasoning": "..."},
+    {"provider": "gemini/gemini-3.6-flash", "score": 0.81, "reasoning": "..."}
   ],
   "mean": 0.8066666667,
   "stddev": 0.0152752523,
@@ -147,8 +149,8 @@ Sample output excerpt:
 ```bash
 optimize-anything optimize --no-seed \
   --objective "Draft a concise support policy prompt" \
-  --model openai/gpt-4o-mini \
-  --judge-model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
+  --judge-model openai/gpt-5.6-luna \
   --budget 25
 ```
 
@@ -161,7 +163,7 @@ optimize-anything optimize --no-seed \
 ```bash
 optimize-anything optimize strategy.md \
   --evaluator-command bash eval_unbounded.sh \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --objective "Maximize reward" \
   --score-range any
 ```
diff --git a/PROTOCOL.md b/PROTOCOL.md
index 143a944..a694f65 100644
--- a/PROTOCOL.md
+++ b/PROTOCOL.md
@@ -27,7 +27,7 @@ Protocol v2 extends the input payload with optional metadata and example context
 {
   "_protocol_version": 2,
   "candidate": "<text>",
-  "task_model": "openai/gpt-4o-mini",
+  "task_model": "openai/gpt-5.6-luna",
   "example": {
     "...": "dataset example object"
   }
diff --git a/README.md b/README.md
index 82d8187..537e4bc 100644
--- a/README.md
+++ b/README.md
@@ -18,9 +18,9 @@ optimize-anything generate-evaluator seed.txt \
 
 # 4) Optimize
 optimize-anything optimize seed.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Improve clarity and specificity" \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 20 \
   --workers 4 \
   --cache \
@@ -30,6 +30,10 @@ optimize-anything optimize seed.txt \
 
 CLI stdout returns a JSON summary — see [Result Contract](#result-contract) for the full shape.
 
+### Model defaults
+
+The proposer defaults to `openai/gpt-5.6-sol` after checking `--model` and `OPTIMIZE_ANYTHING_MODEL`. Generated judge and composite evaluators default to `openai/gpt-5.6-luna`. LLM judge calls use each provider's sampling defaults unless an explicit temperature is supplied.
+
 ## How It Works
 
 optimize-anything runs a GEPA (Guided Evolutionary Prompt Algorithm) loop: propose → evaluate → reflect, repeating until budget is exhausted or early stopping kicks in.
@@ -55,11 +59,11 @@ Use `--dataset` for multi-task optimization (one evaluator call per example). Ad
 
 ```bash
 optimize-anything optimize prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Generalize across customer request types" \
   --dataset data/train.jsonl \
   --valset data/val.jsonl \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 120 --workers 6 --cache --run-dir runs
 ```
 
@@ -69,11 +73,13 @@ Cross-check one artifact with multiple judge providers:
 
 ```bash
 optimize-anything validate result.txt \
-  --providers openai/gpt-4o-mini anthropic/claude-sonnet-4-5 google/gemini-2.0-flash \
+  --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
   --objective "Score clarity, constraints, and robustness" \
   --intake-file intake.json
 ```
 
+Gemini uses LiteLLM's `gemini/` provider prefix.
+
 ### Seedless mode
 
 No seed file required; GEPA bootstraps from objective.
@@ -81,8 +87,8 @@ No seed file required; GEPA bootstraps from objective.
 ```bash
 optimize-anything optimize --no-seed \
   --objective "Draft a concise, testable API prompt" \
-  --model openai/gpt-4o-mini \
-  --judge-model openai/gpt-4o-mini
+  --model openai/gpt-5.6-sol \
+  --judge-model openai/gpt-5.6-luna
 ```
 
 `--no-seed` requires both `--objective` and `--model`.
@@ -110,7 +116,7 @@ to copying the prior fitness cache.
 ```bash
 optimize-anything optimize seed.txt \
   --evaluator-command bash eval.sh \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 150 \
   --cache --cache-from runs/run-20260303-120000 \
   --run-dir runs \
@@ -126,7 +132,7 @@ For command/HTTP evaluators:
 ```bash
 optimize-anything optimize seed.txt \
   --evaluator-command bash eval.sh \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --score-range any
 ```
 
@@ -226,7 +232,7 @@ The plugin includes three skills that Claude Code can invoke automatically:
 /optimize-anything:quick prompt.txt "improve clarity and specificity"
   → runs analyze + optimize with sensible defaults, shows diff
 
-/optimize-anything:validate result.txt --providers openai/gpt-4o anthropic/claude-sonnet-4-5
+/optimize-anything:validate result.txt --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash
   → cross-checks the result with multiple judges
 ```
 
@@ -278,12 +284,12 @@ Pass it inline or from a file:
 # Inline
 optimize-anything optimize seed.txt \
   --intake-json '{"quality_dimensions": ["clarity", "specificity"], "hard_constraints": ["max 100 words"]}' \
-  --judge-model openai/gpt-4o-mini
+  --judge-model openai/gpt-5.6-luna
 
 # From file
 optimize-anything optimize seed.txt \
   --intake-file intake.json \
-  --judge-model openai/gpt-4o-mini
+  --judge-model openai/gpt-5.6-luna
 ```
 
 `optimize-anything intake` normalizes and validates these keys:
diff --git a/WALKTHROUGH.md b/WALKTHROUGH.md
index 34feb5d..395e01a 100644
--- a/WALKTHROUGH.md
+++ b/WALKTHROUGH.md
@@ -57,7 +57,7 @@ echo '{"_protocol_version":2,"candidate":"test"}' | python evaluators/eval.py
 
 ```bash
 uv run optimize-anything score seed.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Score clarity and constraints"
 ```
 
@@ -65,9 +65,9 @@ uv run optimize-anything score seed.txt \
 
 ```bash
 uv run optimize-anything optimize seed.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Improve clarity and specificity" \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 40 \
   --cache --run-dir runs --diff
 ```
@@ -80,7 +80,7 @@ Notes:
 
 ```bash
 uv run optimize-anything score runs/run-<TIMESTAMP>/best_artifact.txt \
-  --judge-model anthropic/claude-sonnet-4-5 \
+  --judge-model anthropic/claude-sonnet-5 \
   --objective "Score clarity, constraints, and usefulness"
 ```
 
@@ -88,20 +88,22 @@ uv run optimize-anything score runs/run-<TIMESTAMP>/best_artifact.txt \
 
 ```bash
 uv run optimize-anything validate runs/run-<TIMESTAMP>/best_artifact.txt \
-  --providers openai/gpt-4o-mini anthropic/claude-sonnet-4-5 google/gemini-2.0-flash \
+  --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
   --objective "Score clarity, constraints, and robustness"
 ```
 
+Gemini uses LiteLLM's `gemini/` provider prefix.
+
 ## Step 9: Iterate (dataset/generalization optional)
 
 For multi-task optimization:
 
 ```bash
 uv run optimize-anything optimize seed.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Generalize across support scenarios" \
   --dataset data/train.jsonl --valset data/val.jsonl \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 120 --cache --run-dir runs
 ```
 
@@ -114,7 +116,7 @@ evaluators, cap concurrency with workers:
 uv run optimize-anything optimize seed.txt \
   --evaluator-command bash evaluators/eval.sh \
   --objective "Improve quality" \
-  --model openai/gpt-4o-mini \
+  --model openai/gpt-5.6-sol \
   --budget 100 \
   --workers 8
 ```
diff --git a/commands/analyze.md b/commands/analyze.md
index 5d66cb6..19674ed 100644
--- a/commands/analyze.md
+++ b/commands/analyze.md
@@ -10,14 +10,14 @@ Use an LLM to discover relevant quality dimensions for a given artifact and opti
 ## Usage
 
 ```bash
-optimize-anything analyze SEED_FILE --judge-model openai/gpt-4o-mini --objective "Quality"
+optimize-anything analyze SEED_FILE --judge-model openai/gpt-5.6-luna --objective "Quality"
 ```
 
 ## Example
 
 ```bash
 optimize-anything analyze my-prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Score for clarity and persuasiveness"
 ```
 
@@ -26,7 +26,7 @@ After dimension discovery, proceed directly to optimization using the returned `
 
 ```bash
 optimize-anything optimize my-prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Score for clarity and persuasiveness" \
   --intake-json '<paste intake_json from analyze>' \
   --budget 50 --diff --run-dir runs/ --early-stop
diff --git a/commands/quick.md b/commands/quick.md
index 5cf538c..bd9ac57 100644
--- a/commands/quick.md
+++ b/commands/quick.md
@@ -9,10 +9,10 @@ Run a no-questions-asked fast optimization.
 
 ## Behavior (do not ask follow-up questions)
 1. Run analysis to discover dimensions:
-   - `optimize-anything analyze <file> --judge-model openai/gpt-4o-mini --objective "<objective>"`
+   - `optimize-anything analyze <file> --judge-model openai/gpt-5.6-luna --objective "<objective>"`
    - If analyze fails, skip dimension discovery and run optimize with `--judge-model` directly using the objective as-is.
 2. Run optimization using LLM judge with:
-   - `--judge-model openai/gpt-4o-mini`
+   - `--judge-model openai/gpt-5.6-luna`
    - `--budget 50`
    - `--diff`
    - `--early-stop`
diff --git a/commands/score.md b/commands/score.md
index d8dee91..bae0e14 100644
--- a/commands/score.md
+++ b/commands/score.md
@@ -11,7 +11,7 @@ Score a single artifact file using a command evaluator, HTTP evaluator, or LLM j
 
 ```
 optimize-anything score SEED_FILE --evaluator-command bash eval.sh
-optimize-anything score SEED_FILE --judge-model openai/gpt-4o-mini --objective "Score clarity"
+optimize-anything score SEED_FILE --judge-model openai/gpt-5.6-luna --objective "Score clarity"
 ```
 
 ## Example
@@ -21,7 +21,7 @@ optimize-anything score my-prompt.txt \
   --evaluator-command bash evaluators/clarity.sh
 
 optimize-anything score my-prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Score for persuasiveness"
 ```
 
diff --git a/commands/validate.md b/commands/validate.md
index 90812a0..d09dd37 100644
--- a/commands/validate.md
+++ b/commands/validate.md
@@ -12,10 +12,12 @@ Use multiple LLM judges to verify that a quality improvement is not provider-spe
 ## Usage
 ```bash
 optimize-anything validate <file> \
-  --providers openai/gpt-4o-mini anthropic/claude-sonnet-4-5 \
+  --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
   --objective "Score for clarity and constraint adherence"
 ```
 
+Gemini uses LiteLLM's `gemini/` provider prefix.
+
 Optional:
 - `--intake-json` or `--intake-file` for shared dimensions/constraints
 - `--api-base` for custom provider endpoints
diff --git a/evaluator-cookbook.md b/evaluator-cookbook.md
index 2a84697..8d2bdc3 100644
--- a/evaluator-cookbook.md
+++ b/evaluator-cookbook.md
@@ -15,7 +15,7 @@ Your evaluator reads a JSON object from **stdin** (command) or as a **POST body*
   "_protocol_version": 2,
   "candidate": "the text being optimized",
   "example": {"input": "...", "expected": "..."},
-  "task_model": "openai/gpt-4o-mini"
+  "task_model": "openai/gpt-5.6-luna"
 }
 ```
 
@@ -39,7 +39,7 @@ The `example` field is populated **per-call** when using `--dataset` mode. Each
 For shell-command evaluators, `task_model` is also available as an environment variable:
 
 ```bash
-echo "$OPTIMIZE_ANYTHING_TASK_MODEL"   # e.g. openai/gpt-4o-mini
+echo "$OPTIMIZE_ANYTHING_TASK_MODEL"   # e.g. openai/gpt-5.6-luna
 ```
 
 ### Output
@@ -183,14 +183,14 @@ Use an LLM judge when you want to score text qualitatively (clarity, style, pers
 **Score a single artifact:**
 ```bash
 uv run optimize-anything score my-prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Score clarity and persuasiveness"
 ```
 
 **Optimize with an LLM judge:**
 ```bash
 uv run optimize-anything optimize my-prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Maximize clarity and persuasiveness" \
   --budget 15
 ```
@@ -198,7 +198,7 @@ uv run optimize-anything optimize my-prompt.txt \
 **Discover quality dimensions, then optimize:**
 ```bash
 uv run optimize-anything analyze my-prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Quality"
 ```
 
@@ -206,7 +206,7 @@ Outcome: The `analyze` command returns named dimensions (e.g., clarity, specific
 
 ```bash
 uv run optimize-anything optimize prompt.txt \
-  --judge-model openai/gpt-4o-mini \
+  --judge-model openai/gpt-5.6-luna \
   --objective "Maximize quality" \
   --intake-json '{
     "quality_dimensions": [
@@ -254,12 +254,12 @@ def main():
         print(json.dumps({"score": 0.0, "error": "example missing input field"}))
         return 0
 
-    model = os.getenv("JUDGE_MODEL", "openai/gpt-4o-mini")
+    model = os.getenv("JUDGE_MODEL", "openai/gpt-5.6-luna")
     try:
         resp = completion(model=model, messages=[
             {"role": "system", "content": candidate},
             {"role": "user", "content": user_input},
-        ], temperature=0, max_tokens=256)
+        ], max_tokens=256)
         output = (resp.choices[0].message.content or "").lower()
     except Exception as exc:
         print(json.dumps({"score": 0.0, "error": f"llm_call_failed: {exc}"}))
@@ -337,7 +337,7 @@ def llm_judge(candidate, model):
         resp = completion(model=model, messages=[
             {"role": "system", "content": "Return JSON: {score, clarity, completeness, reasoning} all 0-1."},
             {"role": "user", "content": f"Rate clarity and completeness:\n\n{candidate}"},
-        ], temperature=0, response_format={"type": "json_object"})
+        ], response_format={"type": "json_object"})
         p = json.loads(resp.choices[0].message.content or "{}")
         return clamp01(float(p.get("score", 0))), p
     except Exception as e:
@@ -346,7 +346,7 @@ def llm_judge(candidate, model):
 def main():
     payload = json.load(sys.stdin)
     candidate = str(payload.get("candidate", ""))
-    model = os.getenv("JUDGE_MODEL", "openai/gpt-4o-mini")
+    model = os.getenv("JUDGE_MODEL", "openai/gpt-5.6-luna")
 
     h_score, h_side = heuristic_score(candidate.strip())
     if h_score < 0.2:  # short-circuit: skip LLM
@@ -462,7 +462,7 @@ else:
 ```bash
 uv run optimize-anything optimize prompt.txt \
   --evaluator-command python3 eval.py \
-  --task-model openai/gpt-4o-mini \
+  --task-model openai/gpt-5.6-luna \
   --objective "Optimize for the target model" \
   --budget 15
 ```
diff --git a/scripts/live_integration.py b/scripts/live_integration.py
index 7938822..a767156 100644
--- a/scripts/live_integration.py
+++ b/scripts/live_integration.py
@@ -5,18 +5,19 @@
 structured JSON. Designed to be driven by Claude Code as an interactive observer.
 
 Usage:
-    # GREEN phase: optimize an artifact
+    # GREEN phase: optimize with the proposer-tier default
     python scripts/live_integration.py --phase green \
         --artifact skills/generate-evaluator/SKILL.md \
-        --evaluator-command bash evaluators/skill_clarity.sh \
         --budget 15 --objective "Improve clarity" \
-        --run-dir integration_runs
+        --run-dir integration_runs \
+        --model openai/gpt-5.6-sol \
+        --evaluator-command bash evaluators/skill_clarity.sh
 
-    # RED phase: score with multiple providers
+    # RED phase: validate with the canonical judge-tier provider set
     python scripts/live_integration.py --phase red \
         --artifact skills/generate-evaluator/SKILL.md \
         --objective "Score skill quality" \
-        --providers openai/gpt-5.1 anthropic/claude-sonnet-4-5-20250929 \
+        --providers openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash \
         --evaluator-command bash evaluators/skill_clarity.sh
 """
 from __future__ import annotations
@@ -45,16 +46,16 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--run-dir", help="GREEN: directory to save run artifacts")
     parser.add_argument(
         "--model",
-        help="GREEN: LLM model string for the proposer (e.g. 'openai/gpt-4o-mini')",
+        help="GREEN proposer model (e.g. 'openai/gpt-5.6-sol')",
     )
     parser.add_argument(
         "--providers",
         nargs="+",
-        help="RED: LLM provider model strings for multi-provider scoring",
+        help="RED validation models (e.g. openai/gpt-5.6-luna anthropic/claude-sonnet-5 gemini/gemini-3.6-flash)",
     )
     parser.add_argument(
         "--judge-model",
-        help="GREEN: LLM judge model for meta-evaluator optimization",
+        help="GREEN judge model (e.g. 'openai/gpt-5.6-luna')",
     )
     parser.add_argument(
         "--judge-objective",
@@ -261,7 +262,7 @@ def _run_red(args: argparse.Namespace) -> int:
     providers = args.providers or []
     for provider in providers:
         judge_score = _score_with_judge(artifact_path, provider, args.objective)
-        # Use a clean key: "openai/gpt-5.1" -> "openai_gpt_5_1"
+        # Use a clean key: "openai/gpt-5.6-luna" -> "openai_gpt_5_6_luna"
         key = provider.replace("/", "_").replace("-", "_").replace(".", "_")
         scores[key] = judge_score
 
diff --git a/scripts/plugin_eval.sh b/scripts/plugin_eval.sh
index c93887b..09bbb45 100755
--- a/scripts/plugin_eval.sh
+++ b/scripts/plugin_eval.sh
@@ -81,17 +81,17 @@ scenario_budget() {
 
 scenario_analyze() {
   run_scenario "analyze" \
-    "Use the optimize-anything plugin to analyze the artifact at $SEED_FILE for quality dimensions. Run: optimize-anything analyze $SEED_FILE --judge-model openai/gpt-4o-mini --objective 'Score the quality of this system prompt'"
+    "Use the optimize-anything plugin to analyze the artifact at $SEED_FILE for quality dimensions. Run: optimize-anything analyze $SEED_FILE --judge-model openai/gpt-5.6-luna --objective 'Score the quality of this system prompt'"
 }
 
 scenario_validate() {
   run_scenario "validate" \
-    "Use the optimize-anything plugin to validate the artifact at $SEED_FILE across multiple providers. Run: optimize-anything validate $SEED_FILE --providers openai/gpt-4o-mini anthropic/claude-haiku-4-5-20251001 --objective 'Score the quality and clarity of this system prompt'"
+    "Use the optimize-anything plugin to validate the artifact at $SEED_FILE across multiple providers. Run: optimize-anything validate $SEED_FILE --providers openai/gpt-5.6-luna anthropic/claude-haiku-4-5-20251001 --objective 'Score the quality and clarity of this system prompt'"
 }
 
 scenario_quick() {
   run_scenario "quick" \
-    "Use the optimize-anything plugin to quickly optimize the seed at $SEED_FILE. Run: optimize-anything optimize $SEED_FILE --judge-model openai/gpt-4o-mini --objective 'Improve clarity and specificity of this system prompt' --budget 5 --model openai/gpt-4o-mini --output $RESULTS_DIR/quick-best.txt"
+    "Use the optimize-anything plugin to quickly optimize the seed at $SEED_FILE. Run: optimize-anything optimize $SEED_FILE --judge-model openai/gpt-5.6-luna --objective 'Improve clarity and specificity of this system prompt' --budget 5 --model openai/gpt-5.6-sol --output $RESULTS_DIR/quick-best.txt"
 }
 
 case "$SCENARIO" in
diff --git a/scripts/plugin_regression.py b/scripts/plugin_regression.py
index ef46a14..2192317 100644
--- a/scripts/plugin_regression.py
+++ b/scripts/plugin_regression.py
@@ -104,7 +104,7 @@ def _assert_contains(result: str, scenario: str, needles: list[str]) -> None:
 def scenario_analyze(repo_root: Path, output_dir: Path, seed_path: Path) -> dict[str, Any]:
     prompt = (
         f"Use the optimize-anything plugin to analyze the artifact at {seed_path} for quality dimensions. "
-        "Run: optimize-anything analyze runs/zo-eval/seed.txt --judge-model openai/gpt-4o-mini "
+        "Run: optimize-anything analyze runs/zo-eval/seed.txt --judge-model openai/gpt-5.6-luna "
         "--objective 'Score the quality of this system prompt'"
     )
     payload = _run_claude(repo_root, prompt, output_dir / "analyze.json", output_dir / "analyze.stderr.log")
@@ -121,7 +121,7 @@ def scenario_analyze(repo_root: Path, output_dir: Path, seed_path: Path) -> dict
 def scenario_validate(repo_root: Path, output_dir: Path, seed_path: Path) -> dict[str, Any]:
     prompt = (
         f"Use the optimize-anything plugin to validate the artifact at {seed_path} across multiple providers. "
-        "Run: optimize-anything validate runs/zo-eval/seed.txt --providers openai/gpt-4o-mini "
+        "Run: optimize-anything validate runs/zo-eval/seed.txt --providers openai/gpt-5.6-luna "
         "anthropic/claude-haiku-4-5-20251001 --objective 'Score the quality and clarity of this system prompt'"
     )
     payload = _run_claude(repo_root, prompt, output_dir / "validate.json", output_dir / "validate.stderr.log")
@@ -139,9 +139,9 @@ def scenario_quick(repo_root: Path, output_dir: Path, seed_path: Path) -> dict[s
     best_path = repo_root / "runs" / "plugin-eval" / "quick-best.txt"
     prompt = (
         f"Use the optimize-anything plugin to quickly optimize the seed at {seed_path}. "
-        f"Run: optimize-anything optimize runs/zo-eval/seed.txt --judge-model openai/gpt-4o-mini "
+        f"Run: optimize-anything optimize runs/zo-eval/seed.txt --judge-model openai/gpt-5.6-luna "
         f"--objective 'Improve clarity and specificity of this system prompt' --budget 5 "
-        f"--model openai/gpt-4o-mini --output {best_path}"
+        f"--model openai/gpt-5.6-sol --output {best_path}"
     )
     payload = _run_claude(repo_root, prompt, output_dir / "quick.json", output_dir / "quick.stderr.log")
     result = _assert_success(payload, "quick")
diff --git a/skills/evaluator-patterns/SKILL.md b/skills/evaluator-patterns/SKILL.md
index edc57eb..c1b0e58 100644
--- a/skills/evaluator-patterns/SKILL.md
+++ b/skills/evaluator-patterns/SKILL.md
@@ -82,7 +82,7 @@ Return JSON exactly like:
 }}
 """.strip()
 
-    model = os.getenv("JUDGE_MODEL", "openai/gpt-4o-mini")
+    model = os.getenv("JUDGE_MODEL", "openai/gpt-5.6-luna")
 
     try:
         resp = completion(
@@ -91,7 +91,6 @@ Return JSON exactly like:
                 {"role": "system", "content": system_prompt},
                 {"role": "user", "content": user_prompt},
             ],
-            temperature=0,
             response_format={"type": "json_object"},
         )
         content = resp.choices[0].message.content
diff --git a/skills/generate-evaluator/SKILL.md b/skills/generate-evaluator/SKILL.md
index 7ec2e1f..ecd9f10 100644
--- a/skills/generate-evaluator/SKILL.md
+++ b/skills/generate-evaluator/SKILL.md
@@ -61,7 +61,7 @@ Generate a judge evaluator and test it:
 # Generate
 optimize-anything generate-evaluator seed.txt \
   --objective "Score clarity and specificity" \
-  --model openai/gpt-4o-mini > eval_judge.py
+  --model openai/gpt-5.6-luna > eval_judge.py
 
 # Test it
 echo '{"candidate":"Your artifact text here"}' | python3 eval_judge.py
diff --git a/src/optimize_anything/cli.py b/src/optimize_anything/cli.py
index 71eb44c..905fe05 100644
--- a/src/optimize_anything/cli.py
+++ b/src/optimize_anything/cli.py
@@ -14,6 +14,7 @@
     _preflight_command_evaluator,
     _preflight_http_evaluator,
 )
+from optimize_anything.model_defaults import DEFAULT_EVALUATOR_MODEL
 
 EvaluatorFn = Callable[..., tuple[float, dict[str, Any]]]
 EvaluatorFactory = Callable[..., EvaluatorFn]
@@ -60,16 +61,16 @@ def main(argv: list[str] | None = None) -> int:
     opt_parser.add_argument(
         "--model",
         help=(
-            "LiteLLM model string for the proposer LLM "
-            "(e.g. 'openai/gpt-4o-mini', 'claude-sonnet-4-6'). "
+            "LiteLLM model string for the proposer LLM. Proposer: "
+            "openai/gpt-5.6-sol or anthropic/claude-sonnet-5. "
             "Falls back to OPTIMIZE_ANYTHING_MODEL env var."
         ),
     )
     opt_parser.add_argument(
         "--judge-model",
         help=(
-            "LiteLLM model string for built-in LLM-as-judge evaluation "
-            "(e.g. 'openai/gpt-4o-mini'). "
+            "LiteLLM model string for built-in LLM-as-judge evaluation. "
+            "Judge: openai/gpt-5.6-luna. "
             "Mutually exclusive with --evaluator-command and --evaluator-url."
         ),
     )
@@ -187,8 +188,11 @@ def main(argv: list[str] | None = None) -> int:
     )
     gen_parser.add_argument(
         "--model",
-        default="openai/gpt-4o-mini",
-        help="LiteLLM model to hardcode in generated judge/composite evaluators",
+        default=DEFAULT_EVALUATOR_MODEL,
+        help=(
+            "LiteLLM model to hardcode in generated judge/composite evaluators "
+            f"(default: {DEFAULT_EVALUATOR_MODEL})"
+        ),
     )
     gen_parser.add_argument(
         "--dataset",
@@ -254,8 +258,8 @@ def main(argv: list[str] | None = None) -> int:
     score_parser.add_argument(
         "--judge-model",
         help=(
-            "LiteLLM model string for LLM-as-judge scoring "
-            "(e.g. 'openai/gpt-5.1'). "
+            "LiteLLM model string for LLM-as-judge scoring. "
+            "Judge: openai/gpt-5.6-luna. "
             "Mutually exclusive with --evaluator-command and --evaluator-url."
         ),
     )
@@ -298,7 +302,11 @@ def main(argv: list[str] | None = None) -> int:
         "--providers",
         nargs="+",
         required=True,
-        help="Two or more LiteLLM provider model strings (e.g. openai/gpt-4o-mini anthropic/claude-sonnet-4-5)",
+        help=(
+            "Two or more LiteLLM provider model strings. Validation: "
+            "openai/gpt-5.6-luna anthropic/claude-sonnet-5 "
+            "gemini/gemini-3.6-flash"
+        ),
     )
     validate_parser.add_argument(
         "--objective",
@@ -327,7 +335,7 @@ def main(argv: list[str] | None = None) -> int:
     analyze_parser.add_argument(
         "--judge-model",
         required=True,
-        help="LiteLLM model string for the LLM judge (e.g. 'openai/gpt-4o-mini')",
+        help="LiteLLM model string for the LLM judge. Judge: openai/gpt-5.6-luna.",
     )
     analyze_parser.add_argument(
         "--objective",
@@ -341,8 +349,11 @@ def main(argv: list[str] | None = None) -> int:
     analyze_parser.add_argument(
         "--temperature",
         type=float,
-        default=0.0,
-        help="Sampling temperature for LLM calls (default: 0.0)",
+        default=None,
+        help=(
+            "Sampling temperature for LLM calls "
+            "(default: provider model default)"
+        ),
     )
     analyze_parser.add_argument(
         "--timeout",
diff --git a/src/optimize_anything/cli_optimize.py b/src/optimize_anything/cli_optimize.py
index 5860d94..05140cc 100644
--- a/src/optimize_anything/cli_optimize.py
+++ b/src/optimize_anything/cli_optimize.py
@@ -5,7 +5,6 @@
 import argparse
 import copy
 import json
-import os
 import sys
 from pathlib import Path
 from typing import Any
@@ -17,6 +16,7 @@
     _save_run_dir,
     _timestamped_run_dir,
 )
+from optimize_anything.model_defaults import resolve_proposer_model
 
 OptimizeInputs = tuple[
     argparse.Namespace,
@@ -67,7 +67,7 @@ def _cmd_optimize(args: argparse.Namespace) -> int:
         file=sys.stderr,
     )
 
-    model = args.model or os.environ.get("OPTIMIZE_ANYTHING_MODEL")
+    model = resolve_proposer_model(args.model)
     config, gepa_run_dir, early_stop_active, runtime_error = _build_optimize_runtime(
         args,
         model=model,
@@ -134,7 +134,7 @@ def _cmd_optimize(args: argparse.Namespace) -> int:
 def _build_optimize_runtime(
     args: argparse.Namespace,
     *,
-    model: str | None,
+    model: str,
 ) -> tuple[Any, str | None, bool, str | None]:
     """Build GEPA runtime config plus run-dir state for optimize."""
     from optimize_anything.stop import plateau_stop_callback
@@ -184,14 +184,11 @@ def _build_optimize_runtime(
             n=args.proposals_per_iteration
         )
     engine = EngineConfig(**engine_kwargs)
-    if model:
-        config = GEPAConfig(
-            engine=engine,
-            reflection=ReflectionConfig(reflection_lm=model),
-            stop_callbacks=stop_callbacks,
-        )
-    else:
-        config = GEPAConfig(engine=engine, stop_callbacks=stop_callbacks)
+    config = GEPAConfig(
+        engine=engine,
+        reflection=ReflectionConfig(reflection_lm=model),
+        stop_callbacks=stop_callbacks,
+    )
 
     return config, gepa_run_dir, early_stop_active, None
 
diff --git a/src/optimize_anything/evaluator_generator.py b/src/optimize_anything/evaluator_generator.py
index 8aa5464..ce5207b 100644
--- a/src/optimize_anything/evaluator_generator.py
+++ b/src/optimize_anything/evaluator_generator.py
@@ -6,6 +6,8 @@
 from numbers import Real
 from typing import Any
 
+from optimize_anything.model_defaults import DEFAULT_EVALUATOR_MODEL
+
 
 def generate_evaluator_script(
     *,
@@ -13,7 +15,7 @@ def generate_evaluator_script(
     objective: str,
     evaluator_type: str | None = None,
     intake: Mapping[str, Any] | None = None,
-    model: str = "openai/gpt-4o-mini",
+    model: str = DEFAULT_EVALUATOR_MODEL,
     dataset: bool = False,
 ) -> str:
     """Generate an evaluator script that reads input JSON and outputs score JSON."""
@@ -414,7 +416,7 @@ def _generate_judge_evaluator(
     template_family: str,
     rubric_summary: str,
     quality_dimensions: list[tuple[str, float]],
-    model: str = "openai/gpt-4o-mini",
+    model: str,
     dataset: bool = False,
 ) -> str:
     """Generate a Python LLM-judge evaluator script using litellm."""
@@ -437,7 +439,7 @@ def _generate_judge_evaluator(
         def _build_prompt(candidate: str, example: object | None) -> str:
             dimensions_text = "\\n".join([f"- {{name}} (weight={{weight}})" for name, weight in QUALITY_DIMENSIONS])
             example_text = json.dumps(example, ensure_ascii=False, indent=2) if example is not None else "(none)"
-            return f\"\"\"## Objective\n{{OBJECTIVE}}\n\n## Template Family\n{{TEMPLATE_FAMILY}}\n\n## Rubric Summary\n{{RUBRIC_SUMMARY}}\n\n## Quality Dimensions\n{{dimensions_text}}\n\n## Example Context (optional)\n{{example_text}}\n\n## Artifact to Evaluate\n```\n{{candidate}}\n```\n\nReturn JSON with keys: score, reasoning, and one key per quality dimension name. score must be in [0,1].\"\"\"
+            return f\"\"\"## Objective\\n{{OBJECTIVE}}\\n\\n## Template Family\\n{{TEMPLATE_FAMILY}}\\n\\n## Rubric Summary\\n{{RUBRIC_SUMMARY}}\\n\\n## Quality Dimensions\\n{{dimensions_text}}\\n\\n## Example Context (optional)\\n{{example_text}}\\n\\n## Artifact to Evaluate\\n```\\n{{candidate}}\\n```\\n\\nReturn JSON with keys: score, reasoning, and one key per quality dimension name. score must be in [0,1].\"\"\"
 
         def _api_key_available() -> bool:
             key_vars = ["OPENAI_API_KEY", "ANTHROPIC_API_KEY", "GEMINI_API_KEY", "OPENROUTER_API_KEY"]
@@ -446,7 +448,7 @@ def _api_key_available() -> bool:
         def _strip_code_fences(text: str) -> str:
             cleaned = text.strip()
             if cleaned.startswith("```"):
-                first_newline = cleaned.index("\n") if "\n" in cleaned else len(cleaned)
+                first_newline = cleaned.index("\\n") if "\\n" in cleaned else len(cleaned)
                 cleaned = cleaned[first_newline + 1:]
                 if cleaned.rstrip().endswith("```"):
                     cleaned = cleaned.rstrip()[:-len("```")].rstrip()
@@ -478,7 +480,6 @@ def main() -> int:
                         {{"role": "system", "content": JUDGE_SYSTEM_PROMPT}},
                         {{"role": "user", "content": prompt}},
                     ],
-                    temperature=0.0,
                     timeout=60.0,
                     response_format={{"type": "json_object"}},
                 )
@@ -524,7 +525,7 @@ def _generate_composite_evaluator(
     template_family: str,
     rubric_summary: str,
     quality_dimensions: list[tuple[str, float]],
-    model: str = "openai/gpt-4o-mini",
+    model: str,
     dataset: bool = False,
 ) -> str:
     """Generate composite evaluator with hard constraints + judge scoring."""
@@ -544,6 +545,8 @@ def _generate_composite_evaluator(
         import sys
 
         # Composite evaluator: hard constraints first, then LLM judge.
+        MODEL = {model!r}
+
         def _constraint_non_empty(candidate: str) -> tuple[bool, str]:
             if candidate.strip():
                 return True, ""
@@ -564,7 +567,7 @@ def _constraint_no_placeholder(candidate: str) -> tuple[bool, str]:
         def _strip_code_fences(text: str) -> str:
             cleaned = text.strip()
             if cleaned.startswith("```"):
-                first_newline = cleaned.index("\n") if "\n" in cleaned else len(cleaned)
+                first_newline = cleaned.index("\\n") if "\\n" in cleaned else len(cleaned)
                 cleaned = cleaned[first_newline + 1:]
                 if cleaned.rstrip().endswith("```"):
                     cleaned = cleaned.rstrip()[:-len("```")].rstrip()
diff --git a/src/optimize_anything/llm_judge.py b/src/optimize_anything/llm_judge.py
index 99d72d6..23b4168 100644
--- a/src/optimize_anything/llm_judge.py
+++ b/src/optimize_anything/llm_judge.py
@@ -85,7 +85,7 @@ def llm_judge_evaluator(
     quality_dimensions: list[dict[str, Any]] | None = None,
     hard_constraints: list[str] | None = None,
     timeout: float = 60.0,
-    temperature: float = 0.0,
+    temperature: float | None = None,
     api_base: str | None = None,
     task_model: str | None = None,
 ) -> Callable[[str, Any | None], tuple[float, dict[str, Any]]]:
@@ -113,10 +113,11 @@ def evaluate(candidate: str, example: Any | None = None) -> tuple[float, dict[st
                     {"role": "system", "content": JUDGE_SYSTEM_PROMPT},
                     {"role": "user", "content": prompt},
                 ],
-                "temperature": temperature,
                 "timeout": timeout,
                 "response_format": {"type": "json_object"},
             }
+            if temperature is not None:
+                completion_kwargs["temperature"] = temperature
             if api_base:
                 completion_kwargs["base_url"] = api_base
 
@@ -321,7 +322,7 @@ def analyze_for_dimensions(
     *,
     api_base: str | None = None,
     timeout: float = 60.0,
-    temperature: float = 0.0,
+    temperature: float | None = None,
 ) -> dict[str, Any]:
     """Score an artifact then discover quality dimensions for refinement.
 
@@ -346,10 +347,11 @@ def analyze_for_dimensions(
             {"role": "system", "content": JUDGE_SYSTEM_PROMPT},
             {"role": "user", "content": score_prompt},
         ],
-        "temperature": temperature,
         "timeout": timeout,
         "response_format": {"type": "json_object"},
     }
+    if temperature is not None:
+        completion_kwargs["temperature"] = temperature
     if api_base:
         completion_kwargs["base_url"] = api_base
 
@@ -375,10 +377,11 @@ def analyze_for_dimensions(
             {"role": "system", "content": ANALYZE_SYSTEM_PROMPT},
             {"role": "user", "content": analyze_prompt},
         ],
-        "temperature": temperature,
         "timeout": timeout,
         "response_format": {"type": "json_object"},
     }
+    if temperature is not None:
+        analyze_kwargs["temperature"] = temperature
     if api_base:
         analyze_kwargs["base_url"] = api_base
 
diff --git a/src/optimize_anything/model_defaults.py b/src/optimize_anything/model_defaults.py
new file mode 100644
index 0000000..d9cffae
--- /dev/null
+++ b/src/optimize_anything/model_defaults.py
@@ -0,0 +1,23 @@
+"""Project-owned defaults for LLM roles."""
+
+from __future__ import annotations
+
+import os
+from collections.abc import Mapping
+from typing import Final
+
+DEFAULT_PROPOSER_MODEL: Final[str] = "openai/gpt-5.6-sol"
+DEFAULT_EVALUATOR_MODEL: Final[str] = "openai/gpt-5.6-luna"
+
+
+def resolve_proposer_model(
+    explicit_model: str | None,
+    environ: Mapping[str, str] | None = None,
+) -> str:
+    """Resolve CLI, environment, then project default proposer precedence."""
+    environment = os.environ if environ is None else environ
+    return (
+        explicit_model
+        or environment.get("OPTIMIZE_ANYTHING_MODEL")
+        or DEFAULT_PROPOSER_MODEL
+    )
diff --git a/tests/test_cli.py b/tests/test_cli.py
index 9e0eb93..ec8c6ca 100644
--- a/tests/test_cli.py
+++ b/tests/test_cli.py
@@ -23,6 +23,16 @@ def test_optimize_help(self, capsys):
         assert "--intake-json" in captured.out
         assert "--intake-file" in captured.out
         assert "--evaluator-cwd" in captured.out
+        assert "openai/gpt-5.6-sol" in captured.out
+
+    def test_generate_evaluator_help(self, capsys):
+        """Verify generated evaluator help recommends the current judge model."""
+        try:
+            main(["generate-evaluator", "--help"])
+        except SystemExit as e:
+            assert e.code == 0
+        captured = capsys.readouterr()
+        assert "openai/gpt-5.6-luna" in captured.out
 
     def test_explain(self, tmp_path: Path, capsys):
         seed_file = tmp_path / "seed.txt"
@@ -856,6 +866,44 @@ def fake_optimize(**kwargs):
         cfg = captured_config["config"]
         assert cfg.reflection.reflection_lm == "gemini/gemini-2.0-flash"
 
+    def test_optimize_uses_default_proposer_model(
+        self, tmp_path: Path, capsys, monkeypatch
+    ):
+        seed_file = tmp_path / "seed.txt"
+        seed_file.write_text("test")
+        captured_config = {}
+
+        class DummyResult:
+            best_candidate = "x"
+            total_metric_calls = 1
+
+        def fake_optimize(**kwargs):
+            captured_config["config"] = kwargs.get("config")
+            return DummyResult()
+
+        monkeypatch.setattr(
+            "optimize_anything.evaluators.command_evaluator",
+            lambda command, cwd=None, **kwargs: lambda c: (0.5, {}),
+        )
+        monkeypatch.setattr(
+            "gepa.optimize_anything.optimize_anything",
+            fake_optimize,
+        )
+        monkeypatch.setattr(
+            "optimize_anything.cli._preflight_command_evaluator",
+            lambda command, cwd=None: None,
+        )
+        monkeypatch.delenv("OPTIMIZE_ANYTHING_MODEL", raising=False)
+
+        result = main([
+            "optimize", str(seed_file),
+            "--evaluator-command", "bash", "eval.sh",
+            "--budget", "1",
+        ])
+        assert result == 0
+        cfg = captured_config["config"]
+        assert cfg.reflection.reflection_lm == "openai/gpt-5.6-sol"
+
     def test_optimize_prints_progress_to_stderr(
         self, tmp_path: Path, capsys, monkeypatch
     ):
diff --git a/tests/test_evaluator_generator.py b/tests/test_evaluator_generator.py
index 82c1687..3324684 100644
--- a/tests/test_evaluator_generator.py
+++ b/tests/test_evaluator_generator.py
@@ -1,11 +1,60 @@
 """Tests for evaluator generator."""
 import json
 import subprocess
+import sys
 from pathlib import Path
 
+import pytest
+
 from optimize_anything.evaluator_generator import generate_evaluator_script
 
 
+@pytest.mark.parametrize("evaluator_type", ["judge", "composite"])
+def test_generated_python_evaluator_compiles(evaluator_type: str) -> None:
+    script = generate_evaluator_script(
+        seed="hello",
+        objective="score quality",
+        evaluator_type=evaluator_type,
+    )
+
+    compile(script, f"<generated-{evaluator_type}-evaluator>", "exec")
+
+
+@pytest.mark.parametrize("evaluator_type", ["judge", "composite"])
+def test_generated_python_evaluator_reports_missing_key_offline(
+    evaluator_type: str,
+    tmp_path: Path,
+) -> None:
+    script = generate_evaluator_script(
+        seed="hello",
+        objective="score quality",
+        evaluator_type=evaluator_type,
+    )
+    script_path = tmp_path / f"{evaluator_type}_evaluator.py"
+    script_path.write_text(script, encoding="utf-8")
+
+    proc = subprocess.run(
+        [sys.executable, str(script_path)],
+        input=json.dumps({"candidate": "hello"}),
+        capture_output=True,
+        text=True,
+        check=True,
+        env={},
+    )
+
+    result = json.loads(proc.stdout)
+    assert result["error"] == "missing_api_key"
+
+
+def test_generated_judge_uses_provider_sampling_defaults() -> None:
+    script = generate_evaluator_script(
+        seed="hello",
+        objective="score quality",
+        evaluator_type="judge",
+    )
+    assert "temperature=" not in script
+
+
 class TestGenerateEvaluatorScript:
     def test_command_evaluator_is_bash(self):
         script = generate_evaluator_script(seed="hello", objective="improve clarity", evaluator_type="command")
@@ -39,6 +88,22 @@ def test_judge_evaluator_contains_litellm_and_objective(self):
         assert "from litellm import completion" in script
         assert objective in script
 
+    def test_default_judge_uses_current_evaluator_model(self) -> None:
+        script = generate_evaluator_script(
+            seed="hello",
+            objective="score quality",
+            evaluator_type="judge",
+        )
+        assert "MODEL = 'openai/gpt-5.6-luna'" in script
+
+    def test_default_composite_uses_current_evaluator_model(self) -> None:
+        script = generate_evaluator_script(
+            seed="hello",
+            objective="score quality",
+            evaluator_type="composite",
+        )
+        assert "MODEL = 'openai/gpt-5.6-luna'" in script
+
     def test_judge_evaluator_handles_missing_api_key_gracefully(self):
         script = generate_evaluator_script(seed="hello", objective="test", evaluator_type="judge")
         assert "Missing API key" in script
diff --git a/tests/test_llm_judge.py b/tests/test_llm_judge.py
index a4e569b..1e45605 100644
--- a/tests/test_llm_judge.py
+++ b/tests/test_llm_judge.py
@@ -164,6 +164,24 @@ def test_temperature_passed_through_to_litellm(self):
         call_kwargs = mock_call.call_args.kwargs
         assert call_kwargs["temperature"] == 0.7
 
+    def test_default_temperature_is_omitted_from_litellm(self):
+        captured: dict[str, object] = {}
+
+        def fake_completion(**kwargs):
+            captured.update(kwargs)
+            return self._make_mock_completion(
+                '{"score": 0.8, "reasoning": "Good"}'
+            )
+
+        with patch("litellm.completion", side_effect=fake_completion):
+            evaluate = llm_judge_evaluator(
+                "Score quality",
+                model="anthropic/claude-sonnet-5",
+            )
+            evaluate("candidate")
+
+        assert "temperature" not in captured
+
     def test_model_string_passed_through(self):
         response_content = json.dumps({"score": 0.5, "reasoning": "ok"})
         evaluator = llm_judge_evaluator(
@@ -582,9 +600,12 @@ def test_makes_exactly_two_llm_calls(self):
             ]
         })
 
+        captured_calls: list[dict[str, object]] = []
         call_count = 0
+
         def mock_completion(**kwargs):
             nonlocal call_count
+            captured_calls.append(kwargs)
             call_count += 1
             if call_count == 1:
                 return self._make_mock_response(score_response)
@@ -594,6 +615,31 @@ def mock_completion(**kwargs):
             analyze_for_dimensions("text", "obj", "openai/gpt-4o-mini")
 
         assert call_count == 2
+        assert all("temperature" not in kwargs for kwargs in captured_calls)
+
+    def test_explicit_temperature_is_forwarded_to_both_llm_calls(self):
+        score_response = json.dumps({"score": 0.5, "reasoning": "ok"})
+        dims_response = json.dumps({
+            "dimensions": [
+                {"name": "a", "weight": 1.0, "score": 0.5, "description": "d"},
+            ]
+        })
+        responses = iter((score_response, dims_response))
+        captured_calls: list[dict[str, object]] = []
+
+        def mock_completion(**kwargs):
+            captured_calls.append(kwargs)
+            return self._make_mock_response(next(responses))
+
+        with patch("litellm.completion", side_effect=mock_completion):
+            analyze_for_dimensions(
+                "text",
+                "obj",
+                "openai/gpt-4o-mini",
+                temperature=0.7,
+            )
+
+        assert [call["temperature"] for call in captured_calls] == [0.7, 0.7]
 
     def test_scoring_failure_raises_runtime_error(self):
         with patch("litellm.completion", side_effect=RuntimeError("API down")):
diff --git a/tests/test_model_defaults.py b/tests/test_model_defaults.py
new file mode 100644
index 0000000..148cd1e
--- /dev/null
+++ b/tests/test_model_defaults.py
@@ -0,0 +1,115 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from optimize_anything.model_defaults import (
+    DEFAULT_EVALUATOR_MODEL,
+    DEFAULT_PROPOSER_MODEL,
+    resolve_proposer_model,
+)
+
+
+ROOT = Path(__file__).resolve().parents[1]
+EXECUTABLE_MODEL_REFERENCE_FILES = (
+    "src/optimize_anything/cli.py",
+    "src/optimize_anything/evaluator_generator.py",
+    "scripts/live_integration.py",
+    "scripts/plugin_eval.sh",
+    "scripts/plugin_regression.py",
+    "skills/evaluator-patterns/SKILL.md",
+    "skills/generate-evaluator/SKILL.md",
+)
+DOCUMENTATION_MODEL_REFERENCE_FILES = (
+    "AGENTS.md",
+    "CLAUDE.md",
+    "EXAMPLES.md",
+    "PROTOCOL.md",
+    "README.md",
+    "WALKTHROUGH.md",
+    "evaluator-cookbook.md",
+    "commands/analyze.md",
+    "commands/quick.md",
+    "commands/score.md",
+    "commands/validate.md",
+)
+STALE_CANONICAL_MODEL_STRINGS = (
+    "openai/gpt-4o",
+    "openai/gpt-4o-mini",
+    "openai/gpt-5.1",
+    "claude-sonnet-4-6",
+    "anthropic/claude-sonnet-4-5",
+    "anthropic/claude-sonnet-4-5-20250929",
+    "anthropic/claude-sonnet-4-6",
+    "google/gemini-2.0-flash",
+    "gemini/gemini-2.0-flash",
+)
+
+
+@pytest.mark.parametrize(
+    "relative_path",
+    EXECUTABLE_MODEL_REFERENCE_FILES + DOCUMENTATION_MODEL_REFERENCE_FILES,
+)
+def test_user_facing_surfaces_do_not_recommend_stale_models(
+    relative_path: str,
+) -> None:
+    text = (ROOT / relative_path).read_text(encoding="utf-8")
+    assert not [
+        model for model in STALE_CANONICAL_MODEL_STRINGS if model in text
+    ]
+
+
+def test_primary_docs_show_current_validation_models() -> None:
+    for relative_path in ("README.md", "EXAMPLES.md", "WALKTHROUGH.md"):
+        text = (ROOT / relative_path).read_text(encoding="utf-8")
+        assert "openai/gpt-5.6-luna" in text
+        assert "anthropic/claude-sonnet-5" in text
+        assert "gemini/gemini-3.6-flash" in text
+
+
+def test_default_models_match_current_capability_tiers() -> None:
+    assert DEFAULT_PROPOSER_MODEL == "openai/gpt-5.6-sol"
+    assert DEFAULT_EVALUATOR_MODEL == "openai/gpt-5.6-luna"
+
+
+@pytest.mark.parametrize(
+    ("explicit_model", "environ", "expected"),
+    [
+        (
+            "anthropic/claude-sonnet-5",
+            {"OPTIMIZE_ANYTHING_MODEL": "gemini/gemini-3.6-flash"},
+            "anthropic/claude-sonnet-5",
+        ),
+        (
+            None,
+            {"OPTIMIZE_ANYTHING_MODEL": "gemini/gemini-3.6-flash"},
+            "gemini/gemini-3.6-flash",
+        ),
+        (None, {}, "openai/gpt-5.6-sol"),
+    ],
+)
+def test_resolve_proposer_model_precedence(
+    explicit_model: str | None,
+    environ: dict[str, str],
+    expected: str,
+) -> None:
+    assert resolve_proposer_model(explicit_model, environ) == expected
+
+
+@pytest.mark.parametrize(
+    ("model", "provider"),
+    [
+        ("openai/gpt-5.6-sol", "openai"),
+        ("openai/gpt-5.6-luna", "openai"),
+        ("anthropic/claude-sonnet-5", "anthropic"),
+        ("gemini/gemini-3.6-flash", "gemini"),
+    ],
+)
+def test_recommended_models_route_through_pinned_litellm(
+    model: str,
+    provider: str,
+) -> None:
+    import litellm
+
+    assert litellm.get_llm_provider(model)[1] == provider