diff --git a/configs/endpoints.toml b/configs/endpoints.toml
deleted file mode 100644
index 10f9f5d479..0000000000
--- a/configs/endpoints.toml
+++ /dev/null
@@ -1,307 +0,0 @@
-[[endpoint]]
-endpoint_id = "olmo3-32b-t"
-model = "allenai/olmo-3-32b-think"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "olmo3-7b-i"
-model = "allenai/olmo-3-7b-instruct"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "olmo3-7b-t"
-model = "allenai/olmo-3-7b-think"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "trinity-mini"
-model = "arcee/trinity-mini"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "haiku"
-model = "claude-haiku-4-5"
-url = "https://api.anthropic.com"
-key = "ANTHROPIC_API_KEY"
-type = "anthropic_messages"
-
-[[endpoint]]
-endpoint_id = "sonnet"
-model = "claude-sonnet-4-5"
-url = "https://api.anthropic.com"
-key = "ANTHROPIC_API_KEY"
-type = "anthropic_messages"
-
-[[endpoint]]
-endpoint_id = "opus"
-model = "claude-opus-4-5"
-url = "https://api.anthropic.com"
-key = "ANTHROPIC_API_KEY"
-type = "anthropic_messages"
-
-[[endpoint]]
-endpoint_id = "deepseek-chat"
-model = "deepseek-chat"
-url = "https://api.deepseek.com/v1"
-key = "DEEPSEEK_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "deepseek-reasoner"
-model = "deepseek-reasoner"
-url = "https://api.deepseek.com/v1"
-key = "DEEPSEEK_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "deepseek-chat-anth"
-model = "deepseek-chat"
-url = "https://api.deepseek.com/anthropic"
-key = "DEEPSEEK_API_KEY"
-type = "anthropic_messages"
-
-[[endpoint]]
-endpoint_id = "deepseek-reasoner-anth"
-model = "deepseek-reasoner"
-url = "https://api.deepseek.com/anthropic"
-key = "DEEPSEEK_API_KEY"
-type = "anthropic_messages"
-
-[[endpoint]]
-endpoint_id = "gemini-2.5-flash"
-model = "google/gemini-2.5-flash"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gemini-2.5-pro"
-model = "google/gemini-2.5-pro"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gemini-3-flash"
-model = "google/gemini-3-flash"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gemini-3-pro"
-model = "google/gemini-3-pro-preview"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gemini-3-pro-exp"
-model = "google/gemini-3-pro-preview"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-30b-i"
-model = "qwen/qwen3-30b-a3b-instruct-2507"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-30b-t"
-model = "qwen/qwen3-30b-a3b-thinking-2507"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-235b-i"
-model = "qwen/qwen3-235b-a22b-instruct-2507"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-235b-t"
-model = "qwen/qwen3-235b-a22b-thinking-2507"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-vl-30b-i"
-model = "qwen/qwen3-vl-30b-a3b-instruct"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-vl-30b-t"
-model = "qwen/qwen3-vl-30b-a3b-thinking"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-vl-235b-i"
-model = "qwen/qwen3-vl-235b-a22b-instruct"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "qwen3-vl-235b-t"
-model = "qwen/qwen3-vl-235b-a22b-thinking"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "kimi-k2"
-model = "moonshotai/kimi-k2-0905"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "kimi-k2-t"
-model = "moonshotai/kimi-k2-thinking"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-oss-120b"
-model = "openai/gpt-oss-120b"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-oss-20b"
-model = "openai/gpt-oss-20b"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5.4-nano"
-model = "openai/gpt-5.4-nano"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5.4-mini"
-model = "openai/gpt-5.4-mini"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5.4"
-model = "openai/gpt-5.4"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5.5"
-model = "openai/gpt-5.5"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-4.1-nano"
-model = "gpt-4.1-nano"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-4.1-mini"
-model = "gpt-4.1-mini"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-4.1"
-model = "gpt-4.1"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5-nano"
-model = "gpt-5-nano"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5-mini"
-model = "gpt-5-mini"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5"
-model = "gpt-5"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5.1"
-model = "gpt-5.1"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "gpt-5.2"
-model = "gpt-5.2"
-url = "https://api.openai.com/v1"
-key = "OPENAI_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "glm-4.5"
-model = "z-ai/glm-4.5"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "glm-4.5-air"
-model = "z-ai/glm-4.5-air"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "glm-4.6"
-model = "z-ai/glm-4.6"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
-
-[[endpoint]]
-endpoint_id = "glm-4.7"
-model = "z-ai/glm-4.7"
-url = "https://api.pinference.ai/api/v1"
-key = "PRIME_API_KEY"
-type = "openai_chat_completions"
diff --git a/configs/eval/gpt-oss.toml b/configs/eval/gpt-oss.toml
deleted file mode 100644
index cb7d40f8a1..0000000000
--- a/configs/eval/gpt-oss.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# gpt-oss models. Uncomment exactly one model.
-model = "openai/gpt-oss-20b"
-# model = "openai/gpt-oss-120b"
-
-num_examples = 20
-rollouts_per_example = 1
-max_tokens = 1024
-
-[[eval]]
-env_id = "primeintellect/reverse-text"
-
-# [[eval]]
-# env_id = "primeintellect/wordle"
-
-# [[eval]]
-# env_id = "primeintellect/wiki-search"
-
-# [[eval]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[eval]]
-# env_id = "primeintellect/opencode-harbor"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/eval/llama-3.toml b/configs/eval/llama-3.toml
deleted file mode 100644
index 02ebaa77e5..0000000000
--- a/configs/eval/llama-3.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# Llama 3.2 models. Uncomment exactly one model.
-model = "meta-llama/Llama-3.2-1B-Instruct"
-# model = "meta-llama/Llama-3.2-3B-Instruct"
-
-num_examples = 20
-rollouts_per_example = 1
-max_tokens = 1024
-
-[[eval]]
-env_id = "primeintellect/reverse-text"
-
-# [[eval]]
-# env_id = "primeintellect/wordle"
-
-# [[eval]]
-# env_id = "primeintellect/wiki-search"
-
-# [[eval]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[eval]]
-# env_id = "primeintellect/opencode-harbor"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/eval/nemotron-3.toml b/configs/eval/nemotron-3.toml
deleted file mode 100644
index 51d3e300cd..0000000000
--- a/configs/eval/nemotron-3.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# Nemotron 3 models. Uncomment exactly one model.
-model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
-# model = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
-
-num_examples = 20
-rollouts_per_example = 1
-max_tokens = 2048
-
-[[eval]]
-env_id = "primeintellect/wiki-search"
-
-# [[eval]]
-# env_id = "primeintellect/reverse-text"
-
-# [[eval]]
-# env_id = "primeintellect/wordle"
-
-# [[eval]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[eval]]
-# env_id = "primeintellect/opencode-harbor"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/eval/qwen-3-5-moe.toml b/configs/eval/qwen-3-5-moe.toml
deleted file mode 100644
index b7bf74012a..0000000000
--- a/configs/eval/qwen-3-5-moe.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# Qwen3.5 MoE models. Uncomment exactly one model.
-model = "Qwen/Qwen3.5-35B-A3B"
-# model = "Qwen/Qwen3.5-122B-A10B"
-# model = "Qwen/Qwen3.5-397B-A17B"
-
-num_examples = 20
-rollouts_per_example = 1
-max_tokens = 2048
-
-[[eval]]
-env_id = "primeintellect/wiki-search"
-
-# [[eval]]
-# env_id = "primeintellect/reverse-text"
-
-# [[eval]]
-# env_id = "primeintellect/wordle"
-
-# [[eval]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[eval]]
-# env_id = "primeintellect/opencode-harbor"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/eval/qwen-3-5.toml b/configs/eval/qwen-3-5.toml
deleted file mode 100644
index 4793861c00..0000000000
--- a/configs/eval/qwen-3-5.toml
+++ /dev/null
@@ -1,26 +0,0 @@
-# Qwen3.5 dense models. Uncomment exactly one model.
-# model = "Qwen/Qwen3.5-0.8B"
-# model = "Qwen/Qwen3.5-2B"
-model = "Qwen/Qwen3.5-4B"
-# model = "Qwen/Qwen3.5-9B"
-
-num_examples = 20
-rollouts_per_example = 1
-max_tokens = 1024
-
-[[eval]]
-env_id = "primeintellect/reverse-text"
-
-# [[eval]]
-# env_id = "primeintellect/wordle"
-
-# [[eval]]
-# env_id = "primeintellect/wiki-search"
-
-# [[eval]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[eval]]
-# env_id = "primeintellect/opencode-harbor"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/gepa/gpt-oss.toml b/configs/gepa/gpt-oss.toml
deleted file mode 100644
index 9ecece4d89..0000000000
--- a/configs/gepa/gpt-oss.toml
+++ /dev/null
@@ -1,33 +0,0 @@
-# gpt-oss models. Update both fields when switching sizes.
-model = "openai/gpt-oss-20b"
-reflection_model = "openai/gpt-oss-20b"
-# model = "openai/gpt-oss-120b"
-# reflection_model = "openai/gpt-oss-120b"
-
-[[env]]
-env_id = "primeintellect/reverse-text"
-
-# [[env]]
-# env_id = "primeintellect/wordle"
-
-# [[env]]
-# env_id = "primeintellect/wiki-search"
-
-# [[env]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[env]]
-# env_id = "primeintellect/opencode-harbor"
-
-[gepa]
-max_calls = 500
-num_train = 100
-num_val = 50
-minibatch_size = 3
-max_concurrent = 32
-
-[sampling]
-max_tokens = 1024
-reasoning_effort = "low"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/gepa/llama-3.toml b/configs/gepa/llama-3.toml
deleted file mode 100644
index 024b72636f..0000000000
--- a/configs/gepa/llama-3.toml
+++ /dev/null
@@ -1,32 +0,0 @@
-# Llama 3.2 models. Update both fields when switching sizes.
-model = "meta-llama/Llama-3.2-1B-Instruct"
-reflection_model = "meta-llama/Llama-3.2-1B-Instruct"
-# model = "meta-llama/Llama-3.2-3B-Instruct"
-# reflection_model = "meta-llama/Llama-3.2-3B-Instruct"
-
-[[env]]
-env_id = "primeintellect/reverse-text"
-
-# [[env]]
-# env_id = "primeintellect/wordle"
-
-# [[env]]
-# env_id = "primeintellect/wiki-search"
-
-# [[env]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[env]]
-# env_id = "primeintellect/opencode-harbor"
-
-[gepa]
-max_calls = 500
-num_train = 100
-num_val = 50
-minibatch_size = 3
-max_concurrent = 32
-
-[sampling]
-max_tokens = 1024
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/gepa/nemotron-3.toml b/configs/gepa/nemotron-3.toml
deleted file mode 100644
index 8bc49f45e9..0000000000
--- a/configs/gepa/nemotron-3.toml
+++ /dev/null
@@ -1,32 +0,0 @@
-# Nemotron 3 models. Update both fields when switching sizes.
-model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
-reflection_model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
-# model = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
-# reflection_model = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
-
-[[env]]
-env_id = "primeintellect/wiki-search"
-
-# [[env]]
-# env_id = "primeintellect/reverse-text"
-
-# [[env]]
-# env_id = "primeintellect/wordle"
-
-# [[env]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[env]]
-# env_id = "primeintellect/opencode-harbor"
-
-[gepa]
-max_calls = 500
-num_train = 100
-num_val = 50
-minibatch_size = 3
-max_concurrent = 32
-
-[sampling]
-max_tokens = 2048
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/gepa/qwen-3-5-moe.toml b/configs/gepa/qwen-3-5-moe.toml
deleted file mode 100644
index 9bb3d3162b..0000000000
--- a/configs/gepa/qwen-3-5-moe.toml
+++ /dev/null
@@ -1,34 +0,0 @@
-# Qwen3.5 MoE models. Update both fields when switching sizes.
-model = "Qwen/Qwen3.5-35B-A3B"
-reflection_model = "Qwen/Qwen3.5-35B-A3B"
-# model = "Qwen/Qwen3.5-122B-A10B"
-# reflection_model = "Qwen/Qwen3.5-122B-A10B"
-# model = "Qwen/Qwen3.5-397B-A17B"
-# reflection_model = "Qwen/Qwen3.5-397B-A17B"
-
-[[env]]
-env_id = "primeintellect/wiki-search"
-
-# [[env]]
-# env_id = "primeintellect/reverse-text"
-
-# [[env]]
-# env_id = "primeintellect/wordle"
-
-# [[env]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[env]]
-# env_id = "primeintellect/opencode-harbor"
-
-[gepa]
-max_calls = 500
-num_train = 100
-num_val = 50
-minibatch_size = 3
-max_concurrent = 32
-
-[sampling]
-max_tokens = 2048
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/gepa/qwen-3-5.toml b/configs/gepa/qwen-3-5.toml
deleted file mode 100644
index cce5985aad..0000000000
--- a/configs/gepa/qwen-3-5.toml
+++ /dev/null
@@ -1,36 +0,0 @@
-# Qwen3.5 dense models. Update both fields when switching sizes.
-# model = "Qwen/Qwen3.5-0.8B"
-# reflection_model = "Qwen/Qwen3.5-0.8B"
-# model = "Qwen/Qwen3.5-2B"
-# reflection_model = "Qwen/Qwen3.5-2B"
-model = "Qwen/Qwen3.5-4B"
-reflection_model = "Qwen/Qwen3.5-4B"
-# model = "Qwen/Qwen3.5-9B"
-# reflection_model = "Qwen/Qwen3.5-9B"
-
-[[env]]
-env_id = "primeintellect/wiki-search"
-
-# [[env]]
-# env_id = "primeintellect/reverse-text"
-
-# [[env]]
-# env_id = "primeintellect/wordle"
-
-# [[env]]
-# env_id = "primeintellect/dspy-flights"
-
-# [[env]]
-# env_id = "primeintellect/opencode-harbor"
-
-[gepa]
-max_calls = 500
-num_train = 100
-num_val = 50
-minibatch_size = 3
-max_concurrent = 32
-
-[sampling]
-max_tokens = 1024
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/rl/gpt-oss.toml b/configs/rl/gpt-oss.toml
deleted file mode 100644
index d8b0ee360a..0000000000
--- a/configs/rl/gpt-oss.toml
+++ /dev/null
@@ -1,40 +0,0 @@
-# gpt-oss models. Uncomment exactly one model.
-model = "openai/gpt-oss-20b"
-# model = "openai/gpt-oss-120b"
-
-max_steps = 100
-batch_size = 128
-rollouts_per_example = 8
-
-[sampling]
-max_tokens = 1024
-reasoning_effort = "low"
-
-# Start with one environment. Uncomment more [[env]] blocks for multi-env training.
-[[env]]
-id = "primeintellect/reverse-text"
-
-# [[env]]
-# id = "primeintellect/wordle"
-
-# [[env]]
-# id = "primeintellect/wiki-search"
-
-# [[env]]
-# id = "primeintellect/dspy-flights"
-
-# [[env]]
-# id = "primeintellect/opencode-harbor"
-
-# Online evaluations during training. Uncomment to add [[eval.env]] blocks.
-
-# [eval]
-# interval = 20
-# num_examples = 64
-# rollouts_per_example = 1
-# eval_base_model = true
-
-# [[eval.env]]
-# id = "primeintellect/reverse-text"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
\ No newline at end of file
diff --git a/configs/rl/llama.toml b/configs/rl/llama.toml
deleted file mode 100644
index 2be8fdda56..0000000000
--- a/configs/rl/llama.toml
+++ /dev/null
@@ -1,39 +0,0 @@
-# Llama 3.2 models. Uncomment exactly one model.
-model = "meta-llama/Llama-3.2-1B-Instruct"
-# model = "meta-llama/Llama-3.2-3B-Instruct"
-
-max_steps = 100
-batch_size = 128
-rollouts_per_example = 8
-
-[sampling]
-max_tokens = 1024
-
-# Start with one environment. Uncomment more [[env]] blocks for multi-env training.
-[[env]]
-id = "primeintellect/reverse-text"
-
-# [[env]]
-# id = "primeintellect/wordle"
-
-# [[env]]
-# id = "primeintellect/wiki-search"
-
-# [[env]]
-# id = "primeintellect/dspy-flights"
-
-# [[env]]
-# id = "primeintellect/opencode-harbor"
-
-# Online evaluations during training. Uncomment to add [[eval.env]] blocks.
-
-# [eval]
-# interval = 20
-# num_examples = 64
-# rollouts_per_example = 1
-# eval_base_model = true
-
-# [[eval.env]]
-# id = "primeintellect/reverse-text"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/rl/nemotron.toml b/configs/rl/nemotron.toml
deleted file mode 100644
index 545f7aecbf..0000000000
--- a/configs/rl/nemotron.toml
+++ /dev/null
@@ -1,40 +0,0 @@
-# Nemotron 3 models. Uncomment exactly one model.
-model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
-# model = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
-
-max_steps = 100
-batch_size = 128
-rollouts_per_example = 8
-
-[sampling]
-max_tokens = 1024
-enable_thinking = false
-
-# Start with one environment. Uncomment more [[env]] blocks for multi-env training.
-[[env]]
-id = "primeintellect/reverse-text"
-
-# [[env]]
-# id = "primeintellect/wordle"
-
-# [[env]]
-# id = "primeintellect/wiki-search"
-
-# [[env]]
-# id = "primeintellect/dspy-flights"
-
-# [[env]]
-# id = "primeintellect/opencode-harbor"
-
-# Online evaluations during training. Uncomment to add [[eval.env]] blocks.
-
-# [eval]
-# interval = 20
-# num_examples = 64
-# rollouts_per_example = 1
-# eval_base_model = true
-
-# [[eval.env]]
-# id = "primeintellect/reverse-text"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/rl/qwen-moe.toml b/configs/rl/qwen-moe.toml
deleted file mode 100644
index 98904553d0..0000000000
--- a/configs/rl/qwen-moe.toml
+++ /dev/null
@@ -1,42 +0,0 @@
-# Qwen3.5/3.6 MoE models. Uncomment exactly one model.
-model = "Qwen/Qwen3.6-35B-A3B"
-# model = "Qwen/Qwen3.5-35B-A3B"
-# model = "Qwen/Qwen3.5-122B-A10B"
-# model = "Qwen/Qwen3.5-397B-A17B"
-
-max_steps = 100
-batch_size = 256
-rollouts_per_example = 8
-
-[sampling]
-max_tokens = 1024
-enable_thinking = false
-
-# Start with one environment. Uncomment more [[env]] blocks for multi-env training.
-[[env]]
-id = "primeintellect/reverse-text"
-
-# [[env]]
-# id = "primeintellect/wordle"
-
-# [[env]]
-# id = "primeintellect/wiki-search"
-
-# [[env]]
-# id = "primeintellect/dspy-flights"
-
-# [[env]]
-# id = "primeintellect/opencode-harbor"
-
-# Online evaluations during training. Uncomment to add [[eval.env]] blocks.
-
-# [eval]
-# interval = 20
-# num_examples = 64
-# rollouts_per_example = 1
-# eval_base_model = true
-
-# [[eval.env]]
-# id = "primeintellect/reverse-text"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/rl/qwen.toml b/configs/rl/qwen.toml
deleted file mode 100644
index 152dec897b..0000000000
--- a/configs/rl/qwen.toml
+++ /dev/null
@@ -1,42 +0,0 @@
-# Qwen3.5 dense models. Uncomment exactly one model.
-# model = "Qwen/Qwen3.5-0.8B"
-# model = "Qwen/Qwen3.5-2B"
-model = "Qwen/Qwen3.5-4B"
-# model = "Qwen/Qwen3.5-9B"
-
-max_steps = 100
-batch_size = 128
-rollouts_per_example = 8
-
-[sampling]
-max_tokens = 1024
-enable_thinking = false
-
-# Start with one environment. Uncomment more [[env]] blocks for multi-env training.
-[[env]]
-id = "primeintellect/reverse-text"
-
-# [[env]]
-# id = "primeintellect/wordle"
-
-# [[env]]
-# id = "primeintellect/wiki-search"
-
-# [[env]]
-# id = "primeintellect/dspy-flights"
-
-# [[env]]
-# id = "primeintellect/opencode-harbor"
-
-# Online evaluations during training. Uncomment to add [[eval.env]] blocks.
-
-# [eval]
-# interval = 20
-# num_examples = 64
-# rollouts_per_example = 1
-# eval_base_model = true
-
-# [[eval.env]]
-# id = "primeintellect/reverse-text"
-
-# Hosted Training docs: https://docs.primeintellect.ai/hosted-training
diff --git a/configs/zero3.yaml b/configs/zero3.yaml
deleted file mode 100644
index 21ab374fb9..0000000000
--- a/configs/zero3.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-compute_environment: LOCAL_MACHINE
-debug: false
-deepspeed_config:
-  deepspeed_multinode_launcher: standard
-  offload_optimizer_device: none
-  offload_param_device: none
-  zero3_init_flag: true
-  zero3_save_16bit_model: true
-  zero_stage: 3
-distributed_type: DEEPSPEED
-downcast_bf16: 'no'
-machine_rank: 0
-main_training_function: main
-mixed_precision: bf16
-num_machines: 1
-num_processes: 8
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
\ No newline at end of file
diff --git a/pyproject.toml b/pyproject.toml
index 888a3ac443..c39a9ab95f 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -107,33 +107,6 @@ required-version = ">=0.11.1"
 exclude-newer = "7 days"
 default-groups = ["dev", "examples"]
 
-# NeMo Gym declares its full platform stack; this integration only starts resource servers.
-[[tool.uv.dependency-metadata]]
-name = "nemo-gym"
-version = "0.4.0"
-requires-python = ">=3.12"
-requires-dist = [
-    "aiohttp>=3.14.1",
-    "fastapi",
-    "gprof2dot",
-    "hydra-core",
-    "itsdangerous",
-    "mcp>=1.27,<2",
-    "omegaconf",
-    "openai>=2.9",
-    "orjson",
-    "pandas",
-    "pydantic",
-    "pydot",
-    "ray>=2.55.1",
-    "requests",
-    "rich",
-    "typing-extensions",
-    "uvicorn",
-    "wandb",
-    "yappi",
-]
-
 [[tool.uv.index]]
 name = "pypi"
 url = "https://pypi.org/simple"
@@ -244,8 +217,6 @@ addopts = [
 ]
 markers = [
     "e2e: marks end-to-end eval-run tests (need a model API key; skipped without one)",
-    "integration: marks tests as integration tests",
-    "slow: marks slow-running tests (deselect with -m 'not slow')",
     # v1 e2e matrix axes — select subsets with `-m` (see tests/v1/conftest.py). A mark is applied
     # per axis, so it matches every case touching that value on any axis.
     "subprocess: v1 e2e cases touching the subprocess (host) runtime",
@@ -264,11 +235,7 @@ markers = [
     "pi: v1 e2e cases on the pi harness",
     "pool: v1 e2e cases on the pool harness",
     "codex: v1 e2e cases on the codex harness",
-    "unit: marks tests as unit tests",
     "asyncio: marks tests as async tests",
-    "parsers: marks tests for parser components",
-    "rubrics: marks tests for rubric components",
-    "environments: marks tests for environment components",
 ]
 filterwarnings = [
     "ignore::DeprecationWarning",
@@ -276,7 +243,6 @@ filterwarnings = [
     "ignore::UserWarning:transformers.*",
     "ignore:The 'repr' attribute.*:UserWarning:pydantic.*",
     "ignore:The 'frozen' attribute.*:UserWarning:pydantic.*",
-    "ignore:.*is.*with.*str.*literal:SyntaxWarning:cellpylib.*",
 ]
 asyncio_mode = "auto"
 norecursedirs = [".git", ".tox", "dist", "build", "*.egg", "__pycache__"]
diff --git a/tests/v1/test_configs.py b/tests/v1/test_configs.py
index ac47253d95..36a6f52120 100644
--- a/tests/v1/test_configs.py
+++ b/tests/v1/test_configs.py
@@ -1,9 +1,7 @@
 """Every checked-in v1 eval config parses.
 
 Mirrors prime-rl's config test: glob the configs and assert each validates into its config
-type. The root `configs/*.toml` are the `uv run eval @ <file>` v1 configs (EvalConfig);
-`endpoints.toml` isn't an eval config, and `configs/eval|rl|gepa/` are the legacy
-`vf-eval` / training formats (different, non-v1 config classes), so both are out of scope here.
+type. The root `configs/*.toml` are the `uv run eval @ <file>` v1 configs (EvalConfig).
 """
 
 import tomllib
@@ -13,11 +11,7 @@
 
 from verifiers.v1.configs.cli.eval import EvalConfig
 
-CONFIGS = sorted(
-    p
-    for p in (Path(__file__).resolve().parents[2] / "configs").glob("*.toml")
-    if p.name != "endpoints.toml"
-)
+CONFIGS = sorted((Path(__file__).resolve().parents[2] / "configs").glob("*.toml"))
 
 
 @pytest.mark.parametrize("path", CONFIGS, ids=lambda p: p.name)
diff --git a/uv.lock b/uv.lock
index 1bf6ed66d4..3281d11a28 100644
--- a/uv.lock
+++ b/uv.lock
@@ -26,14 +26,6 @@ prime-pydantic-config = false
 renderers = false
 ruff = "2026-07-28T00:00:00Z"
 
-[manifest]
-
-[[manifest.dependency-metadata]]
-name = "nemo-gym"
-version = "0.4.0"
-requires-dist = ["aiohttp>=3.14.1", "fastapi", "gprof2dot", "hydra-core", "itsdangerous", "mcp>=1.27,<2", "omegaconf", "openai>=2.9", "orjson", "pandas", "pydantic", "pydot", "ray>=2.55.1", "requests", "rich", "typing-extensions", "uvicorn", "wandb", "yappi"]
-requires-python = ">=3.12"
-
 [[package]]
 name = "aiofile"
 version = "3.12.3"
diff --git a/verifiers/v1/agent.py b/verifiers/v1/agent.py
index 5e6a212ed9..456ac84ffe 100644
--- a/verifiers/v1/agent.py
+++ b/verifiers/v1/agent.py
@@ -9,7 +9,7 @@
 
 import asyncio
 import logging
-from collections.abc import AsyncIterator, Callable, Iterator, Mapping
+from collections.abc import AsyncIterator, Callable, Mapping
 from contextlib import asynccontextmanager, nullcontext
 from dataclasses import dataclass
 from typing import Self
@@ -733,9 +733,3 @@ def __getattr__(self, name: str) -> Agent:
                 f"{sorted(agents) if agents else []}"
             )
         return agents[name]
-
-    def __iter__(self) -> Iterator[Agent]:
-        return iter(self._agents.values())
-
-    def __len__(self) -> int:
-        return len(self._agents)
diff --git a/verifiers/v1/cli/eval/main.py b/verifiers/v1/cli/eval/main.py
index b245e42f4a..338087cb91 100644
--- a/verifiers/v1/cli/eval/main.py
+++ b/verifiers/v1/cli/eval/main.py
@@ -46,11 +46,9 @@ def main(argv: list[str] | None = None) -> None:
                 narrow_config(EvalConfig, argv)
             )  # full option help, narrowed to the given ids
         return
-    # An env-block flag (or a since-moved flat axis) skips the usage gate so the
-    # typed parse renders its did-you-mean instead of a bare usage line.
-    typed_axis = any(
-        a.startswith(("--env.", "--taskset.", "--harness.", "--serve.")) for a in argv
-    )
+    # An env-block flag skips the usage gate so the typed parse renders its
+    # did-you-mean instead of a bare usage line.
+    typed_axis = any(a.startswith(("--env.", "--serve.")) for a in argv)
     if (
         not extract_id(argv, "env.taskset")
         and not references_config_file(argv)
diff --git a/verifiers/v1/cli/gepa.py b/verifiers/v1/cli/gepa.py
index d0a0599f29..cf09b5c24e 100644
--- a/verifiers/v1/cli/gepa.py
+++ b/verifiers/v1/cli/gepa.py
@@ -42,7 +42,7 @@ def main(argv: list[str] | None = None) -> None:
                 narrow_config(GEPAConfig, argv)
             )  # full option help, narrowed to the given ids
         return
-    typed_axis = any(a.startswith(("--env.", "--taskset.", "--harness.")) for a in argv)
+    typed_axis = any(a.startswith("--env.") for a in argv)
     if (
         not extract_id(argv, "env.taskset")
         and not references_config_file(argv)
diff --git a/verifiers/v1/cli/output.py b/verifiers/v1/cli/output.py
index f651bd5844..979d5f523b 100644
--- a/verifiers/v1/cli/output.py
+++ b/verifiers/v1/cli/output.py
@@ -72,13 +72,10 @@ def attempt_log_file(run_dir: Path) -> Path:
 
 
 def saved_config_path(run_dir: Path) -> Path | None:
-    """The run's saved resolved config (`configs/resolved/<cli>.json`; legacy runs
-    kept it at `configs/<cli>.json`), None if absent."""
-    for config_dir in (run_dir / CONFIG_DIR / RESOLVED_DIR, run_dir / CONFIG_DIR):
-        candidates = sorted(config_dir.glob("*.json")) if config_dir.is_dir() else []
-        if candidates:
-            return candidates[0]
-    return None
+    """The run's saved resolved config (`configs/resolved/<cli>.json`), None if absent."""
+    config_dir = run_dir / CONFIG_DIR / RESOLVED_DIR
+    candidates = sorted(config_dir.glob("*.json")) if config_dir.is_dir() else []
+    return candidates[0] if candidates else None
 
 
 # Compiling an adapter is the expensive part; run output reuses only a few model classes.
diff --git a/verifiers/v1/harness.py b/verifiers/v1/harness.py
index 1c00ab4e32..05483daa83 100644
--- a/verifiers/v1/harness.py
+++ b/verifiers/v1/harness.py
@@ -117,31 +117,6 @@ async def install_skills(self, runtime: Runtime, dest: str) -> None:
                 # `write` moves bytes, not modes; restore the execute bits scripts need.
                 await runtime.run(["chmod", "+x", *executables], {})
 
-    async def run(
-        self,
-        ctx: ModelContext,
-        trace: Trace,
-        runtime: Runtime,
-        endpoint: str,
-        secret: str,
-        mcp_urls: dict[str, str],
-        data: TaskData,
-        messages: Messages | None = None,
-    ) -> None:
-        """Run ONE segment of the exchange without retaining a process: the program
-        from launch (or, with `messages`, the user's next turn(s) via `resume`) until
-        it yields. The rollout loop owns the exchange across segments."""
-        async with boundary(HarnessError, f"harness {self.config.id!r}"):
-            if messages is None:
-                result = await self.launch(
-                    ctx, trace, runtime, endpoint, secret, mcp_urls, data
-                )
-            else:
-                result = await self.resume(
-                    ctx, trace, runtime, endpoint, secret, mcp_urls, data, messages
-                )
-        await self._check_result(trace, runtime, result)
-
     async def _check_result(
         self, trace: Trace, runtime: Runtime, result: ProgramResult
     ) -> None:
diff --git a/verifiers/v1/harnesses/prime_agent/harness.py b/verifiers/v1/harnesses/prime_agent/harness.py
index bcbe2b701e..2882f51c44 100644
--- a/verifiers/v1/harnesses/prime_agent/harness.py
+++ b/verifiers/v1/harnesses/prime_agent/harness.py
@@ -97,7 +97,6 @@ class PrimeAgentHarnessConfig(HarnessConfig):
 class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
     APPENDS_SYSTEM_PROMPT = True
     SUPPORTS_MCP = True
-    SUPPORTS_RESUME = True
     SUPPORTS_SKILLS = True
 
     def acp_turn_result(self, trace: Trace, result: ACPTurn) -> None:
diff --git a/verifiers/v1/mcp/launch.py b/verifiers/v1/mcp/launch.py
index 82e2da2698..eaf7020da8 100644
--- a/verifiers/v1/mcp/launch.py
+++ b/verifiers/v1/mcp/launch.py
@@ -201,9 +201,8 @@ async def _install_in_sandbox(server: ServerBase, runtime: Runtime) -> str:
         await runtime.write(env_remote, env_data)
     venv = str(PurePosixPath(workdir) / ".vf-venv")
     root_q, temp_q, cache_q, venv_q = map(shlex.quote, (root, temp, cache, venv))
-    extras = ",".join(type(server).EXTRAS)
     vf_source = shlex.quote(vf_remote)
-    env_source = shlex.quote(env_remote + (f"[{extras}]" if extras else ""))
+    env_source = shlex.quote(env_remote)
     setup = (
         f"set -e; mkdir -p {root_q} {temp_q} {cache_q}; "
         f"export TMPDIR={temp_q} UV_CACHE_DIR={cache_q}; "
@@ -279,12 +278,7 @@ async def serve_in_runtime(
         env["MCP_PORT_FILE"] = port_file
     python = sys.executable
     if runtime.type != "subprocess":
-        # Prebuilt runtime images may already contain the server and its Python
-        # environment. Let those servers bypass redundant per-launch source
-        # uploads and installation while preserving the existing default.
-        python = getattr(type(server), "RUNTIME_PYTHON", None)
-        if python is None:
-            python = await _install_in_sandbox(server, runtime)
+        python = await _install_in_sandbox(server, runtime)
     command = [python, "-m", type(server).__module__]
     if runtime.type != "subprocess":
         # Providers may invoke uv after the install shell exits, so preserve its PATH.
diff --git a/verifiers/v1/mcp/server.py b/verifiers/v1/mcp/server.py
index d58f4dafc1..2ea101e338 100644
--- a/verifiers/v1/mcp/server.py
+++ b/verifiers/v1/mcp/server.py
@@ -134,9 +134,6 @@ class ServerBase(Generic[ConfigT, StateT]):
     tools bare (no `<server>_` prefix); name collisions across servers are then the taskset
     author's concern."""
 
-    EXTRAS: ClassVar[tuple[str, ...]] = ()
-    """Package extras the server's module needs, applied at sandbox install."""
-
     def __init__(self, config: ConfigT) -> None:
         self.config = config
         self._state_cls = state_cls(type(self))