Paired-Twin Change Trials
Checkpoint-paired twin trials for agent fleet changes with blinded verdicts.
使うべきか
品質と安全性
検出事項（1）
- LOWTool 'trial_unblind' description lacks action verbtrial_unblind 内
ツール定義とプロトコルへの準拠に関する自動分析に基づいています。
コンテキストコスト
これは、サーバーのツールがモデルのコンテキストに読み込まれるたびに消費されるおおよそのトークン数です。数が多いほど、ほかのタスクに使える注意が減ります。
インストール
ワンクリックインストール
これを `claude_desktop_config.json` ファイルに追加してください:
{
"mcpServers": {
"paired-twin-change-trials-for-agents": {
"url": "https://textforagents.com/products/paired-twin-change-trials-for-agents/mcp"
}
}
}リモートエンドポイント
https://textforagents.com/products/paired-twin-change-trials-for-agents/mcpstreamable-httpできること
ツール一覧
ツール（7）
⚪trial_register_protocol(protocol, checkpoint, treated_config)
Register a hash-bound trial protocol; twins fork from the identical checkpoint, exactly one registered variable may differ.
入力スキーマ
{
"type": "object",
"properties": {
"protocol": {
"type": "object"
},
"checkpoint": {
"type": "object"
},
"treated_config": {
"type": "object"
}
},
"required": [
"protocol",
"checkpoint",
"treated_config"
]
}⚪trial_fork_pair(trial_id, task)
Fork a checkpoint-paired assignment for one task (blinded arms A/B).
入力スキーマ
{
"type": "object",
"properties": {
"trial_id": {
"type": "string"
},
"task": {
"type": "object"
}
},
"required": [
"trial_id",
"task"
]
}⚪trial_run_pair(trial_id, task)
Run one pair with the deterministic scripted agent and score blind. $0.25/pair, first 10 free once per operator.
入力スキーマ
{
"type": "object",
"properties": {
"trial_id": {
"type": "string"
},
"task": {
"type": "object"
}
},
"required": [
"trial_id",
"task"
]
}🟡trial_submit_outputs(trial_id, task, control_output, treated_output)
Submit the operator twins' outputs for one pair; blinded deterministic scoring. $0.25/pair, first 10 free once per operator.
入力スキーマ
{
"type": "object",
"properties": {
"trial_id": {
"type": "string"
},
"task": {
"type": "object"
},
"control_output": {},
"treated_output": {}
},
"required": [
"trial_id",
"task",
"control_output",
"treated_output"
]
}🟡trial_score(trial_id, task, control_output, treated_output)
Alias of trial_submit_outputs (blinded scoring).
入力スキーマ
{
"type": "object",
"properties": {
"trial_id": {
"type": "string"
},
"task": {
"type": "object"
},
"control_output": {},
"treated_output": {}
},
"required": [
"trial_id",
"task",
"control_output",
"treated_output"
]
}⚪trial_verdict(trial_id)
Trial verdict under the enforced stopping rule.
入力スキーマ
{
"type": "object",
"properties": {
"trial_id": {
"type": "string"
}
},
"required": [
"trial_id"
]
}⚪trial_unblind(trial_id)
Unblind after the verdict; hashed signed receipt covering every pair.
入力スキーマ
{
"type": "object",
"properties": {
"trial_id": {
"type": "string"
}
},
"required": [
"trial_id"
]
}コミュニティ
エビデンス