01 / Access Flow

Access 是受治理的 developer path。

API 不是简单发放 key。Dev Credits、Dev Schooling、key provisioning、SDK download、本地 workspace 和 first call 组成一条受控路线。

Access contract
  • API 使用需要 Dev Credits。
  • Dev Schooling provision API key、private key 和 public key reference。
  • SDK download 包含 Engine、TerminalRunner、Extension example、带 12 个 Python files 的 PythonRunner 以及 command documentation。
六步访问路线
  1. 01

    Dev Credits

    Value gate

    Dev Credits 会解锁 API 使用,并在任何 operation 运行前把 execution 绑定到 developer 余额。

  2. 03

    Keys

    Access 材料

    该 workflow 会 provision API access 材料,同时让 private key 处理保持在 request payloads 之外。

  3. 04

    SDK Download

    Runtime package

    SDK 包含 Engine、TerminalRunner、Extension example,以及带 12 个 Python files 的 PythonRunner。

  4. 05

    Workspace

    本地控制

    Developer 从本地 SDK workspace 开始,而不是从彼此分离的 snippets 开始。

  5. 06

    First API Call

    Execution 启动

    第一次 call 会针对 M13 gateway 使用 command、input、run ID、stage ID 和 operation ID。

02 / Curl Commands

选择一个 command 并测试 API。

每个 command 都使用同一个 M13 gateway contract:Authorization header、public key reference、JSON payload,以及可追踪的 run、stage、operation IDs。

m13_fast.curl
1
03 / Payload Explorer

检查真实的 Attack response。

此示例向真实 policy decision 发送 attack command,并显示完整 public response:content、artifact、Server IRSS、usage evidence、Dev Credit billing、ledger reference 和 local persistence status。

探索此次 call
选中的层

Attack request

该 request 展示 API command、被 Attack 的 decision object,以及紧凑的 output constraint。

这是由用户控制的 entry point:command、object 和 output budget。
attack_request.json
1{2  "command": "attack",3  "input": "Attack the U.S. government decision dated January 15, 2025, known as the Framework for Artificial Intelligence Diffusion, to restrict and regulate access to advanced computing ICs and model weights of advanced AI systems for several countries.\n\nHard limit: max 1,000 characters total. Do not exceed this limit.\nUse compact sections only:\nverdict\nstrategic_effectiveness\nmarket_distortion\nenforcement_realism\nally_friction\nfinal_pressure_judgment",4  "run_id": "run_attack_ai_diffusion_rule_1000_...",5  "stage_id": "stage_attack_ai_diffusion_rule_1000_...",6  "operation_id": "op_attack_ai_diffusion_rule_1000_..."7}
public_response.json
1{2  "ok": true,3  "status": "success",4  "run_id": "run_attack_ai_diffusion_rule_1000_...",5  "stage_id": "stage_attack_ai_diffusion_rule_1000_...",6  "operation_id": "op_attack_ai_diffusion_rule_1000_...",7  "content": "verdict: damaged\n\nstrategic_effectiveness: Tiered country classification creates binary in/out logic that ignores fluid geopolitical alignment shifts. Adversary access routes through tier-2 jurisdictions remain open. Controls lag chip architecture cycles by 12-18 months minimum.\n\nmarket_distortion: Concentrates compute supply advantage in incumbent US firms while accelerating sovereign chip programs in excluded nations. Pushes allied-nation AI labs toward non-US architectures, fragmenting interoperability long-term.\n\nenforcement_realism: Model weights are copyable digital artifacts. Export control frameworks built for physical goods collapse against weightfile diffusion via public code mirrors, academic channels, and offshore fine-tuning. Enforcement gap is structural, not resourcing.\n\nally_friction: Tier-2 allied democracies face compliance burdens indistinguishable from adversary treatment, generating diplomatic cost disproportionate to security gain. Drives procurement diversification away from US supply chains.\n\nfinal_pressure_judgment: Framework applies Cold War export-control logic to infinitely reproducible digital assets. Kill zone is enforcement feasibility against model weight diffusion. Survives only as signaling mechanism; collapses as actual containment architecture.",8  "artifact": {9    "artifact_id": "art_...",10    "artifact_type": "generic",11    "artifact_version": "v1",12    "status": "created",13    "origin": {14      "space": "api",15      "route": "/api/m13/llm",16      "run_id": "run_attack_ai_diffusion_rule_1000_...",17      "stage_id": "stage_attack_ai_diffusion_rule_1000_...",18      "operation_id": "op_attack_ai_diffusion_rule_1000_...",19      "ledger_entry_id": "led_...",20      "server_irss_ref": null21    },22    "output": {23      "kind": "text",24      "format": "markdown"25    },26    "payload": {27      "content": "verdict: damaged\n\nstrategic_effectiveness: Tiered country classification creates binary in/out logic that ignores fluid geopolitical alignment shifts. Adversary access routes through tier-2 jurisdictions remain open. Controls lag chip architecture cycles by 12-18 months minimum.\n\nmarket_distortion: Concentrates compute supply advantage in incumbent US firms while accelerating sovereign chip programs in excluded nations. Pushes allied-nation AI labs toward non-US architectures, fragmenting interoperability long-term.\n\nenforcement_realism: Model weights are copyable digital artifacts. Export control frameworks built for physical goods collapse against weightfile diffusion via public code mirrors, academic channels, and offshore fine-tuning. Enforcement gap is structural, not resourcing.\n\nally_friction: Tier-2 allied democracies face compliance burdens indistinguishable from adversary treatment, generating diplomatic cost disproportionate to security gain. Drives procurement diversification away from US supply chains.\n\nfinal_pressure_judgment: Framework applies Cold War export-control logic to infinitely reproducible digital assets. Kill zone is enforcement feasibility against model weight diffusion. Survives only as signaling mechanism; collapses as actual containment architecture."28    },29    "metadata_ref": null,30    "created_at": "2026-07-03T10:47:22.578Z",31    "integrity": {32      "payload_hash": "sha256:...",33      "metadata_hash": null,34      "artifact_hash": "sha256:..."35    }36  },37  "server_irss": {38    "system": "M13",39    "server_irss_version": "server_irss_v1",40    "space": "api_execution",41    "route": "/api/m13/llm",42    "timestamp": "2026-07-03T10:47:22.578Z",43    "runtime": {44      "runtime_env": "staging",45      "billing_context": "staging_dev",46      "ledger_context": "staging",47      "checkout_context": "disabled_admin_grant",48      "balance_source": "staging_dev_token_balance"49    },50    "identity": {51      "user_account_ref": "usr_...",52      "api_workspace_id": "ws_...",53      "public_key_ref": "sha256:...",54      "api_key_ref": "key_...",55      "auth_status": "accepted"56    },57    "operation": {58      "run_id": "run_attack_ai_diffusion_rule_1000_...",59      "stage_id": "stage_attack_ai_diffusion_rule_1000_...",60      "operation_id": "op_attack_ai_diffusion_rule_1000_...",61      "command": "attack",62      "operation_type": "single_llm_call"63    },64    "llm": {65      "called": true,66      "provider": null,67      "adapter": "anthropic_foundry",68      "deployment": null,69      "runtime_model": "claude-opus-4-6",70      "stop_reason": "end_turn"71    },72    "usage": {73      "input_tokens": 107,74      "output_tokens": 268,75      "cached_input_tokens": 11178,76      "cache_creation_input_tokens": 0,77      "total_tokens": 11553,78      "billable_tokens": 149379    },80    "billing": {81      "required": true,82      "status": "debited",83      "debited": true,84      "charged_amount": 4,85      "billing_unit": "dev_credits",86      "balance_after": 94087    },88    "artifact": {89      "artifact_id": "art_...",90      "artifact_type": "generic",91      "status": "created",92      "payload_hash": "sha256:..."93    },94    "artifact_metadata": {95      "metadata_ref": null,96      "metadata_status": "not_created",97      "metadata_hash": null98    },99    "ledger": {100      "ledger_entry_id": "led_...",101      "ledger_entry_created": true,102      "entry_type": "llm_execution"103    },104    "execution_context": {105      "execution_context_ref": null,106      "execution_context_status": "not_required",107      "execution_type": "direct_api_call",108      "current_step_id": "llm_call",109      "current_step_kind": "llm_call",110      "current_operation_type": "single_llm_call"111    },112    "drift": {113      "state": "none",114      "origin": "none",115      "risk": "none",116      "signals": []117    },118    "status": {119      "execution_status": "success",120      "ledger_entry_created": true,121      "local_persistence_required": true,122      "api_error_code": null123    }124  },125  "execution_context": null,126  "usage": {127    "input_tokens": 107,128    "output_tokens": 268,129    "cached_input_tokens": 11178,130    "cache_creation_input_tokens": 0,131    "total_tokens": 11553,132    "billable_tokens": 1493133  },134  "billing": {135    "required": true,136    "status": "debited",137    "debited": true,138    "charged_amount": 4,139    "billing_unit": "dev_credits",140    "balance_after": 940141  },142  "ledger_entry": {143    "id": "led_...",144    "status": "created"145  },146  "local_persistence_required": true147}
04 / Start

从 Dev Credits 开始。

当 access、keys 和 Dev Credits 准备好时,API 就绪。本地 SDK 属于 Build——这里的下一步是 activation。

Ready to run

获取 API access,通过 Dev Credits 计费 calls,并在 workspace 准备好后本地构建。

Dev Schooling prepares access and key handling. Dev Credits make execution visible and billable. Build explains the SDK, workspace, and local runtime in depth.

Public billing
Dev Credits

面向客户的 billing 通过 charged_amount、billing_unit 和 balance_after 展示。

Dev path
Dev Schooling

API 使用会在 Access、Key Setup 和负责的 workspace orientation 完成后开始。

API surface
IRSS + Ledger

API surface 可通过 response content、Server IRSS、billing、ledger 和 artifacts 进行检查。