harn-vm 0.10.126

# ── Provider data controls (retention / training) ────────────────────────────
#
# One owner for a per-provider fact hosts otherwise re-derive: what a provider
# lets a caller decide per request about retention and training, what happens
# when Harn sets nothing, and the documentation backing both.
#
# `control_scope`:
#   per_request — a documented header or body field Harn can set per call.
#   account     — the posture exists but is selected on the account, project,
#                 organization, or contract. Harn must not pretend a request
#                 can change it.
#   none        — the provider documents no control at either level.
#
# `retention_default` / `training_default` describe the provider's behavior
# with NOTHING set. `unspecified` means the provider publishes no answer; it is
# deliberately not the same claim as `not_retained` / `does_not_train`.
#
# Harn applies `request_controls` only when a caller asks for the
# `strictest_available` posture, and records a receipt on every request saying
# which of applied / no_control_available / provider_unresearched / not_
# requested occurred. See `llm/api/data_controls.rs`.
#
# Every provider in the catalog is either declared here or named in
# `[data_controls_audit].unverified`. Absence is not a classification.

[data_controls_audit]
reviewed_on = "2026-08-30"
expires_on = "2026-11-30"
tracking_issue = 7674
unverified = [
  "atlas",
  "baseten",
  "cloudflare_ai_gateway",
  "dashscope",
  "deepinfra",
  "flexai",
  "friendli",
  "github_models",
  "huggingface",
  "hunyuan",
  "hyperbolic",
  "inception",
  "minimax",
  "moonshot",
  "nebius",
  "nvidia",
  "parasail",
  "qianfan",
  "sambanova",
  "siliconflow",
  "vercel_ai_gateway",
  "volcengine_ark",
  "zai",
]

# ── Per-request controls ─────────────────────────────────────────────────────

# `store` is documented for both the Responses API and Chat Completions. The
# Responses API defaults to storing 30 days of application state; Chat
# Completions stores none by default. Setting `store: false` does not end the
# separate 30-day abuse-monitoring retention, which is an account-level (ZDR)
# concern with no request header.
[providers.openai.data_controls]
control_scope = "per_request"
retention_default = "retained"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://developers.openai.com/api/docs/guides/your-data"]
note = "store: false ends application-state retention only; the 30-day abuse-monitoring store is ended by an approved org/project ZDR configuration, which has no per-request form."

[[providers.openai.data_controls.request_controls]]
location = "body"
name = "store"
value = false
effect = "retention"
caveat = "Abuse-monitoring retention continues regardless; a ZDR organization forces store to false server-side."

# Azure exposes OpenAI's `store` as the stored-completions opt-in. The 30-day
# abuse-monitoring store is separate and is disabled only via the Modified
# Abuse Monitoring application.
[providers.azure_openai.data_controls]
control_scope = "per_request"
retention_default = "abuse_monitoring_only"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = [
  "https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/stored-completions",
  "https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/abuse-monitoring",
]
note = "Stored completions are off unless store is true; the separate abuse-monitoring store is an account-level application, not a request field."

[[providers.azure_openai.data_controls.request_controls]]
location = "body"
name = "store"
value = false
effect = "retention"

# OpenRouter is the only catalog provider with a per-request control over
# training as well as retention: `provider.data_collection` selects providers
# by their data policy, and `provider.zdr` restricts routing to zero-retention
# endpoints. Both are routing restrictions, so the effective posture still
# depends on the upstream provider actually serving the call.
[providers.openrouter.data_controls]
control_scope = "per_request"
retention_default = "unspecified"
training_default = "unspecified"
checked_on = "2026-08-30"
sources = [
  "https://openrouter.ai/docs/features/provider-routing",
  "https://openrouter.ai/docs/features/privacy-and-logging",
]
note = "Both controls restrict which upstream providers may serve the call; retention and training behavior is otherwise the routed provider's."

[[providers.openrouter.data_controls.request_controls]]
location = "body"
name = "provider.data_collection"
value = "deny"
effect = "training"
caveat = "Defaults to allow. Restricts routing to providers that do not collect prompts; it does not retroactively bind one that does."

[[providers.openrouter.data_controls.request_controls]]
location = "body"
name = "provider.zdr"
value = true
effect = "retention"

# Gemini's `store` is an Interactions API field. `generateContent` documents no
# equivalent, so the control is scoped to the Interactions dialect rather than
# invented on the sibling wire.
[providers.gemini.data_controls]
control_scope = "per_request"
retention_default = "abuse_monitoring_only"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = [
  "https://ai.google.dev/gemini-api/docs/zdr",
  "https://ai.google.dev/gemini-api/docs/logs-policy",
]
note = "Paid tier only. Full zero-data-retention is a per-project approval; Search grounding retains for 30 days regardless."

[[providers.gemini.data_controls.request_controls]]
location = "body"
name = "store"
value = false
effect = "retention"
applies_to = ["gemini_interactions_sse"]

# ── Account- or contract-scoped providers ────────────────────────────────────

[providers.anthropic.data_controls]
control_scope = "account"
retention_default = "abuse_monitoring_only"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = [
  "https://platform.claude.com/docs/en/api/messages",
  "https://platform.claude.com/docs/en/manage-claude/api-and-data-retention",
]
note = "The Messages API documents no retention parameter and no retention-related anthropic-beta value; zero data retention is enabled per organization by the account team. Some models require 30-day retention and reject ZDR organizations."

[providers.vertex.data_controls]
control_scope = "account"
retention_default = "abuse_monitoring_only"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://cloud.google.com/vertex-ai/generative-ai/docs/data-governance"]
note = "Governed by project configuration; implicit context caching is on by default in the serving region."

[providers.bedrock.data_controls]
control_scope = "account"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html"]
note = "Retention is set by the account/project data_retention mode, not by the Converse request Harn sends. AWS documents store: false on its Responses surface as insufficient for guaranteed zero retention."

[providers.mistral.data_controls]
control_scope = "account"
retention_default = "abuse_monitoring_only"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = [
  "https://docs.mistral.ai/admin/monitor-comply/privacy-data-controls",
  "https://docs.mistral.ai/admin/monitor-comply/zero-data-retention",
]
note = "Paid API traffic is excluded from training by default; free-tier training is opted out in the Admin Console. ZDR is a contract."

[providers.groq.data_controls]
control_scope = "account"
retention_default = "abuse_monitoring_only"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://console.groq.com/docs/your-data"]
note = "Inference content is not retained by default; troubleshooting and abuse logs are kept up to 30 days."

[providers.xai.data_controls]
control_scope = "account"
retention_default = "retained"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://docs.x.ai/developers/faq/security"]
note = "30-day encrypted retention. Zero data retention is a team-level arrangement; xAI documents no way to enable it for a specific API key."

[providers.cohere.data_controls]
control_scope = "account"
retention_default = "abuse_monitoring_only"
training_default = "trains"
checked_on = "2026-08-30"
sources = ["https://cohere.com/enterprise-data-commitments"]
note = "Training is opted in by default and is turned off in the dashboard under Settings, Data Controls; there is no request field."

[providers.together.data_controls]
control_scope = "account"
retention_default = "not_retained"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://docs.together.ai/docs/privacy-and-security"]
note = "Inputs and outputs are not stored by default; training is opt-in."

[providers.fireworks.data_controls]
control_scope = "account"
retention_default = "not_retained"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://docs.fireworks.ai/guides/security_compliance/data_handling"]
note = "Zero retention by default on the chat-completions surface Harn dials. Fireworks documents store on its Responses API, which this route does not use."

[providers.cerebras.data_controls]
control_scope = "none"
retention_default = "not_retained"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://support.cerebras.net/articles/1811589793-does-cerebras-retain-my-data"]
note = "Cerebras documents that it retains no prompts, requests, responses, or logs, so there is nothing to opt out of."

[providers.deepseek.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "trains"
checked_on = "2026-08-30"
sources = [
  "https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html",
  "https://cdn.deepseek.com/policies/en-US/deepseek-open-platform-terms-of-service.html",
]
note = "The privacy policy asserts a general opt-out right but publishes no API mechanism, and does not separate API traffic from consumer chat. Treat as training by default."

# ── Local and self-hosted endpoints ──────────────────────────────────────────
#
# No third party receives the request. Retention is whatever the operator's own
# server does, which is why these declare `unspecified` rather than
# `not_retained`: Harn is not in a position to assert the operator's behavior.

[providers.ollama.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://github.com/ollama/ollama/blob/main/docs/api.md"]
note = "Local endpoint. The API documents no retention or training control because no third party receives the request."

[providers.llamacpp.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md"]
note = "Self-hosted endpoint. Retention is the operator's server behavior."

[providers.vllm.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html"]
note = "Self-hosted endpoint. Retention is the operator's server behavior."

[providers.tgi.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://huggingface.co/docs/text-generation-inference/en/reference/api_reference"]
note = "Self-hosted endpoint. Retention is the operator's server behavior."

[providers.mlx.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://github.com/ml-explore/mlx-lm/blob/main/mlx_lm/SERVER.md"]
note = "Self-hosted endpoint. Retention is the operator's server behavior."

[providers.local.data_controls]
control_scope = "none"
retention_default = "unspecified"
training_default = "does_not_train"
checked_on = "2026-08-30"
sources = ["https://platform.openai.com/docs/api-reference/chat"]
note = "Generic user-configured OpenAI-compatible endpoint. The wire contract it speaks documents no retention control beyond store, which a local server is not required to honor."