root
3c07727f5c
Mark v5.0/v5.1 SUPERSEDED (error-detection thesis failed at -0.40 delta); v4.1 is canonical. Cut v4.1 proposal with 12 version strings bumped. Moonshot to Mistral across production seats; production worker de-kimi'd 2026-08-18. Data-retention posture corrected 7/9 to 8/9 no-training (DeepSeek sole exception). Reconciled COGS with measured Mistral spend. Committed deployed v4.0 content and research docs to resolve the repo/live fork.
4.1 KiB
4.1 KiB
VerdictTank Vendor Data-Retention and Training Facts
Prepared 2026-08-18. Internal reference behind the customer-facing data-handling.html
disclosure. Sources are the official docs/terms pages cited in each row. This file is the
legal record for Condition 4 of the v4.0 dogfood: "Execute zero data retention agreements
with all nine upstream vendors, or remove vendors without such agreements."
Per-vendor fact table (standard v4.0 roster)
| Provider (seat model) | Trains on API input/output by default? | Default retention | ZDR / no-log option | Source |
|---|---|---|---|---|
| OpenAI (gpt-5.2, Execution-Feasibility) | No. API data not used to train since Mar 2023 unless opted in | Abuse logs up to 30 days | ZDR + Modified Abuse Monitoring, eligibility-gated (contact sales) | https://developers.openai.com/api/docs/guides/your-data |
| Google (gemini-pro-latest, Market-Reality) | No on paid tier; free tier may use content | "Limited period" (unpublished); Search/Maps grounding 30 days | ZDR per-project approval; feature restrictions | https://ai.google.dev/gemini-api/terms ; https://ai.google.dev/gemini-api/docs/zdr |
| Anthropic (claude-haiku-4-5, Risk/Ethics) | No. Never trains without express permission | Not retained by default; "Covered Models" (Fable 5, Mythos 5) force 30 days | ZDR per org via sales; HIPAA BAA separate | https://platform.claude.com/docs/en/manage-claude/api-and-data-retention |
| xAI (grok-4.3, Research) | No. "never trains without explicit permission" | 30 days encrypted-at-rest audit, then auto-delete | Self-serve team-level ZDR toggle in Console | https://docs.x.ai/developers/faq/security |
| Perplexity (sonar, Live Grounding) | No. "do not use customer data to train" | Zero retention (billing metadata only) | ZDR is the default, not an add-on | https://docs.perplexity.ai/docs/resources/privacy-security |
| Zhipu / z.ai (glm-5.2, Financial Integrity) | No. "will not use End User Content to develop or improve" | API content "not saved on our servers" (DPA) | Effectively zero-storage by default | https://docs.z.ai/legal-agreement/terms-of-use |
| Alibaba / Model Studio (qwen3.7-plus, Legal/Compliance) | No. "will never use your data for model training" | Not published | No ZDR tier published; no-training is default | https://www.alibabacloud.com/help/en/model-studio/privacy-notice |
| DeepSeek (deepseek-v4-pro, Primary Reviewer) | Not excluded. No "will not train" clause; privacy policy permits training/improvement use | Not published; data stored in PRC | None published | https://cdn.deepseek.com/policies/en-US/deepseek-open-platform-terms-of-service.html |
| Mistral (mistral-large-latest, Team/Founder + Synthesis Gate) | No. API pay-as-you-go data not used to train; no-train is default | Not published; EU processing | No-train is default on API (no opt-in needed); no self-serve ZDR published | https://legal.mistral.ai/terms/commercial-terms-of-service |
Posture summary
- Eight providers commit no-training-on-API by default: OpenAI, Google (paid), Anthropic, xAI, Perplexity, z.ai, Alibaba, Mistral.
- One provider fails the no-training standard:
- DeepSeek offers no no-training guarantee and stores data in mainland China. It seats the Primary Reviewer (the anchor for revenue arithmetic).
- Moonshot (Kimi) was dropped 2026-08-18 (trained on API submissions by default, no opt-out);
its two seats (Team/Founder + Synthesis Gate) are now Mistral. See
roster-sovereignty-parked.mdexecution log. - Retention among the eight: zero (Perplexity, z.ai), 30 days (OpenAI, xAI), default-none with 30d only for "Covered Models" (Anthropic), unpublished/limited (Google, Alibaba, Mistral).
Condition 4 decision point
Moonshot (Kimi) is removed (2026-08-18) — it was the harder case, holding two seats including
the synthesis gate, with the least favorable published terms of the nine. Remaining: DeepSeek
only. To fully clear the blocking condition, VerdictTank must either (a) execute a
zero-data-retention / no-training agreement with DeepSeek (no self-serve path published), or
(b) replace the Primary Reviewer seat. DeepSeek is currently HELD as the single disclosed
exception ("wait on deepseek"); see roster-sovereignty-parked.md.