07 — Trust Architecture

Security by architecture,
not by promise.

Last Reviewed
May 22, 2026
Version
Protocol v1.0
Section 01

Every tool call requires your approval.

Dropstone CLI treats every model output as untrusted text. Whatever the model writes — code, shell commands, URLs to fetch, file edits — does nothing on your machine until you explicitly approve it. In Build mode, every destructive action prompts you with three choices: Allow once, Allow always, or Reject.

This is the security boundary. It is the same control surface that Anthropic, Databricks, NIST AI RMF, and every serious agentic-AI security framework converges on. The industry term is Human-in-the-Loop (HITL); the engineering principle is treat-model-as-untrusted-input.

What this means in practice

If a model were to suggest a malicious shell command, file write, or network call, it would appear in your terminal as a proposed action. Nothing executes. You read it, decide, and either approve or reject. There is no path from model output to your filesystem that bypasses you.

Note: Accept-All mode auto-approves tool calls for the current session. It is intentionally opt-in and clearly labeled in amber. Use it only when you have already reviewed the agent's plan.

Section 02

Inference runs on US-hosted providers only.

Every Dropstone chat traverses one of three providers, all US-hosted and selected for their published latency, capacity, and data-handling guarantees:

  • DeepInfra — primary for Dropstone Fast and Dropstone Heavy
  • Fireworks AI — primary for Dropstone Pro
  • Together AI — secondary failover for all three tiers

Provider selection is enforced in code, not configuration — China-hosted endpoints (DeepSeek's native API, Moonshot's native API) are excluded by design and cannot be selected at runtime. Failover between US providers is permitted; falling out of the US-only set is not.

Section 03

Zero data retention, enforced at the API layer.

Every upstream call to OpenRouter is sent with data_collection: 'deny'. The request fails closed if any provider in the routing chain would retain the prompt or completion. This is a contractual guarantee at the API boundary, not a promise in marketing copy.

The Dropstone server itself does not persist prompt or completion content. Request logs capture metadata only: model identifier, token counts, latency, cost, status code. Prompt text, attached files, and model responses are never written to disk and never sent to analytics platforms.

What we log

  • Model used, plan tier, token counts, cost
  • Request timestamp, latency, status code
  • IP address and user agent (for fraud and abuse detection)
  • Cache hit / miss counters for the vision preprocessor

What we do not log

  • Prompt text or completion text
  • Attached file contents
  • Image bytes or captions
  • Tool call arguments or results
Section 04

The best open-weight model, every month.

Dropstone Fast 1.5, Pro 1.5, and Heavy 1.5 are branded names. The underlying model behind each tier is selected monthly by Blankline Research's open-weight frontier evaluation, and the brand stays stable as the model changes underneath.

The version suffix 1.5 encodes Year 1, Month 5 — May 2026. Each monthly cycle re-evaluates the open-weight frontier on cost-adjusted SWE-bench, agentic tool-use, and vision benchmarks and pins the best fit to each tier.

Current selection — May 2026

  • Dropstone Fast 1.5 — built on DeepSeek V4 Flash
  • Dropstone Pro 1.5 — built on DeepSeek V4 Pro
  • Dropstone Heavy 1.5 — built on Moonshot Kimi K2.6

The underlying model is an implementation detail. The brand is the stable interface. When a better model emerges, the brand follows.

Section 05

Vision pipeline with cryptographic caching.

Images attached to any Dropstone tier are processed through Dropstone Vision 1.5, currently built on Gemini 3.5 Flash (84% MMMU-Pro, tied for industry leadership). The caption is then forwarded as text to the user's selected reasoning model.

Captions are keyed by SHA-256 of the image bytes. The same image attached twice returns the cached caption with sub-millisecond latency and zero additional API cost. Cache entries expire within 24 hours and do not persist beyond the server process lifetime.

Section 06

What we cannot prove.

We cannot mathematically prove that any frontier foundation model is free of embedded behaviors. Goldwasser, Kim, Vaikuntanathan & Zamir (2022) proved that no party can — this applies equally to Claude, GPT-5, Gemini, and every closed-weight model in production today. The limit is cryptographic, not engineering.

We acknowledge this openly. Our security guarantees do not depend on proving the model is clean. They depend on the runtime architecture above: the model cannot affect your machine without your explicit approval, and inference cannot retain your data because we forbid it at the API boundary.

Model origin is a security non-issue when the runtime treats every model as adversarial. We do.

Section 07

Subprocessors.

The complete list of third parties that may process customer data on Dropstone's behalf:

  • OpenRouter (United States) — model gateway and provider routing
  • DeepInfra (United States) — primary inference provider for Fast and Heavy tiers
  • Fireworks AI (United States) — primary inference provider for Pro tier
  • Together AI (United States) — secondary inference failover
  • Google AI Studio (United States) — Gemini 3.5 Flash for the vision preprocessor
  • Stripe (United States) — payment processing

This list is the full set. Changes are reflected here within seven business days of any addition or removal.

Section 08

Compliance & certifications.

Dropstone is a thin authenticated proxy — we do not persist customer code, prompts, completions, or attached files. Customer data is processed only by subprocessors that maintain their own current security certifications. The full chain is audited end-to-end through inheritance:

SubprocessorCertifications
OpenRouterSOC 2 Type II
DeepInfraSOC 2 Type II, HIPAA-ready
Fireworks AISOC 2 Type II, HIPAA
Together AISOC 2 Type II
Google AI StudioSOC 1 / 2 / 3, ISO 27001 / 17 / 18, FedRAMP High, HIPAA, PCI DSS
StripePCI DSS Level 1, SOC 1, SOC 2, ISO 27001

A Dropstone-level SOC 2 Type I audit is planned for our enterprise GA. We commission it earlier on request — enterprise prospects with procurement requirements can write to [email protected] to schedule.

Inherited compliance is not a substitute for our own audit — it is the floor we operate above. Certifications listed reflect each subprocessor's current published status; verify directly with the provider before relying on them for procurement.

Section 09

Enterprise & regulated environments.

For organizations operating under FedRAMP-adjacent, DoD-procurement, or specific regulated-finance requirements where open-weight model provenance restrictions apply, Dropstone offers an Enterprise tier built on US-trained open-weight models. The runtime architecture is identical; only the underlying model selection differs.

VPC deployment, BYOK encryption, and SOC 2 documentation are available on request.

Contact [email protected].

Section 10

Reporting security issues.

If you discover a vulnerability or unexpected behavior in Dropstone CLI, the Dropstone server, or any infrastructure listed in Section 07, please report it to [email protected]. We acknowledge all reports within 48 hours.

We do not currently operate a bug bounty program. Coordinated disclosure with named credit is offered for any substantive report.