Building openly, with the reasons.

In active development. Items here have an engineer assigned and a target ship date — not a promise, but our best estimate of when this lands.

• First customer pilot — synthetic clinical-trial workload

  Building

  A controller-side pilot running synthetic German clinical-trial data through the full split-knowledge pipeline. Real receipts, real witness signatures, real verifier flow — synthetic data, by design.

• Sanitizer canary tokens

  Building

  Pre-prod safety: inject known-fake PII tokens at a sampled rate, alert if any reach Sandbox B. Required for measured-recall claims in the DPO evidence package.

• Sanitizer adversarial second-pass

  Building

  A second sanitizer model deliberately tries to recover redacted entities. Failure to recover is the property we want; success means the manifest gets revised before the AI sees the prompt.

• TypeScript verifier SDK (beta)

  Designed

  ~50 lines of caller code to validate a Veil receipt without any roundtrip to our infrastructure. Beta this quarter; offline by design.

• DPIA auto-generator (v1)

  Designed

  Generates a customer-specific Data Protection Impact Assessment section from the receipts produced during a 30-day measurement window. v1 covers one workflow shape; full packaging follows.

Designed, prioritized, and queued. Order can shift if a pilot uncovers something more urgent — that is the whole point of going pilot-first.

• Python + Go verifier SDKs

  Designed

  Same surface as the TypeScript verifier; same offline guarantees. Required for customers running non-Node infrastructure.

• Async-job certificate UX

  Designed

  For long-running model calls (streaming responses, multi-step agent runs) the receipt is finalized at completion rather than per-token. Sidesteps the streaming-DLP trap cleanly.

• Agent identity (JWKS, scoped tokens)

  Designed

  For multi-step agent workflows: per-agent identity with key rotation, scoped tokens, explicit trust tiers. Required for the agent product to scale beyond pilot.

• EU AI Act Annex IV evidence bundle

  Designed

  Bundle of receipts, sanitizer manifests, and witness logs that maps directly to Annex IV documentation requirements. Pulled in during pilot scoping when a customer asks.

• Hardware-attested Sandbox B tier

  Researching

  Confidential-computing-attested processing for the highest-tier customers — a separate pricing line. Early conversations with a confidential-computing vendor.

• MCP control plane

  Designed

  Server inventory, approval state, kill-switch, analytics. Treats every Model Context Protocol server like a tenant and gives admins one panel to manage them.

Customer-triggered or strategic. We won't build these on speculation; they unlock when a customer's contract or compliance milestone makes them load-bearing.

• Binder PDF assembly endpoint

  Designed

  One endpoint that compiles DPIA + AVV + architecture review + Annex IV bundle into a single signed PDF — the "deployment binder" customers can hand to legal.

• AVV / DPA template generator

  Researching

  German Auftragsverarbeitungsvertrag plus English Data Processing Agreement filled from tenant configuration. Lawyer-reviewed templates, not legal advice.

• Multi-tenant key rotation

  Designed

  Per-customer signing keys rotated on a customer-controlled schedule. Required for some banking workloads; not required for the first pilots.

• Format-preserving tokenization

  Researching

  Tokens that look like the original (IBAN-shaped, phone-shaped) so legacy systems pass downstream validation. Customer-triggered.

• Markov anomaly detection on re-linkage flow

  Researching

  Models the legitimate distribution of the governance flow; flags low-probability sequences (skipped reviews, biased approvers). Cheap to build, defensible by design.

• DPIA auto-generator (full packaging)

  Designed

  Beyond v1 — multi-workflow, multi-tenant, customer-edit-able output that compiles to a binder PDF. Customer-triggered, not always-on.