JOURNEY_SENTINEL_STATUS
What we can prove, and what we cannot.
sdlc.cc is pre-release. It is not generally available, there is no completed customer deployment, and there is no design partner yet. This page says exactly where the line is, because a privacy product that overstates itself is worth nothing.
Proven — verifiable in the source
These are properties of the code. A prospective partner can read or run them.
| Claim | How you check it |
|---|---|
| Every governed model-write route shares one admission path: authenticate, resolve policy, quota, transform, dispatch once, record | Enforcement tests, roughly 310 of them |
| Model paths without a governed handler are refused, not forwarded unprotected | Router conformance tests |
| Streaming emits no provider bytes until the response is fully validated | Buffered response tests |
| No database column can hold a prompt or a completion | Read the migrations |
| Clients fail closed and never render a gateway response body | Client surface tests |
| Packages carry a checksum and commit provenance | Packaging gate |
The hosted gateway answers its health check and rejects unauthenticated requests. That shows it runs, routes and authenticates. It shows nothing about detection.
Not proven — we will not claim these
| Item | Why | What would settle it |
|---|---|---|
| Detection accuracy for any data class | A per-class corpus, a stated bar and a tested runner are now published, but no detector fixtures and no per-class regression run against a gateway have produced a result — so no class has a measured recall or false-positive rate | A published report from a real run, recorded with the commit and image digest |
| Anyone using it in production | No design partner, no pilot cohort | A named partner with dated acceptance |
| Self-serve purchase | Deliberately sales-assisted: public CTAs request operator activation; Lemon is optional and does not provision keys | Signed webhook fulfilment if self-serve is chosen later |
| Availability or an SLA | Single instance on a beta platform | Multi-instance topology with tested recovery |
| Browser store distribution | Not submitted or approved | Publisher confirmation |
| Pricing | Figures are illustrative, not validated with buyers | Buyer interviews |
| Enterprise identity | One shared admin credential; no SSO, SCIM or roles | The identity work on the roadmap |
The gap that matters
Everything in the proven column describes a pipeline: content is routed, policy is applied, failures close, evidence is recorded, prompts are never stored. That pipeline is well built and well tested.
What is not published is whether the step in the middle reliably finds the sensitive data. There is now a per-class corpus and a stated bar — 100% recall on checksum-validated classes, 95% on structured ones, and zero tolerated false positives — so you can read what would be measured and disagree with it before any number exists. What there is not yet is a result.
Until there is, the honest framing is that detection quality gets measured during your pilot, in observe mode, on your data, before anything is enforced.