diff --git a/docs/VERIFICATION-STANCE.adoc b/docs/VERIFICATION-STANCE.adoc
new file mode 100644
index 0000000..2678555
--- /dev/null
+++ b/docs/VERIFICATION-STANCE.adoc
@@ -0,0 +1,73 @@
+= VCL-total Verification Stance
+:toc:
+:sectnums:
+
+// SPDX-License-Identifier: PMPL-1.0-or-later
+// Copyright (c) 2026 Jonathan D.A. Jewell (hyperpolymath) <j.d.a.jewell@open.ac.uk>
+
+WARNING: vcl-ut is the interface between *hypatia* and *verisim*. Both sides
+trust safety claims made here. This document states *precisely* what is and is
+not currently proven, so that neither consumer over-trusts the boundary. It
+takes precedence over any "verified" / "type-safe" / "proven" phrasing in
+`README.adoc`, `PROOF-NEEDS.md`, or module doc-comments.
+
+== Summary
+
+The 10-level safety story is, at present, *architecturally typed but not
+end-to-end proven*. The dependent-type apparatus exists but is *not connected*
+to the runtime checker or the FFI boundary that hypatia/verisim consume.
+
+== What IS genuinely established
+
+* `src/interface/abi/Types.idr` — `SafetyLevel`, `VclTotalError`, `QueryMode`
+  are real ADTs with total `DecEq` instances and total int<->tag round-trip
+  functions. The `Verify` namespace proves tag ranges with `So`. These are
+  sound, non-vacuous proofs.
+* `QueryHandle` is a genuine opaque type with an `So (ptr /= 0)` non-null
+  witness; it cannot be constructed null through the public API.
+* `src/core/Levels.idr` `SafetyCertificate` / `CanWeaken` define a real
+  subsumption lattice *at the type level*.
+
+== What is NOT proven (the honest gaps)
+
+. *The certificate is disconnected from the checker.* `src/core/Checker.idr`
+  decides every level with `Bool`-returning `checkLevelN` functions and never
+  constructs a `SafetyCertificate`. A query reported "Level 10" carries *no*
+  `Levels.idr` proof object. The dependent types are decorative w.r.t. the
+  actual decision procedure.
+. *Several level predicates are vacuous.* `NoRawUserInput` has the single
+  constructor `AllParameterised : NoRawUserInput stmt` (provable for *any*
+  statement, including injection-bearing ones); `AllNullableFieldsGuarded`'s
+  `GuardedNull` accepts *any* expression; `AllSelectItemsTyped`'s `ConsTyped`
+  requires no actual per-item typing evidence. As written, Levels 3, 4 and 5
+  prove nothing about the property they name.
+. *The FFI boundary is stringly/int-typed.* `src/interface/abi/Foreign.idr`
+  returns a bare `Bits32` error tag; the `parse` wrapper is an explicit
+  placeholder (returns `Left InternalError` on success). hypatia/verisim
+  receive an *asserted* level integer, not a transported proof.
+. *Layout/size proofs are vacuous.* `HasSize` / `HasAlignment` constructors
+  accept any `n`; `queryPlanHeaderSize : ... HasSize QueryPlanHeader 24` type-
+  checks regardless of the real layout. The 24-byte/8-align claim is unverified.
+
+== Consumer guidance
+
+* *hypatia / verisim*: treat the FFI level integer as an *advisory claim*, not
+  a proof carrier. Do not weaken your own checks on the strength of a VCL-total
+  level tag until gap (1) and (3) are closed.
+* Do not cite vcl-ut as "formally verified query safety" in downstream docs.
+
+== Closing the gaps
+
+The high-value, low-risk path (per `PROOF-NEEDS.md`): make `checkLevelN` return
+`Dec (Ln_... )` instead of `Bool`, give the vacuous predicates real
+constructors carrying evidence, and have the checker emit a `SafetyCertificate`
+that the Zig FFI transports as an opaque proof token. Tracked under
+hyperpolymath/standards#124.
+
+== Provenance
+
+Audit 2026-05-18 (corpus grep + read; no proof-escape symbols
+`believe_me`/`assert_total`/`postulate`/`sorry` are *applied* anywhere in
+`src/**` — the only match is a self-attesting comment in `Composition.idr`).
+The risk here is *not* proof escapes; it is *absent* proof masquerading as
+present proof at a trusted interface.