diff --git a/docs/research/2026-08-07-s6-perfectclip-containment-contract.md b/docs/research/2026-08-07-s6-perfectclip-containment-contract.md
new file mode 100644
index 00000000..c928ef45
--- /dev/null
+++ b/docs/research/2026-08-07-s6-perfectclip-containment-contract.md
@@ -0,0 +1,66 @@
+# S6 contract (Campaign S) — contain the undecodable PerfectClip TOI tails (AP-83 / AP-91)
+
+**Date:** 2026-08-07. **Implementer:** one Sonnet agent, after S4b lands.
+**Review:** one Opus pass (small slice). **No visual gate** — fully automated.
+
+## What this is, and is not
+
+AP-83 and AP-91 record that the PerfectClip time-of-impact tails in
+`CylCollideWithPoint` and `SphereCollideWithPoint`/`FindSphereTimeOfCollision`
+(TransitionTypes.cs) were decoded via ACE because the retail x87 sequences do
+not decompile. **There is no retail text to port — this slice does NOT "fix"
+them.** The deliverable is proof about reachability plus a loud guard, so the
+rows become auditable instead of aspirational (the AP-22 resolution shape).
+
+## Premise correction, pinned at scoping — the rows' "no mover sets
+PerfectClip" is FALSE as stated
+
+`PhysicsCameraCollisionProbe.cs:69` sets `ObjectInfoState.PerfectClip` on the
+CAMERA mover (faithful to retail's camera flags). So containment must be
+proven as: **the camera's resolve can never reach the Cyl/Sphere PerfectClip
+TOI tails.** Candidate arguments to verify (do not assume any):
+1. The viewer exemption block in `FindObjCollisions`/`CollisionExemption`
+   (viewer-vs-creature skip, `CollisionExemption.cs:83-85` region) — does it
+   skip ALL shadow objects for a viewer mover, or only creatures? A static
+   prop with CylSpheres in the camera's sweep would reach the tail if not.
+2. Whether the camera resolve enumerates shadow objects at all (retail's
+   `update_viewer` sweep may be environment-only — check our port's call
+   shape at the camera probe's ResolveWithTransition arguments).
+If BOTH arguments fail — the camera genuinely can reach a Cyl/Sphere TOI
+tail — that is a FINDING, not a failure: the ACE-derived math is then LIVE
+production code for the camera, and the rows get rewritten to say so with
+the camera named as the reachable population (severity: camera-feel only).
+
+## Deliverables
+
+1. **Reachability proof or refutation**, written into a short doc: for each
+   of the two tails, the exact call chain from every PerfectClip-setting
+   mover (grep-proof of the full setter population first — today it is the
+   camera probe alone, verify), and where each chain is cut (file:line), or
+   not.
+2. **The loud guard:** at the head of each TOI tail, a debug-only assertion
+   (or diagnostics counter + one-shot log, matching house style for
+   invariants) that fires if the tail executes with a mover whose reachability
+   was proven impossible — so a future flag change cannot silently start
+   executing ACE-derived math nobody re-verified. If the camera IS reachable,
+   the guard instead records the tail as camera-live (counter, not assert).
+3. **Tests:** one per tail driving the production resolve with a
+   PerfectClip mover against a Cyl/Sphere-shaped shadow entity, asserting
+   whichever reachability the proof established (reached → the guard counter
+   increments and the ACE-derived result is pinned as a golden; unreached →
+   the exemption cuts it and the guard stays silent). Sabotage: disable the
+   exemption that cuts the chain and watch the reachability flip.
+4. Register rewrite text for AP-83/AP-91 handed to the session lead
+   (rows stay ACTIVE either way — the math remains ACE-derived — but their
+   population claim becomes measured).
+
+## Scope — OUT
+
+Any change to the TOI math itself. The camera's flag set (retail-faithful).
+Register/ISSUES edits (session lead).
+
+## Acceptance
+
+Scoped build/test green (Core + Core.Tests; clean-room for the verdict run);
+sabotage reported verbatim; nothing committed. Absolute paths, no subagents,
+contradictions → STOP.