- RESEARCH ATLAS / 13 LABSGo deeper into the counterexamples and experiments ↗Connected MCP, action replay, mail targets, real screens, verification, and more.
-
NEW / CONNECTED MCP PILOTOne call crosses the wire. Zero pass the rule.
An official SDK client talks to a child MCP server over stdio. A misleading hint lets an outside call reach the local handler once; an AML host rule stops the same proposal before the server. Inspect six measured cases and the local audit.
Two actual Node workers each reach a simulated callback when their claim stores are separate. A private shared claim lets one through; the other receives a replay error. Inspect four executed cases.
NEW / FALSE PERFECT-SCORE COUNTEREXAMPLE13/13 can still be wrong.
A canned adapter copies fixed answers and passes every public case. Six seeded transformations expose it at 13/19 while the project reference executes 19/19. Inspect the case-by-case evidence.
OUTSIDE IMPLEMENTER TOOLShow your own boundary result.
Run an adapter against 13 published cases plus six seeded probes and export a standalone visual report of actual verdicts and diagnostics. The published 13-case example is a project reference self-test.
Eight executed cases reveal a BCC added by host routing after an earlier check. Moving request construction before policy changes simulated provider callbacks from one to zero.
Seven executed cases compare a To-only function-tool mapper with a complete-recipient host adapter. An added BCC goes from one simulated callback to zero.
An unreviewed exact route admits a changed body. See what a digest-bound host decision changes in eight executable cases.
Inspect the counterexample ↗
+ RESEARCH ATLAS / 13 LABSFind the experiment behind the boundary ↗Search by failure mode or choose a path through the 13 labs.
+
13 labs shown
+
+
NEW / CONNECTED MCP PILOTOne call crosses the wire. Zero pass the rule.
An official SDK client talks to a child MCP server over stdio. A misleading hint lets an outside call reach the local handler once; an AML host rule stops the same proposal before the server. Inspect six measured cases and the local audit.
Two actual Node workers each reach a simulated callback when their claim stores are separate. A private shared claim lets one through; the other receives a replay error. Inspect four executed cases.
NEW / FALSE PERFECT-SCORE COUNTEREXAMPLE13/13 can still be wrong.
A canned adapter copies fixed answers and passes every public case. Six seeded transformations expose it at 13/19 while the project reference executes 19/19. Inspect the case-by-case evidence.
OUTSIDE IMPLEMENTER TOOLShow your own boundary result.
Run an adapter against 13 published cases plus six seeded probes and export a standalone visual report of actual verdicts and diagnostics. The published 13-case example is a project reference self-test.
Eight executed cases reveal a BCC added by host routing after an earlier check. Moving request construction before policy changes simulated provider callbacks from one to zero.
Seven executed cases compare a To-only function-tool mapper with a complete-recipient host adapter. An added BCC goes from one simulated callback to zero.