diff --git a/BUILT.md b/BUILT.md
index 1d2fda1..807fa9d 100644
--- a/BUILT.md
+++ b/BUILT.md
@@ -269,6 +269,44 @@ that split them would break the write check silently. The test covers both.
 Cost, measured: a 50M-iteration dependency chain over a 1024-element array runs at 0.11–0.12s checked against 0.12–0.13s
 unchecked. Indistinguishable.
 
+## Sanitizers, and why hand-written IR does not get them for free
+
+`--sanitize` builds the whole program under ASan and UBSan — the runtime's C, the generated shim, and the Flan. The last
+of those is not what passing `-fsanitize=address` to the clang run over the `.ll` gets you, and the gap is silent.
+
+**AddressSanitizer is an LLVM pass, but it instruments only functions carrying the `sanitize_address` attribute.** That
+attribute is put there by clang's C frontend. `Emit` writes `.ll` by hand, so it wrote none, so the pass walked past
+every Flan function and instrumented `flan_rt.c`. The measurement that settled it: an out-of-bounds read of a `defvar`
+array in a `--no-bounds-checks` build printed its garbage and exited 0; with an `attributes #0 = { sanitize_address }`
+group named on every `define`, the same program reports `global-buffer-overflow in flan.main`. Globals are the exception
+— the module pass redzones them whether or not any function is attributed — which is why the *shape* of a sanitized
+build looked right long before it worked.
+
+**UndefinedBehaviorSanitizer has no equivalent lever.** Its checks are not a pass: the C frontend emits branches to
+`__ubsan_handle_*` inline, and no attribute asks anything to produce them. So UBSan covers the C and nothing else, and
+`(<< 1 32)` is still unremarked under `-fsanitize=undefined`. Shift UB, alignment and the f32→i32 cast on NaN are
+therefore a compiler feature if they are wanted — checks emitted from `Emit` behind the flag, the same shape the bounds
+checks already have — and not a flag away. `test_sanitize` pins both halves with controls: one program that must report
+and one that must not, so either fact changing is a test failure rather than a discovery.
+
+`-fno-sanitize=signed-integer-overflow` is the only exclusion, because wrapping is what this language's arithmetic
+means and without it every program trips on its first `+`.
+
+**The flag deliberately does not force `-O0`,** unlike `--debug`, whose reason (mem2reg deletes the alloca a
+`llvm.dbg.declare` describes) does not apply. The optimiser is half of what is being measured, and `bounds.flan` proves
+it: with checks off, its read past the end of a string constant is reported at `-O0` and silent at `-O2`, because an
+out-of-bounds `inbounds` getelementptr into a constant is poison and LLVM folds the load away. The program then prints a
+wrong answer instead of touching memory. Same family as `(<< 1 32)` compiling to a bare `retq`.
+
+**What ASan covers of the bounds checks' job, since `--sanitize --no-bounds-checks` is the run that asks.** Three of
+`bounds.flan`'s six deliberate out-of-bounds cases are caught. A negative index into a global is not — a global gets a
+right redzone and nothing to its left. Neither is a reversed slice, which computes a negative length and then reads
+nothing at all. ASan sees out-of-*object* access, not out-of-subobject, so a slice into the middle of a larger array can
+overrun its logical bounds without crossing a redzone at all. It is a second net, not a replacement.
+
+The sweep lives on its own dune alias rather than on `dune test`: a sanitized program links to a statically linked 1.8MB
+binary, and twenty-eight of them twice over is minutes against the suite's seconds.
+
 ## Why there is no interpreter
 
 Open decision #7 is settled: **the compiled path is the only backend.** Both arguments for a permanent interpreter had
diff --git a/NEXT.md b/NEXT.md
index eea97ea..bbf519b 100644
--- a/NEXT.md
+++ b/NEXT.md
@@ -62,39 +62,84 @@ the commit that made it.
    `-2851001042534928384` — the same 64 bits, printed unsigned now that `hash-grid`'s `u64` no longer goes through an
    `(i64 …)` cast — and a trap column shifted because the call it names got shorter.
 
-### Queued — the runtime under a sanitizer
+### Landed — the runtime under a sanitizer
 
-Nothing has ever run under ASan or UBSan on the test path. `grep -i 'sanitize\|asan\|valgrind'` over `lib/ bin/ test/
-runtime/ vendor/` returns nothing at all. This is not a big project — the whole C surface is about 1,260 lines
-(`flan_rt.c` 394, `flan_dev.c` 219, `flan_agent.c` 647) plus what `shim.ml` generates.
+`--sanitize` is a build flag beside `--debug`, and `dune build --root . @sanitize` is the sweep: twenty-eight programs
+built twice, plain and sanitized, compared on output and exit status. It is not on `dune test` — a sanitized program is
+a statically linked 1.8MB binary and the sweep is about nine minutes against the suite's seconds.
 
-**The bugs are not in allocation.** There are four `malloc`/`calloc`/`strdup` sites in the entire runtime and every one
-is allocate-once-never-free by design — `rt_args` says so in its own comment — so LeakSanitizer would mostly produce
-suppressions. The risk is **fixed static buffers with bounds arithmetic**, and the ones with no coverage are already
-written down: `scratch[SCRATCH]` and `escaped[ESCAPE_MAX]` in `flan_rt.c`, the 4K result cap, the registry overflow
-guard, and `SNAP_MAX`/`SNAP_NAMES` from the restart snapshot.
+**The flag was the easy half. What the sanitizers reach is not symmetric, and this is the thing to remember:**
 
-There is evidence the sweep pays. Checking `flan_escape_bytes` by hand against lengths 0–1300 under ASan with a red
-zone found the guard correct but its comment understating its own reserve by four bytes — worst output 1021 into 1024.
-That was one buffer, found by looking.
+- **ASan reaches Flan code only because `Emit` now says so.** AddressSanitizer is an LLVM pass but it instruments only
+  functions carrying the `sanitize_address` attribute, which clang's *C frontend* adds and which nothing adds to IR
+  written by hand. Before that attribute, `clang -fsanitize=address` over the `.ll` instrumented `flan_rt.c` and not one
+  instruction of Flan: a read off the end of a `defvar` array in a `--no-bounds-checks` build printed its garbage and
+  exited 0. With it, the same program reports `global-buffer-overflow in flan.main`. Globals get their redzone from the
+  module pass either way. `test_sanitize`'s first control is that exact program, so the day the attribute stops being
+  emitted the sweep fails loudly instead of going quietly green.
+- **UBSan reaches the runtime's C and nothing else, and there is no lever for it.** Its checks are branches to
+  `__ubsan_handle_*` that clang's C frontend emits; no attribute asks a pass to produce them. `(<< 1 32)` under
+  `-fsanitize=undefined` is still unremarked. **Item 3 of the old plan — shift UB, alignment, the f32→i32 cast on NaN —
+  is therefore not answerable this way, and is still open.** The choices are to emit the checks from `Emit` behind the
+  flag (a compiler feature, and the same shape the bounds checks already have) or to leave those to the checker. The
+  second control in `test_sanitize` is a program with unambiguous shift UB that is expected *not* to be caught, so if a
+  future clang changes this, the sweep says so.
 
-The shape:
+`-fno-sanitize=signed-integer-overflow` is the one exclusion, because wrapping is what this language's arithmetic means.
+Nothing else is excluded. The flag does **not** force `-O0` the way `--debug` does; see below for why that earned its
+keep. `--sanitize --debug` together works and gives `-O0` plus source lines in the report.
 
-1. A `--sanitize` flag beside `--debug` in `lib/build.ml`, reaching both the clang run over the `.ll` and the runtime's
-   own C.
-2. Run the existing corpus under it. `test/programs/` is about forty programs with pinned output — a second pass over
-   them is the cheapest coverage available here, and needs no new test written.
-3. **UBSan is worth more than ASan**, with one exclusion that is not optional: arithmetic wraps by design, so
-   `-fno-sanitize=signed-integer-overflow` or every program trips on the first `+`. What is left is real — shift UB
-   (`(<< 1 32)` compiled to a bare `retq` at -O2, see Sharp edges), alignment, and the f32→i32 cast on NaN or an
-   infinity that `floor-f32` guards by hand and nothing else does.
-4. The variant worth its own run: **ASan with `--no-bounds-checks`**. That asks whether the bounds checks are the only
-   thing between the language and corruption, which the checked build cannot ask.
+**The checked sweep is clean: 28 programs, no ASan report, no UBSan report, no divergence from the unsanitized run.**
+What that does and does not prove:
 
-Two limits, so nobody is surprised. raylib and libm are not instrumented, so the windowed examples are noise and the
-headless corpus is the target — `sand-headless`, `values`, `machine`, `virtual-controls-headless`. And ASan does not
-see uninitialised reads, which is where `zeroed` and struct padding live; that wants Valgrind as a slower second pass,
-because MSan needs every dependency instrumented and raylib settles that.
+- It **does** cover `escaped[ESCAPE_MAX]` at its boundary. `println.flan` prints a 1100-character string inside a struct
+  on purpose; the escaped output is 1019 bytes against a hand-verified worst case of 1021 into 1024. That buffer is
+  genuinely exercised.
+- It says **nothing** about `scratch[SCRATCH]`, whose longest reachable output is 20 characters (`%lld`) into 64.
+- It says **nothing at all** about the 4K result cap, the dev registry overflow guard, `SNAP_MAX`/`SNAP_NAMES` or
+  `condition_name[128]`. Those are on the dev and agent paths, which need a daemon and a socket and are not in the
+  sweep's corpus. They were read and the guards are correct; that is reading, not evidence.
+
+**Two defects, both found by reading under the sanitizer rather than by the sanitizer, both fixed.**
+
+- `flan_bytes_to_i64` and `flan_bytes_to_f64` clamped with `(size_t)n < sizeof buf - 1`, and a slice's length is signed:
+  `(slice s 2 1)` is −1, `(size_t)(-1)` is not less than 511, so the `memcpy` read 63 or 511 bytes out of a five-byte
+  string constant. Every other `(ptr, len)` entry point in the runtime already folded a negative length to zero, so this
+  was two exceptions and not a missing convention. A checked build traps on the reversed slice first, which is why it
+  took `--no-bounds-checks` to show. Regression case in `test_sanitize`.
+- The three `snprintf` shims published `scratch` as a slice using snprintf's return, which is what it *would* have
+  written. No format here can reach 64, so it could not fire; it is clamped anyway, because the distance between "cannot
+  fire" and "hands out a length past the end of a static buffer" is one format string.
+
+**Left, deliberately.** `break_loop` prints the condition name with `"%.*s"` and a negative `namelen` would drop the
+precision and read to a NUL that a Flan string does not have — compiler-emitted and not reachable from source, so it is
+recorded rather than fixed. `flan_bytes_to_f64` silently parses only the first 511 bytes, which changes a value rather
+than corrupting memory.
+
+**ASan with `--no-bounds-checks`, the variant that asks whether the checks are the only thing holding the line. They
+mostly are.** Of `bounds.flan`'s six deliberate out-of-bounds cases ASan catches three:
+
+| case | what it does | ASan |
+|---|---|---|
+| `3`, `7` | read and write past a 3-element global | caught |
+| `4` | slice `hi` past the end of a string constant | caught |
+| `-1` | negative index into a global | **silent** — a global gets a right redzone and nothing on its left |
+| `9` | read past the end of a string constant | **silent at -O2, caught at -O0** |
+| `2` | reversed slice, negative length | silent — nothing is read at all |
+
+Three of six is an upper bound on ASan's cover, not a measured fraction of the risk: ASan sees out-of-*object* access,
+not out-of-subobject, so a slice into the middle of a larger array can overrun its logical bounds without crossing a
+redzone, and the `-1` case is one instance of that class. **ASan is not a substitute for the bounds checks.**
+
+Case `9` is why `--sanitize` does not force `-O0`. At `-O2` the access does not happen: an out-of-bounds `inbounds`
+getelementptr into a constant is poison, LLVM folds the load away, and the program prints a wrong answer instead of
+touching anything. Same family as `(<< 1 32)` compiling to a bare `retq` under Sharp edges. Running the sweep at both
+levels is cheap and a divergence between them is itself the finding.
+
+**Still not done.** Valgrind over the headless corpus, for the uninitialised reads ASan cannot see — where `zeroed` and
+struct padding live. MSan is out because it needs every dependency instrumented and raylib settles that. The reload path
+is uninstrumented whatever the flag says: a redefinition module is built by `llc` and `ld`, not by clang, so nothing
+puts a pass over it.
 
 ### Managed classes are planned. Do not start them.
 
diff --git a/test/test_sanitize.ml b/test/test_sanitize.ml
index 770e3e6..a68ab28 100644
--- a/test/test_sanitize.ml
+++ b/test/test_sanitize.ml
@@ -235,10 +235,12 @@ let unchecked_controls () =
 let () =
   match Sys.command "command -v clang > /dev/null 2>&1" with
   | 0 ->
-    (* A read five past the end of a four-element global. Five and not ten:
-       ASan's redzone on a global this small is 16 bytes, so an index far
-       enough past the end lands beyond the redzone and is not seen — which is
-       itself worth knowing about what this tool can do. *)
+    (* A read one past the end of a four-element global. Index 5 and not 9,
+       and the difference is worth knowing: ASan registers this array as
+       "16 bytes in a 32-byte slot", so the poisoned redzone is bytes 16..31.
+       Index 5 is byte 20 and is caught; index 9 is byte 36, past the
+       registration entirely, and is silent. Overrunning a small object by
+       enough lands back in ordinary memory. *)
     control ~expect_report:true "flan-san-ctl-oob"
       ~why:"This sweep is not instrumenting Flan code at all — check that \
             Emit still puts every define in the sanitize_address attribute \