diff --git a/packages/README.md b/packages/README.md
index 3b5067fd..92e5a063 100644
--- a/packages/README.md
+++ b/packages/README.md
@@ -13,6 +13,7 @@ section. The rules are in [ludic.base](ludic.base/README.md).
 | | |
 | --- | --- |
 | [ludic.base](ludic.base/README.md) | the shared vocabulary: Tick and phases, `Queue<T>`, rng streams, the save tree, the system runner (`Systems`, an open registry) |
+| [ludic.anim](ludic.anim/README.md) | glTF animation clips sampled and cross-faded onto a render3d Skin (uses `ludic_render3d`) |
 | [ludic.clock](ludic.clock/README.md) | the hour, the day, the moon and a calendar of seasons |
 | [ludic.crafting](ludic.crafting/README.md) | recipes: what goes in and out, where and how long, making, refunding, hold-to-make |
 | [ludic.effects](ludic.effects/README.md) | timed modifiers that run down in game time (a meal's warmth, a drink's legs) |
diff --git a/packages/ludic.anim/README.md b/packages/ludic.anim/README.md
new file mode 100644
index 00000000..8cf0127e
--- /dev/null
+++ b/packages/ludic.anim/README.md
@@ -0,0 +1,49 @@
+# ludic.anim
+
+The animation clips an animator keyed into a glTF, played on a render3d `Skin` and cross-faded two
+at a time. `ludic.render3d` parses no animations and offers a pose api (`skin_set_quat`,
+`skin_set_offset`, `skin_pose`) rather than a player, so the sampler lives here. Uses
+`ludic_render3d` and nothing else: it reads the document `gltf_load` just parsed and writes a
+Skin's pose. It asks the game nothing, so it has no port.
+
+```ludic
+import "ludic.anim"
+let m = gltf_load("assets/kit/hiker2", "hiker_m.gltf", "hiker_m")
+anim_read_doc("assets/kit/hiker2")                 # straight after the load, before the next
+let walk = anim_find("hiker_m_walk")
+anim_play(sk, walk, t, idle, t2, 0.3)              # 70% walk, 30% idle, into the joint matrices
+```
+
+## API
+
+| | |
+| --- | --- |
+| `Clip { name, dur, n, cnode, cpath, ckeys, coff, cvoff, times, vals }`, `AC_ROT`, `AC_POS` | a clip, its channels flattened: a channel's keys are `times[coff..coff+ckeys]`, its values 4 (rotation) or 3 (translation) per key from `cvoff` |
+| `anim_read_doc(dir) -> int` | every rotation and translation channel of every animation in `gltf_doc`, reading the `.bin` beside it in `dir`; how many clips were new |
+| `anim_find(name) -> Clip`, `anim_count()`, `anim_add(c) -> bool`, `anim_clear()` | the table, by name; a name already held is skipped on read and refused on add |
+| `anim_play(sk, a, ta, b, tb, blend) -> bool` | `a` at `ta`, cross-faded toward `b` at `tb` by `blend` (0 all a, 1 all b; `b` may be null), folded into the joint matrices |
+| `anim_mix(c, sk, t, w, rot, pos, hit)`, `anim_to_pose(sk, rot, pos, hit)`, `anim_key_at(c, ch, t)` | the pieces `anim_play` is made of, for a caller that mixes more than two |
+| `anim_set_positions(on)`, `anim_positions()` | keyed translations applied or not (off: baking keys location on every bone, and rounding down an eighteen-bone chain is a collapsed animal) |
+| `anim_first_key(c) -> float` | the earliest key; a read warns when it is not zero |
+
+## Rules it keeps
+
+- **Clips are named per rig.** A channel addresses its FILE's node indices, two files number their
+  nodes differently, and a name already held is skipped - so one shared name would hand the
+  second body the first one's skeleton. Name them `hiker_m_walk` / `hiker_f_walk`, `bear_walk`.
+- **Read straight after the load.** `gltf_doc` is whichever file was parsed last, and `gltf_load`
+  closes the `.bin`, which `anim_read_doc` reopens (a read through the closed handle is zeros).
+- **Key from frame zero.** A clip wraps on its duration and holds below its first key, so a clip
+  keyed from frame 1 holds its opening pose for 1/30 s every cycle. `anim_read_doc` prints it.
+- **The math.** `skin_pose` takes a delta D in the model frame, `local = (G_p^-1 D G_p) R_rest`,
+  and a channel gives the absolute local L, so D = G_p (L R_rest^-1) G_p^-1.
+- The scratch is made once and grown: a call is per actor per frame.
+
+## Tests
+
+```bash
+ludic test packages/ludic.anim
+```
+
+A two-bone skin made from a glTF document with no GPU: sampling, wrapping, the cross-fade,
+translations off by default, and clips read out of a `.bin` written to the temp directory.
diff --git a/packages/ludic.anim/clip.ludic b/packages/ludic.anim/clip.ludic
new file mode 100644
index 00000000..27c37b98
--- /dev/null
+++ b/packages/ludic.anim/clip.ludic
@@ -0,0 +1,66 @@
+# clip.ludic - a clip: its channels flattened into one run of key times and one of values, and the
+# table of every clip read so far, found by name
+export const AC_ROT: int = 0
+export const AC_POS: int = 1
+
+export property Clip {
+  name:  string = "",
+  dur:   float = 0.0,              # seconds: the last key of any channel
+  n:     int = 0,                  # channels
+  cnode: words,                    # target node per channel (the FILE's node index)
+  cpath: words,                    # AC_ROT or AC_POS
+  ckeys: words,                    # keys per channel
+  coff:  words,                    # where this channel starts in `times`
+  cvoff: words,                    # where it starts in `vals`
+  times: floats,
+  vals:  floats
+}
+
+var clips: []Clip = null
+# Keyed translations are off unless asked for: baking keys location on every bone, and a skeleton
+# does not translate its joints, so applying them turns rounding into a pose.
+var positions: bool = false
+
+function clips_init() -> void {
+  if clips == null { clips = new []Clip }
+}
+
+# the clip of that name, or null; names are per file (per body, per species), never shared
+export function anim_find(name: string) -> Clip {
+  clips_init()
+  for i in 0 .. len(clips) { if clips[i].name == name { return clips[i] } }
+  return null
+}
+
+export function anim_count() -> int {
+  clips_init()
+  return len(clips)
+}
+
+# a clip made outside a glTF (a test, a generated clip); a name already held is refused
+export function anim_add(c: Clip) -> bool {
+  if c == null or anim_find(c.name) != null { return false }
+  push(clips, c)
+  return true
+}
+
+export function anim_clear() -> void { clips = new []Clip }
+export function anim_set_positions(on: bool) -> void { positions = on }
+export function anim_positions() -> bool { return positions }
+
+# Which key pair straddles time t, by halving rather than walking: a grazing clip is three
+# hundred keys long and a valley can have dozens of animals posed in a frame.
+export function anim_key_at(c: Clip, ch: int, t: float) -> int {
+  let n = c.ckeys[ch]
+  let o = c.coff[ch]
+  if n <= 1 { return 0 }
+  if t < c.times[o] { return 0 }
+  if t > c.times[o + n - 1] { return n - 2 }
+  var lo = 0
+  var hi = n - 1
+  while hi - lo > 1 {
+    let mid = (lo + hi) / 2
+    if c.times[o + mid] > t { hi = mid } else { lo = mid }
+  }
+  return lo
+}
diff --git a/packages/ludic.anim/index.ludic b/packages/ludic.anim/index.ludic
new file mode 100644
index 00000000..fb1f8300
--- /dev/null
+++ b/packages/ludic.anim/index.ludic
@@ -0,0 +1,9 @@
+# ludic.anim - playing the clips an animator keyed into a glTF, and blending two of them.
+# render3d parses no animations and offers a POSE api rather than a player, so the sampler is here
+module ludic_anim uses ludic_render3d
+numbers float
+import "ludic.render3d/r3d.ludic"
+import "clip.ludic"
+import "read.ludic"
+import "mix.ludic"
+import "pose.ludic"
diff --git a/packages/ludic.anim/mix.ludic b/packages/ludic.anim/mix.ludic
new file mode 100644
index 00000000..825c12d7
--- /dev/null
+++ b/packages/ludic.anim/mix.ludic
@@ -0,0 +1,69 @@
+# mix.ludic - sample one clip at a time into per-node buffers, mixed in at a weight. `rot` is 4 per
+# node, `pos` 3 per node, and `hit` says which nodes have received a value (bit 0 a rotation, bit 1
+# a translation), so a node the clips never mention stays at rest
+var q_a: floats = null
+var q_b: floats = null
+var q_cur: floats = null
+
+export function anim_mix(c: Clip, sk: Skin, t: float, w: float, rot: floats, pos: floats, hit: words) -> void {
+  if c == null or sk == null { return }
+  if q_a == null {
+    q_a = floats(4)
+    q_b = floats(4)
+    q_cur = floats(4)
+  }
+  var tt = t
+  if c.dur > 0.0 { tt = t % c.dur }
+  for ch in 0 .. c.n {
+    let node = c.cnode[ch]
+    if node < 0 or node >= sk.n_nodes { continue }
+    let n = c.ckeys[ch]
+    let o = c.coff[ch]
+    let i0 = anim_key_at(c, ch, tt)
+    var i1 = i0 + 1
+    if i1 > n - 1 { i1 = n - 1 }
+    var f = 0.0
+    if i1 > i0 {
+      let span = c.times[o + i1] - c.times[o + i0]
+      if span > 0.000001 { f = Math.clamp((tt - c.times[o + i0]) / span, 0.0, 1.0) }
+    }
+    if c.cpath[ch] == AC_ROT { mix_rot(c, ch, node, i0, i1, f, w, rot, hit) } else { mix_pos(c, ch, node, i0, i1, f, w, pos, hit) }
+  }
+}
+
+# Across clips the first contributor lands and the rest are nlerped onto it by their share, which
+# is what makes a cross-fade rather than a jump. Test the ROTATION bit, not the whole word: glTF
+# writes a node's translation before its rotation, and `hit == 0` then skipped every rotation.
+function mix_rot(c: Clip, ch: int, node: int, i0: int, i1: int, f: float, w: float, rot: floats, hit: words) -> void {
+  let vo = c.cvoff[ch]
+  for e in 0 .. 4 {
+    q_a[e] = c.vals[vo + i0 * 4 + e]
+    q_b[e] = c.vals[vo + i1 * 4 + e]
+  }
+  q_nlerp(q_a, q_a, q_b, f)
+  if (hit[node] & 1) == 0 {
+    q_store(rot, node, q_a)
+    hit[node] = hit[node] | 1
+  } else {
+    q_load(q_cur, rot, node)
+    q_nlerp(q_cur, q_cur, q_a, w)
+    q_store(rot, node, q_cur)
+  }
+}
+
+function mix_pos(c: Clip, ch: int, node: int, i0: int, i1: int, f: float, w: float, pos: floats, hit: words) -> void {
+  let vo = c.cvoff[ch]
+  let ax = Math.lerp(c.vals[vo + i0 * 3], c.vals[vo + i1 * 3], f)
+  let ay = Math.lerp(c.vals[vo + i0 * 3 + 1], c.vals[vo + i1 * 3 + 1], f)
+  let az = Math.lerp(c.vals[vo + i0 * 3 + 2], c.vals[vo + i1 * 3 + 2], f)
+  if (hit[node] & 2) == 0 {
+    pos[node * 3] = ax
+    pos[node * 3 + 1] = ay
+    pos[node * 3 + 2] = az
+    hit[node] = hit[node] | 2
+  } else {
+    pos[node * 3] = Math.lerp(pos[node * 3], ax, w)
+    pos[node * 3 + 1] = Math.lerp(pos[node * 3 + 1], ay, w)
+    pos[node * 3 + 2] = Math.lerp(pos[node * 3 + 2], az, w)
+  }
+}
diff --git a/packages/ludic.anim/package.ludic b/packages/ludic.anim/package.ludic
new file mode 100644
index 00000000..110b00c5
--- /dev/null
+++ b/packages/ludic.anim/package.ludic
@@ -0,0 +1,6 @@
+# ludic.anim - the animation clips that ride inside a glTF, read out of the document render3d
+# just parsed, sampled at a time and cross-faded, and folded into a Skin's pose. Uses
+# ludic.render3d (the Skin, the glTF document and the quaternions). See README.md.
+package "ludic.anim"
+version "0.1.0"
+kind source
diff --git a/packages/ludic.anim/pose.ludic b/packages/ludic.anim/pose.ludic
new file mode 100644
index 00000000..a96f6c92
--- /dev/null
+++ b/packages/ludic.anim/pose.ludic
@@ -0,0 +1,78 @@
+# pose.ludic - the sampled LOCAL transforms turned into the model-frame deltas `skin_pose` wants.
+# skin_pose folds a delta D in as local = (G_p^-1 . D . G_p) . R_rest, and a channel gives the
+# absolute local L, so D = G_p . (L . R_rest^-1) . G_p^-1 (a root: D = L . R_rest^-1)
+var parts: []floats = null
+var rot_s: floats = null
+var pos_s: floats = null
+var hit_s: words = null
+var cap: int = 0
+
+function part(i: int) -> floats {
+  if parts == null {
+    parts = new []floats
+    for k in 0 .. 8 { push(parts, floats(4)) }
+  }
+  return parts[i]
+}
+
+export function anim_to_pose(sk: Skin, rot: floats, pos: floats, hit: words) -> void {
+  let l = part(0)
+  let rinv = part(1)
+  let gp = part(2)
+  let gpi = part(3)
+  let tmp = part(4)
+  let d = part(5)
+  let v = part(6)
+  for i in 0 .. sk.n_nodes {
+    let p = sk.par[i]
+    if (hit[i] & 1) != 0 {
+      q_load(l, rot, i)
+      q_load(rinv, sk.rest_r, i)
+      q_conj(rinv, rinv)
+      q_mul(tmp, l, rinv)
+      if p >= 0 {
+        q_load(gp, sk.rest_g, p)
+        q_conj(gpi, gp)
+        q_mul(d, gp, tmp)
+        q_mul(d, d, gpi)
+      } else { q_copy(d, tmp) }
+      q_normalize(d)
+      skin_set_quat(sk, i, d)
+    }
+    if (hit[i] & 2) != 0 and positions {
+      let dx = pos[i * 3] - sk.rest_t[i * 3]
+      let dy = pos[i * 3 + 1] - sk.rest_t[i * 3 + 1]
+      let dz = pos[i * 3 + 2] - sk.rest_t[i * 3 + 2]
+      if p >= 0 {
+        q_load(gp, sk.rest_g, p)
+        v3_set(v, dx, dy, dz)
+        q_rotate(v, gp, v)
+        skin_set_offset(sk, i, v[0], v[1], v[2])
+      } else { skin_set_offset(sk, i, dx, dy, dz) }
+    }
+  }
+}
+
+# The scratch is made ONCE and grown, never allocated per call: a call is per actor per frame.
+function scratch(n: int) -> void {
+  if cap >= n { return }
+  rot_s = floats(n * 4)
+  pos_s = floats(n * 3)
+  hit_s = words(n)
+  cap = n
+}
+
+# Play `a` at time `ta`, cross-faded toward `b` at `tb` by `blend` (0 is all a, 1 is all b), and
+# fold it into the joint matrices. `b` may be null. False when there is no skin or no clip.
+export function anim_play(sk: Skin, a: Clip, ta: float, b: Clip, tb: float, blend: float) -> bool {
+  if sk == null or a == null { return false }
+  let n = sk.n_nodes
+  scratch(n)
+  for i in 0 .. n { hit_s[i] = 0 }
+  skin_reset(sk)
+  anim_mix(a, sk, ta, 1.0, rot_s, pos_s, hit_s)
+  if b != null and blend > 0.001 { anim_mix(b, sk, tb, blend, rot_s, pos_s, hit_s) }
+  anim_to_pose(sk, rot_s, pos_s, hit_s)
+  skin_pose(sk)
+  return true
+}
diff --git a/packages/ludic.anim/read.ludic b/packages/ludic.anim/read.ludic
new file mode 100644
index 00000000..e2f3a518
--- /dev/null
+++ b/packages/ludic.anim/read.ludic
@@ -0,0 +1,112 @@
+# read.ludic - every animation out of the document `gltf_load` just parsed: rotation and
+# translation channels, flattened into a Clip each; a name already held is skipped
+
+# Call it straight after a load and before the next one: `gltf_doc` and `gltf_bin` point at
+# whichever file was read last, so a clip that is not taken now cannot be taken later.
+export function anim_read_doc(dir: string) -> int {
+  clips_init()
+  if gltf_doc == null { return 0 }
+  if value_has(gltf_doc, "animations") == 0 { return 0 }
+  # REOPEN THE .bin. `gltf_load` closes it before it returns, and an accessor read through a closed
+  # handle returns a buffer of the right size full of zeros: right names, right counts, no motion.
+  let buffers = value_get(gltf_doc, "buffers")
+  if value_count(buffers) == 0 { return 0 }
+  let bin_uri = value_as_str(value_get(value_at(buffers, 0), "uri"))
+  gltf_bin = file_open(dir + "/" + bin_uri, "rb")
+  if gltf_bin == null {
+    print(`anim: cannot reopen {dir}/{bin_uri}`)
+    return 0
+  }
+  let arr = value_get(gltf_doc, "animations")
+  var got = 0
+  for ai in 0 .. value_count(arr) {
+    let c = read_one(value_at(arr, ai))
+    if c != null {
+      push(clips, c)
+      got += 1
+    }
+  }
+  file_close(gltf_bin)
+  gltf_bin = null
+  return got
+}
+
+function kept_path(ch: Val) -> string {
+  let tgt = value_get(ch, "target")
+  if value_has(tgt, "node") == 0 { return "" }
+  let path = value_as_str(value_get(tgt, "path"))
+  if path != "rotation" and path != "translation" { return "" }
+  return path
+}
+
+function read_one(a: Val) -> Clip {
+  var nm = ""
+  if value_has(a, "name") != 0 { nm = value_as_str(value_get(a, "name")) }
+  if anim_find(nm) != null { return null }                 # a second rig's file, same clip
+  let chans = value_get(a, "channels")
+  let samps = value_get(a, "samplers")
+  let nc = value_count(chans)
+  # count first: the flat arrays are sized once rather than grown per channel
+  var n_keep = 0
+  var n_t = 0
+  var n_v = 0
+  for ci in 0 .. nc {
+    let ch = value_at(chans, ci)
+    let path = kept_path(ch)
+    if path == "" { continue }
+    let sp = value_at(samps, value_as_int(value_get(ch, "sampler")))
+    let cnt = gltf_accessor_count(value_as_int(value_get(sp, "input")))
+    n_keep += 1
+    n_t += cnt
+    if path == "rotation" { n_v += cnt * 4 } else { n_v += cnt * 3 }
+  }
+  if n_keep == 0 { return null }
+  let c = new Clip
+  c.name = nm
+  c.n = n_keep
+  c.cnode = words(n_keep)
+  c.cpath = words(n_keep)
+  c.ckeys = words(n_keep)
+  c.coff = words(n_keep)
+  c.cvoff = words(n_keep)
+  c.times = floats(n_t)
+  c.vals = floats(n_v)
+  var k = 0
+  var at = 0
+  var av = 0
+  for ci in 0 .. nc {
+    let ch = value_at(chans, ci)
+    let path = kept_path(ch)
+    if path == "" { continue }
+    let sp = value_at(samps, value_as_int(value_get(ch, "sampler")))
+    c.cnode[k] = value_as_int(value_get(value_get(ch, "target"), "node"))
+    if path == "rotation" { c.cpath[k] = AC_ROT } else { c.cpath[k] = AC_POS }
+    let tb = gltf_accessor_floats(value_as_int(value_get(sp, "input")))
+    let nk = gltf_count
+    c.ckeys[k] = nk
+    c.coff[k] = at
+    for i in 0 .. nk { c.times[at + i] = tb[i] }
+    if nk > 0 and c.times[at + nk - 1] > c.dur { c.dur = c.times[at + nk - 1] }
+    at += nk
+    let vb = gltf_accessor_floats(value_as_int(value_get(sp, "output")))
+    var comps = 3
+    if c.cpath[k] == AC_ROT { comps = 4 }
+    c.cvoff[k] = av
+    for i in 0 .. nk * comps { c.vals[av + i] = vb[i] }
+    av += nk * comps
+    k += 1
+  }
+  if anim_first_key(c) > 0.001 { print(`anim: {nm} starts at {anim_first_key(c)} s - key from frame zero, or every cycle holds its opening pose`) }
+  return c
+}
+
+# the earliest key of any channel: a clip keyed from frame 1 has 1/30 here, and a hitch every cycle
+export function anim_first_key(c: Clip) -> float {
+  var t = 0.0
+  for ch in 0 .. c.n {
+    if c.ckeys[ch] == 0 { continue }
+    let s = c.times[c.coff[ch]]
+    if ch == 0 or s < t { t = s }
+  }
+  return t
+}
diff --git a/packages/ludic.anim/tests/anim_test.ludic b/packages/ludic.anim/tests/anim_test.ludic
new file mode 100644
index 00000000..6cfe5c1e
--- /dev/null
+++ b/packages/ludic.anim/tests/anim_test.ludic
@@ -0,0 +1,140 @@
+# anim_test.ludic - clips on a two-bone skin made from a glTF document with no GPU: sampling,
+# looping, the cross-fade, translations off by default, and reading clips out of a .bin
+import "ludic.render3d/r3d.ludic"
+import "ludic.anim"
+program AnimTest {
+  numbers float
+  const HALF: float = 0.70710678
+  # never called: naming Input links the engine's runtime, which render3d's textures stand on
+  function links_runtime() -> bool { return Input.key_pressed(0) }
+
+  # a root and one child 1 m up, no inverse bind matrices, one skin of both
+  function skin() -> Skin {
+    gltf_doc = Json.parse("{\"nodes\": [{\"name\": \"root\", \"children\": [1]}, {\"name\": \"bone\", \"translation\": [0, 1, 0]}], \"skins\": [{\"joints\": [0, 1]}]}")
+    return skin_load(0)
+  }
+
+  # one channel on node 1: identity at 0, a quarter turn about Z at 1 s
+  function turn(name: string) -> Clip {
+    let c = new Clip
+    c.name = name
+    c.n = 1
+    c.cnode = words(1)
+    c.cpath = words(1)
+    c.ckeys = words(1)
+    c.coff = words(1)
+    c.cvoff = words(1)
+    c.cnode[0] = 1
+    c.cpath[0] = AC_ROT
+    c.ckeys[0] = 2
+    c.times = floats(2)
+    c.times[1] = 1.0
+    c.vals = floats(8)
+    c.vals[3] = 1.0
+    c.vals[6] = HALF
+    c.vals[7] = HALF
+    c.dur = 1.0
+    return c
+  }
+
+  # where the child's origin lands after posing: rest is (0, 1, 0), a quarter turn of the root would move it
+  function pose_z(sk: Skin) -> float {
+    let q = floats(4)
+    q_load(q, sk.pose_r, 1)
+    return q[2]
+  }
+
+  test "a clip is found by name, once" {
+    anim_clear()
+    expect(anim_add(turn("hiker_m_walk")))
+    expect(not anim_add(turn("hiker_m_walk")))
+    expect(anim_add(turn("hiker_f_walk")))
+    expect_eq(anim_count(), 2)
+    expect(anim_find("hiker_f_walk") != null)
+    expect(anim_find("bear_walk") == null)
+  }
+
+  test "a key pair is found by halving" {
+    let c = turn("t")
+    expect_eq(anim_key_at(c, 0, 0.5), 0)
+    expect_eq(anim_key_at(c, 0, 2.0), 0)
+    expect_eq(anim_key_at(c, 0, -1.0), 0)
+  }
+
+  test "playing samples between keys, and wraps on the duration" {
+    let sk = skin()
+    let c = turn("t")
+    expect(anim_play(sk, c, 0.0, null, 0.0, 0.0))
+    expect_near(pose_z(sk), 0.0, 0.001)
+    anim_play(sk, c, 1.0 - 0.0001, null, 0.0, 0.0)
+    expect_near(pose_z(sk), HALF, 0.01)
+    anim_play(sk, c, 0.5, null, 0.0, 0.0)
+    let mid = pose_z(sk)
+    anim_play(sk, c, 1.5, null, 0.0, 0.0)
+    expect_near(pose_z(sk), mid, 0.001)
+    expect(mid > 0.3 and mid < 0.45)
+  }
+
+  test "a cross-fade lands between the two clips" {
+    let sk = skin()
+    let a = turn("a")
+    let b = turn("b")
+    anim_play(sk, a, 0.0, b, 0.999, 0.5)
+    let z = pose_z(sk)
+    expect(z > 0.3 and z < 0.45)
+    anim_play(sk, a, 0.0, b, 0.999, 0.0)
+    expect_near(pose_z(sk), 0.0, 0.001)
+  }
+
+  test "no skin or no clip plays nothing" {
+    expect(not anim_play(null, turn("t"), 0.0, null, 0.0, 0.0))
+    expect(not anim_play(skin(), null, 0.0, null, 0.0, 0.0))
+  }
+
+  test "keyed translations are ignored until asked for" {
+    let sk = skin()
+    let c = turn("t")
+    c.cpath[0] = AC_POS
+    c.vals[0] = 0.0
+    c.vals[1] = 3.0
+    c.vals[3] = 0.0
+    c.vals[4] = 3.0
+    anim_play(sk, c, 0.0, null, 0.0, 0.0)
+    expect_near(sk.pose_t[4], 0.0, 0.0001)
+    anim_set_positions(true)
+    anim_play(sk, c, 0.0, null, 0.0, 0.0)
+    expect_near(sk.pose_t[4], 2.0, 0.0001)
+  }
+
+  function put_f(b: []byte, at: int, x: float) -> void {
+    let v = float_bits(x)
+    b[at] = v & 255
+    b[at + 1] = (v >> 8) & 255
+    b[at + 2] = (v >> 16) & 255
+    b[at + 3] = (v >> 24) & 255
+  }
+
+  test "clips are read out of a glTF's .bin, and a late first key is measured" {
+    anim_clear()
+    let dir = Os.temp_dir() + "/ludic-anim-test"
+    Fs.mkdir(dir)
+    let b = buffer(40)
+    put_f(b, 0, 0.0)
+    put_f(b, 4, 1.0)
+    put_f(b, 8 + 12, 1.0)
+    put_f(b, 8 + 24, HALF)
+    put_f(b, 8 + 28, HALF)
+    expect(Fs.write_bytes(dir + "/t.bin", b, 40))
+    gltf_doc = Json.parse("{\"buffers\": [{\"uri\": \"t.bin\"}], \"bufferViews\": [{\"byteOffset\": 0}, {\"byteOffset\": 8}], \"accessors\": [{\"bufferView\": 0, \"count\": 2, \"type\": \"SCALAR\"}, {\"bufferView\": 1, \"count\": 2, \"type\": \"VEC4\"}], \"animations\": [{\"name\": \"bear_walk\", \"samplers\": [{\"input\": 0, \"output\": 1}], \"channels\": [{\"sampler\": 0, \"target\": {\"node\": 1, \"path\": \"rotation\"}}, {\"sampler\": 0, \"target\": {\"node\": 1, \"path\": \"scale\"}}]}]}")
+    expect_eq(anim_read_doc(dir), 1)
+    expect_eq(anim_read_doc(dir), 0)
+    let c = anim_find("bear_walk")
+    expect(c != null)
+    expect_eq(c.n, 1)
+    expect_near(c.dur, 1.0, 0.0001)
+    expect_near(c.vals[7], HALF, 0.0001)
+    expect_near(anim_first_key(c), 0.0, 0.0001)
+    c.times[0] = 1.0 / 30.0
+    expect(anim_first_key(c) > 0.03)
+  }
+}