@voqalize/avatar 0.2.2 → 0.4.0

package/src/gaze.js

@@ -9,8 +9,17 @@
  *      after — accelerating, cruising, braking to a stop (see HEAD_ACCEL).
  *   2. The head only goes part of the way. Real people under-rotate the head
  *      and let the eyes carry the remainder.
- *   3. A blink fires on any large shift. Gaze-evoked blinks are involuntary and
- *      their absence is uncanny even though nobody can name what's wrong.
+ *   3. A blink rides a large shift. Gaze-evoked blinks are involuntary, more
+ *      likely the bigger the shift (Evinger et al. 1994), and their absence
+ *      is uncanny even though nobody can name what's wrong.
+ *
+ * A rig that states what a pose unit is in degrees (`angles`) gets the rest of
+ * the eye-head system too: its own target table, where the head carries most
+ * of a look and the eyes land about a third of the way off centre (Freedman &
+ * Sparks; Pejsa & Andrist's gaze model), and a vestibulo-ocular reflex in the
+ * mixer that holds the eyes on their target while the head moves under them.
+ * Without the reflex every nod and every speech pose change is also a look
+ * somewhere else.
  */
 
 export const GAZE_TARGETS = {
@@ -29,6 +38,12 @@ export const GAZE_TARGETS = {
   SCREEN_BOTTOM: { px:  0.00, py:  0.62, hx:  0.00, hy:  0.24 },
   SCREEN_WORK:     { px: -0.62, py:  0.18, hx: -0.26, hy:  0.12 },
   NOTES:         { px:  0.18, py:  0.72, hx:  0.04, hy:  0.28 },
+  // The agent's own display, just under the camera: where anyone on a video
+  // call looks while they do something for you. Eyes a little down and the
+  // head barely joins, because reading your own screen is an eye movement.
+  // SCREEN_WORK turned the head a quarter of the way to the left and held it
+  // there for the whole task, which read as turning away from the user.
+  OWN_SCREEN:    { px:  0.04, py:  0.30, hx:  0.02, hy:  0.07 },
   // The classic "recalling something" break of eye contact. Keep it for the
   // stylized "let me think" beat — audiences read up-and-away regardless of
   // whether real thinkers do it.
@@ -39,12 +54,64 @@ export const GAZE_TARGETS = {
   // that the lid follow shades the eyes without sealing them; the head
   // carries a share so the pupils stay inside the aperture at tile size.
   AWAY_DOWN:     { px: -0.45, py:  0.42, hx: -0.18, hy:  0.20, roll: 0.04 },
+  // Level and to the side: the other third of measured cognitive aversions
+  // (31.3%, §4.2), neither up nor down. Right, so a thinker who alternates it
+  // with AWAY_DOWN's left is not looking at the same patch of floor twice.
+  AWAY_SIDE:     { px:  0.55, py:  0.08, hx:  0.20, hy:  0.03, roll: -0.03 },
 };
 
 export const GAZE_NAMES = Object.keys(GAZE_TARGETS);
 
 const HEAD_FOLLOW_TAU = 0.34; // roll only — the head lags the eyes badly, on purpose
-const BLINK_THRESHOLD = 0.45; // shift magnitude that triggers a gaze-evoked blink
+// Gaze-evoked blink odds, ramping with the size of the shift: none below the
+// first number, certain above the second. Degrees on a rig that states them —
+// stylised, well under Evinger's 10-40°, because every look here is scaled
+// down to what a head-and-shoulders crop shows: a thinking look (~11°) always
+// blinks, a speaker's glance aside (~5°) about a third of the time, a reading
+// step never. Pose units on one that does not, where 0.45 was once a hard
+// threshold and now blinks a quarter of the time.
+const BLINK_RAMP_DEG = [3, 9];
+const BLINK_RAMP_UNITS = [0.3, 0.9];
+// The upper lid rides the eye, as a share of the pupil's travel from USER. A
+// line face's lid follows down only: its lid travels further than its iris,
+// and one that opened past rest on an upward look bared a band of sclera and
+// read as alarm. A rig whose lid and iris travel alike follows both ways.
+const LID_FOLLOW = { down: 0.34, up: 0 };
+// A turning head dips. Heads travel in arcs, not on a flat plane — a turn that
+// holds its pitch reads as a turntable (Williams, *The Animator's Survival
+// Kit*, on head turns). A share of the turn's length, read off the yaw speed
+// as v²/a — which at the middle of a braked turn *is* its length — so the dip
+// peaks mid-turn, is gone when the head lands, and sizes itself on a rig whose
+// head moves four times faster. A 0.75-unit look on tara sinks ~0.8°.
+const HEAD_DIP = 0.06;
+// A gated aversion waits this long past its due time for a phrase boundary,
+// then goes anyway: a run of speech with no pause in it still gets its look.
+const GATE_WAIT = 1.5;
+// How much of a drift the head joins. Little: a drift is a reader's scan or a
+// thinker's gaze moving where it rests, and both are eye movements. At 0.35
+// the head nodded along with every step of a reading scan, which on a
+// photographic face reads as a bobbing head rather than a moving eye.
+const DRIFT_HEAD = 0.12;
+
+// Where the eyes land inside a fixation on the user: the user's two eyes and,
+// less often, the mouth — the triangle a listener's gaze scans a face with.
+// A fixation that sits on one point is the stare; one that visits these
+// reads as someone looking at a person. Pose units at scanGain 1.
+const FACE_POINTS = [[-0.05, 0], [0.05, 0], [-0.05, 0], [0.05, 0], [0, 0.07], [0, 0]];
+// Default gap between those fixational saccades, seconds. Eyes Alive
+// (docs/research-biomechanics.md §4.5) is that this interval differs by mode,
+// so states set their own through the mixer.
+const SCAN_EVERY = [0.5, 1.6];
+// Fixational jumps inside a look away, as a share of the usual size. The look
+// is one saccade out and one back; big jumps inside it read as searching, and
+// on a rig with a large scanGain one could land the eyes back on the face
+// mid-look, so the look happened and did not read.
+const AVERT_SCAN = 0.35;
+// The least a new fixation moves from the last, pose units at scanGain 1.
+// Drawn independently, one in a few landed on the one before, and two jumps
+// that go nowhere are one long stare: 6 s on a single point, measured, in the
+// state that is waiting for the user to come back.
+const MIN_STEP = 0.025;
 
 // Head follow is ballistic, not exponential. An exponential chase has its peak
 // velocity at t=0 and then creeps forever — motion that starts instantly and
@@ -56,8 +123,129 @@ const BLINK_THRESHOLD = 0.45; // shift magnitude that triggers a gaze-evoked bli
 // 0.55-unit swing in ~0.85s. The mixer still low-passes headYaw/headPitch at
 // τ=0.16 downstream, which rounds the hard stop into a short settle — that
 // cascade is deliberate, so do not "help" by softening the brake here too.
+// A rig whose head unit is a few degrees sets its own (`headAccel`,
+// `headSpeed`): an amble sized in line-face units took a 7° look 1.2 s.
 const HEAD_ACCEL = 4.0; // units/s² — sets both launch and braking firmness
 const HEAD_SPEED = 0.9; // units/s — cruise cap; only long swings ever reach it
+// The head takes only a fraction of what an aversion takes — a brief look-away
+// is an eye movement that the head barely joins, and a head that follows it
+// fully reads as turning away. Sized for a line face, whose pupils travel far
+// inside a big eye; a rig sets its own split (`avertSplit`).
+const HEAD_SHARE = 0.22;
+
+/**
+ * Ballistic follow: steer velocity toward "full speed at the goal, but never
+ * faster than can still brake to a stop within the distance left". The
+ * braking bound is v² = 2·a·d solved for v, with a half-step correction for
+ * discrete time (the −maxA term) so a frame never lands past the goal. As d
+ * shrinks the bound falls to zero, which *is* the deceleration — no separate
+ * easing curve.
+ */
+function chase(pos, vel, gx, gy, dt, accel, speed) {
+  const dx = gx - pos.x;
+  const dy = gy - pos.y;
+  const d = Math.hypot(dx, dy);
+  const maxA = accel * dt;
+  const brake = 0.5 * (Math.sqrt(maxA * maxA + 8 * accel * d) - maxA);
+  const goal = Math.min(speed, brake);
+  let ax = (d ? (dx / d) * goal : 0) - vel.x;
+  let ay = (d ? (dy / d) * goal : 0) - vel.y;
+  const a = Math.hypot(ax, ay);
+  if (a > maxA) { ax *= maxA / a; ay *= maxA / a; }
+  vel.x += ax;
+  vel.y += ay;
+  // A dropped-frame dt could step past the goal; land on it instead. Judged
+  // on the toward-goal component so a mid-retarget frame with sideways
+  // velocity keeps flying rather than teleporting.
+  const along = d ? ((vel.x * dx + vel.y * dy) / d) * dt : 0;
+  if (along >= d && d >= 0) {
+    pos.x = gx;
+    pos.y = gy;
+    vel.x = 0;
+    vel.y = 0;
+  } else {
+    pos.x += vel.x * dt;
+    pos.y += vel.y * dt;
+  }
+}
+
+/**
+ * Aversion profiles: a brief, deliberate break of eye contact that returns to
+ * whatever the gaze target already was.
+ *
+ * This exists because *holding* the user's eyes is not the attentive pose it
+ * looks like. Wang & Gratch (CHI 2010, n=133) ran the condition directly: a
+ * virtual listener that simply stares rated no better than one that visibly
+ * ignored the user (rapport 3.49 vs 3.34, n.s.), was rated the most *tense* of
+ * the three conditions, and raised the speaker's own disfluency rate to
+ * 36.75/min against 22.44 for a responsive listener. All three were rated
+ * equally natural, so it is not an animation-quality artefact. Rossano supplies
+ * the mechanism — sustained mutual gaze is a *demand for more talk*, not a
+ * signal of attention (95% of sequences expanded when both parties kept looking;
+ * 84% closed when both withdrew) — and Binetti (N=498) puts preferred mutual
+ * gaze at 3295 ± 706 ms, which is well short of forever.
+ *
+ * The other wall is just as hard: sustained *aversion* is an ostracism cue, and
+ * measured on an animated character (Chotpitayasunondh & Douglas, N=128) it
+ * costs η²ₚ = .52–.56 with post-hoc d of 1.09–2.69, with partial inattention
+ * costing most of what total inattention costs. So this is not "look away
+ * sometimes" — it is a narrow band, and the numbers below are Andrist's measured
+ * ones rather than a taste call (docs/research-biomechanics.md §4.2). The spike
+ * that arrived at this band, and what it measured on peep, is
+ * docs/research-active-listening.md.
+ *
+ * `every`/`dur` are seconds. `dirs` are unit-ish directions weighted by how
+ * often each is taken; magnitude is scaled per-fire.
+ */
+export const AVERSION = {
+  // Idle, and before a call: present and relaxed rather than attending. The
+  // eyes rest on the user a comfortable bout (preferred mutual gaze 3.3 s,
+  // §4.1) and leave for longer, softer looks than a listener's — level or a
+  // little down, never up, because up is the "recalling" look and there is
+  // nothing to recall. This is the first face anyone sees, and it used to
+  // open looking up and to the left.
+  IDLE: {
+    every: [3.2, 6.0],
+    dur: [1.4, 2.8],
+    mag: [0.30, 0.46],
+    dirs: [[-1, 0.10], [1, 0.10], [-0.8, 0.45], [0.8, 0.40], [-1, 0.25]],
+  },
+  // While listening: 1.14 s (SD 0.27) every 7.21 s (SD 1.88), 57.5% sideways.
+  // Sideways dominates because down reads as submission and up reads as
+  // exasperation on a face this schematic.
+  LISTEN: {
+    every: [5.3, 9.1],
+    dur: [0.85, 1.45],
+    mag: [0.30, 0.44],
+    dirs: [[-1, 0.06], [-1, 0.06], [1, 0.02], [1, 0.02], [-0.7, -0.5], [0.5, 0.35]],
+  },
+  // While speaking: 1.96 s every 4.75 s mid-turn, a turn-start look on 73% of
+  // turns — the planning look, away to find the words and back to deliver
+  // them — and none in the last 2.4 s, because the floor is handed back under
+  // mutual gaze (§4.2; the mixer owns that hold, since only it can see the
+  // cue track). Human speakers hold the listener's eyes for about 41% of a
+  // turn. This is set for 70-80% instead — the
+  // listener here watches a face on a screen, where 41% reads as evasive — so
+  // the looks keep about Andrist's length and come a little less often.
+  // `every` sits under Andrist's 4.75 s because the phrase gate and the
+  // end-of-turn hold both stretch it: at 3.8-5.8 a turn measured 6-8 looks a
+  // minute and 76-82% contact, the top edge of the band rather than its middle.
+  // `gated`: a mid-turn look waits for a phrase boundary, because speakers
+  // break gaze between clauses, not in the middle of a word.
+  SPEAK: {
+    every: [3.4, 5.2],
+    dur: [1.5, 2.3],
+    mag: [0.34, 0.48],
+    dirs: [[-1, 0.12], [1, 0.12], [-0.75, 0.55], [0.7, 0.5], [-0.6, -0.45]],
+    gated: true,
+    start: { p: 0.73, dur: [1.1, 1.8], window: 0.8 },
+  },
+};
+// There is deliberately no THINK profile. The *cognitive* aversion — 3.54 s
+// (SD 1.26), splitting 39.3% down / 29.4% up / 31.3% side (§4.2) — is longer
+// and deeper than the listening kind, and THINKING renders it with its own
+// dwell cycle (a `glance` in STATES), which moves the whole gaze target rather
+// than nudging off it. Two mechanisms producing the same look would fight.
 
 export class GazeLayer {
   constructor() {
@@ -68,88 +256,276 @@ export class GazeLayer {
     this.onLargeShift = null;
     this.jitter = { x: 0, y: 0 };
     this._nextMicro = 0;
+    /** How far a fixational saccade travels. A pupil unit is a different
+     *  angle on every rig, and one sized for a line face is invisible on a
+     *  photographic one — the mixer passes the rig's `saccadeGain`. */
+    this.scanGain = 1;
+    /** The same for conversational aversions (AVERSION), which are sized for
+     *  a line face too. On tara a speaking look-away at 1 is 2-3 px of iris
+     *  travel, and a recorded call read the whole turn as a locked stare. */
+    this.avertGain = 1;
+    /** Where each named look goes, on this rig. The shared table is sized for
+     *  a line face; a rig whose eyes are a fraction of its head's reach
+     *  brings its own, where the head carries the look. */
+    this.targets = GAZE_TARGETS;
+    /** How an aversion divides: `eye` of it stays in the eyes, and `head` of
+     *  it (in head units per pupil unit) goes to the neck. */
+    this.avertSplit = { eye: 1, head: HEAD_SHARE };
+    /** Degrees per pose unit, `{ eye: {x, y}, head: {x, y} }`, or null on a
+     *  rig that has not said — then shifts are measured in pupil units. */
+    this.angles = null;
+    this.lidFollow = LID_FOLLOW;
+    this.headAccel = HEAD_ACCEL;
+    this.headSpeed = HEAD_SPEED;
+    /** [min, max] seconds between fixational saccades, or null for the
+     *  default. Set per state by the mixer. */
+    this.scanEvery = null;
+    this._t = 0;
+    // --- aversion scheduler
+    this.aversion = null;      // one of AVERSION, or null for none
+    /** Set true when eye contact must be held: the floor is about to change
+     *  hands. Andrist prohibits intimacy-regulating aversions near utterance
+     *  end for exactly this reason — the floor is passed with mutual gaze, and
+     *  an avatar that looks away as the user finishes has just declined it. */
+    this.hold = false;
+    this._avNext = 0;
+    this._avUntil = 0;
+    this._avVec = { x: 0, y: 0 };
+    this._avAmt = 0;           // glided 0..1 so the return is a movement, not a cut
+    this._avWant = 0;          // where _avAmt is going; the head chases this
+    this._avHead = { x: 0, y: 0 };
+    this._avVel = { x: 0, y: 0 };
+    this._avProfileRef = undefined;
+    /** Whether now is a phrase boundary, for a `gated` profile. Set by the mixer. */
+    this.gate = true;
+    this._avStartBy = 0;       // a turn-start look may still fire until this time
+    // --- drift: a small held offset about the target, set by the mixer
+    this.drift = { x: 0, y: 0 };
+    this._drift = { x: 0, y: 0 };
   }
 
+  /** Adopt an aversion profile (or null). Cheap to call every frame. */
+  setAversion(p) {
+    if (p === this._avProfileRef) return;
+    this._avProfileRef = p;
+    this.aversion = p || null;
+    // Re-arm rather than inherit: a state that averts must not fire the instant
+    // it is entered off a stale timestamp from one that didn't.
+    this._avNext = this._t + (p ? p.every[0] + Math.random() * (p.every[1] - p.every[0]) : 0);
+    this._avUntil = 0;
+    this._avStartBy = 0;
+  }
+
+  /** A turn begins. If the profile has a turn-start look and the dice take it,
+   *  it fires as soon as `hold` allows inside its window, or not at all — a
+   *  planning look that arrives a second into the words is just a look. */
+  startTurn() {
+    const s = this.aversion && this.aversion.start;
+    if (s && Math.random() < s.p) this._avStartBy = this._t + s.window;
+  }
+
+  _avert(t, dt) {
+    const p = this.aversion;
+    if (!p) {
+      // Glide home even after the profile is gone, so a state change mid-look
+      // returns the eyes instead of snapping them.
+      this._avAmt = Math.max(0, this._avAmt - dt / 0.18);
+      this._avWant = 0;
+      return;
+    }
+    if (this._avUntil && this.hold) {
+      // `hold` ends a look already running, and the schedule starts over
+      // behind it. Suspending the look instead sent the eyes back out the
+      // moment the hold lifted — out, back and out again across one turn
+      // edge, which is three saccades where the floor wanted none.
+      this._avUntil = 0;
+      this._avStartBy = 0;
+      this._avNext = t + p.every[0] + Math.random() * (p.every[1] - p.every[0]);
+    } else if (this._avUntil && t >= this._avUntil) {
+      this._avUntil = 0;
+      this._nextMicro = 0;
+      this._avNext = t + p.every[0] + Math.random() * (p.every[1] - p.every[0]);
+    } else if (!this._avUntil && !this.hold && (t < this._avStartBy || (t >= this._avNext
+        && (this.gate || !p.gated || t >= this._avNext + GATE_WAIT)))) {
+      const dur = t < this._avStartBy ? p.start.dur : p.dur;
+      this._avStartBy = 0;
+      this._avUntil = t + dur[0] + Math.random() * (dur[1] - dur[0]);
+      this._nextMicro = 0;
+      const d = p.dirs[(Math.random() * p.dirs.length) | 0];
+      const m = (p.mag[0] + Math.random() * (p.mag[1] - p.mag[0])) * this.avertGain;
+      this._avVec.x = d[0] * m;
+      this._avVec.y = d[1] * m;
+      // A look away is a gaze shift like any other, and blinks by its size.
+      const s = this.avertSplit;
+      const out = { px: d[0] * m * s.eye, py: d[1] * m * s.eye, hx: d[0] * m * s.head, hy: d[1] * m * s.head };
+      this._shiftBlink({ px: 0, py: 0, hx: 0, hy: 0 }, out);
+    }
+    // `hold` cancels an aversion already running, it does not merely postpone
+    // the next: the turn can end mid-look, and the eyes have to be back.
+    const want = this._avUntil && !this.hold ? 1 : 0;
+    this._avWant = want;
+    // Out fast, back slightly slower. A saccade away is ballistic; the return
+    // to a face is a fraction more deliberate, and symmetric timing here is one
+    // of the things that makes a rig read as a metronome.
+    const rate = want ? dt / 0.055 : dt / 0.11;
+    this._avAmt = want
+      ? Math.min(1, this._avAmt + rate)
+      : Math.max(0, this._avAmt - rate);
+  }
+
+  /** How far off-target the eyes currently are, 0..1. The mixer reads this to
+   *  keep the trunk out of it — an aversion is eyes and a little head, never a
+   *  body turn. */
+  get averted() { return this._avAmt; }
+
   /**
    * @param {string} name  one of GAZE_NAMES
    * @param {{x:number,y:number}} [override] normalized -1..1 escape hatch for
    *        when the server knows exact screen coordinates
+   * @param {boolean} [blink] true to ask for the evoked blink whatever the
+   *        size, false to withhold it, omitted to leave it to the odds
    */
-  set(name, override) {
+  set(name, override, blink) {
     const next = override
       ? { px: override.x, py: override.y, hx: override.x * 0.42, hy: override.y * 0.36 }
-      : GAZE_TARGETS[name] || GAZE_TARGETS.USER;
-    const d = Math.hypot(next.px - this.target.px, next.py - this.target.py);
+      : this.targets[name] || this.targets.USER;
+    const prev = this.target;
     this.target = next;
     this.name = override ? 'CUSTOM' : name;
-    if (d > BLINK_THRESHOLD && this.onLargeShift) this.onLargeShift();
+    if (blink !== false) this._shiftBlink(prev, next, blink);
+  }
+
+  /** Roll the gaze-evoked blink for a shift from one target to another. */
+  _shiftBlink(a, b, force) {
+    if (!this.onLargeShift) return;
+    const g = this.angles;
+    const [lo, hi] = g ? BLINK_RAMP_DEG : BLINK_RAMP_UNITS;
+    const size = g
+      ? Math.hypot((b.px - a.px) * g.eye.x + (b.hx - a.hx) * g.head.x,
+                   (b.py - a.py) * g.eye.y + ((b.hy || 0) - (a.hy || 0)) * g.head.y)
+      : Math.hypot(b.px - a.px, b.py - a.py);
+    // The listener is told which of the two it is: a beat the state named
+    // outranks the blink budget, a roll of the odds does not.
+    //
+    // A shift past `hi` is not a roll. The ramp prices those at probability 1,
+    // so the odds have already decided, and the budget was overruling the
+    // ramp's own certainty: on the call reel the three biggest hops a screen
+    // hunt makes -- 10.9°, 9.5°, 9.4° -- reached the eyelids as a coin toss
+    // that the budget then refused. A saccade that size wants every channel a
+    // head turn has (research-biomechanics §1.2: yaw, a pitch dip, a blink).
+    //
+    // Affordable because it is only the tail. Forcing the *whole* evoked
+    // population is what cost 2.12x the authored rate and is why
+    // EVOKED_EARLIEST exists; this is three shifts in a minute-long run, and
+    // over four seeds 78 blinks fired where the budget alone fired 79.
+    //
+    // The refractory is untouched and refuses half of these, because a hunt
+    // retargets every 0.8-2.0 s and that is faster than 1.5 s. That is the
+    // refractory working as written -- a glance out and back carries one
+    // blink, not two. So this masks the isolated large look and not the
+    // rapid ones inside a hunt, which stay open-eyed.
+    if (force || size >= hi) this.onLargeShift(true);
+    else if (Math.random() < (size - lo) / (hi - lo)) this.onLargeShift(false);
+  }
+
+  /** The lid's share of an eye at `pupilY`, added to the lid channels. */
+  lidBias(pupilY) {
+    const rest = GAZE_TARGETS.USER.py;
+    const f = this.lidFollow;
+    return rest * f.down + (pupilY - rest) * (pupilY > rest ? f.down : f.up);
   }
 
   /**
-   * Micro-saccades. Eyes are never still; a perfectly fixed pupil reads as
-   * dead. Small, frequent, and irregular.
+   * Fixational saccades. Eyes are never still, and they do not wobble about a
+   * point either: they jump and *hold*, a new fixation every half second or
+   * so. On the user they visit the user's eyes and mouth; anywhere else they
+   * land a little way about the target. These used to decay back to the exact
+   * centre within a second, so between jumps every state stared at one point,
+   * and at a photographic face's scale the jumps were under a pixel —
+   * §4.6's "sub-threshold jitter signals nothing", measured.
    */
-  _micro(t, dt) {
-    if (t >= this._nextMicro) {
-      this._nextMicro = t + 0.7 + Math.random() * 1.6;
-      this.jitter.x = (Math.random() - 0.5) * 0.16;
-      this.jitter.y = (Math.random() - 0.5) * 0.11;
+  _micro(t) {
+    // A state's `scan` is [min gap, max gap, amplitude]: a reader's own
+    // line steps carry the movement, so a small third value keeps the
+    // fixational jumps from blurring them.
+    const [a, b, k = 1] = this.scanEvery || SCAN_EVERY;
+    // And a state that scans faster gets it from the frame it starts, not once
+    // the state before it has finished waiting. The deadline is drawn when a
+    // fixation fires, under whatever cadence was in force *then*, so entering a
+    // hunt out of a slower state finds a gap longer than the hunt's own maximum
+    // already on the clock and spends its first second at the old rate. On a
+    // two-second state that is most of the state: eyes that hold dead still and
+    // then jump once, which is read as mechanical rather than as searching.
+    // Clamped, not redrawn: the wait already served still counts, and a state
+    // change must not cost a draw from the shared stream — every blink and
+    // wander in the reel comes off it, and spending one here would move them.
+    if (this._nextMicro > t + b) this._nextMicro = t + b;
+    if (t < this._nextMicro) return;
+    this._nextMicro = t + a + Math.random() * (b - a);
+    const g = this.scanGain * k * (this._avUntil ? AVERT_SCAN : 1);
+    // The face scan is for looking *at* the user. A look away is not one,
+    // and a face-point offset riding on it pulled short looks back onto the
+    // face — the look happened and did not read.
+    const face = (this.name === 'USER' || this.name === 'USER_EAR') && !this._avUntil;
+    let x = 0, y = 0;
+    for (let i = 0; i < 4; i++) {
+      if (face) {
+        const [fx, fy] = FACE_POINTS[(Math.random() * FACE_POINTS.length) | 0];
+        x = (fx + (Math.random() - 0.5) * 0.02) * g;
+        y = (fy + (Math.random() - 0.5) * 0.02) * g;
+      } else {
+        x = (Math.random() - 0.5) * 0.12 * g;
+        y = (Math.random() - 0.5) * 0.08 * g;
+      }
+      if (Math.hypot(x - this.jitter.x, y - this.jitter.y) >= MIN_STEP * g) break;
     }
-    const decay = 1 - Math.exp(-dt / 0.5);
-    this.jitter.x -= this.jitter.x * decay;
-    this.jitter.y -= this.jitter.y * decay;
+    this.jitter.x = x;
+    this.jitter.y = y;
   }
 
   update(t, dt) {
-    this._micro(t, dt);
-
-    // Ballistic head follow: steer velocity toward "full speed at the target,
-    // but never faster than can still brake to a stop within the distance
-    // left". The braking bound is v² = 2·a·d solved for v, with a half-step
-    // correction for discrete time (the −maxA term) so a frame never lands
-    // past the target. As d shrinks the bound falls to zero, which *is* the
-    // deceleration — no separate easing curve.
-    const dx = this.target.hx - this.head.x;
-    const dy = this.target.hy - this.head.y;
-    const d = Math.hypot(dx, dy);
-    const maxA = HEAD_ACCEL * dt;
-    const brake = 0.5 * (Math.sqrt(maxA * maxA + 8 * HEAD_ACCEL * d) - maxA);
-    const goal = Math.min(HEAD_SPEED, brake);
-    let ax = (d ? (dx / d) * goal : 0) - this.vel.x;
-    let ay = (d ? (dy / d) * goal : 0) - this.vel.y;
-    const a = Math.hypot(ax, ay);
-    if (a > maxA) { ax *= maxA / a; ay *= maxA / a; }
-    this.vel.x += ax;
-    this.vel.y += ay;
-    // A dropped-frame dt could step past the target; land on it instead.
-    // Judged on the toward-target component so a mid-retarget frame with
-    // sideways velocity keeps flying rather than teleporting.
-    const along = d ? ((this.vel.x * dx + this.vel.y * dy) / d) * dt : 0;
-    if (along >= d && d >= 0) {
-      this.head.x = this.target.hx;
-      this.head.y = this.target.hy;
-      this.vel.x = 0;
-      this.vel.y = 0;
-    } else {
-      this.head.x += this.vel.x * dt;
-      this.head.y += this.vel.y * dt;
-    }
+    this._t = t;
+    this._avert(t, dt);
+    this._micro(t);
+    // Drift lands like a saccade — quick, then held — not as a slide.
+    const kd = 1 - Math.exp(-dt / 0.04);
+    this._drift.x += (this.drift.x - this._drift.x) * kd;
+    this._drift.y += (this.drift.y - this._drift.y) * kd;
+    const s = this.avertSplit;
+    const avx = this._avVec.x * this._avAmt * s.eye + this._drift.x;
+    const avy = this._avVec.y * this._avAmt * s.eye + this._drift.y;
+    // The head's part of an aversion is a head movement like any other, so it
+    // gets the same ballistic launch and brake as the gaze follow — on its own
+    // follower, so the trunk (which tracks `this.head`) sits a look out.
+    const ahx = this._avVec.x * this._avWant * s.head;
+    const ahy = this._avVec.y * this._avWant * s.head;
+    chase(this._avHead, this._avVel, ahx, ahy, dt, this.headAccel, this.headSpeed);
+    chase(this.head, this.vel, this.target.hx, this.target.hy, dt, this.headAccel, this.headSpeed);
 
     // Roll stays on the exponential: its travels are tiny (≤0.55 and usually
     // ~0.08), far below where the drift tail is visible.
     const k = 1 - Math.exp(-dt / HEAD_FOLLOW_TAU);
     this.head.roll += ((this.target.roll || 0) - this.head.roll) * k;
 
-    const pupilY = this.target.py + this.jitter.y;
+    const vx = this.vel.x + this._avVel.x;
+    const dip = HEAD_DIP * vx * vx / this.headAccel;
     return {
-      pupilX: this.target.px + this.jitter.x,
-      pupilY,
-      headYaw: this.head.x,
-      headPitch: this.head.y,
+      pupilX: this.target.px + this.jitter.x + avx,
+      pupilY: this.target.py + this.jitter.y + avy,
+      headYaw: this.head.x + this._avHead.x + this._drift.x * DRIFT_HEAD,
+      headPitch: this.head.y + this._avHead.y + this._drift.y * DRIFT_HEAD + dip,
       headRoll: this.head.roll,
-      // The upper lid tracks the eye vertically. Without this, looking down
-      // exposes a band of sclera above the iris and the avatar looks startled.
-      lidBias: pupilY * 0.34,
+      // Where the head is *going*, without the looks that ride on it. The
+      // mixer turns the trunk toward this and not toward headYaw: an aversion
+      // is eyes and a little head, never a body turn.
+      trunkYaw: this.head.x,
+      // Where the head will be once it lands — the pose the eyes' target was
+      // authored against. The mixer's reflex holds the eyes on that target
+      // against every difference between this and the head actually drawn:
+      // the follow still under way, the dip, and everything above the gaze
+      // layer that moves the head.
+      aimYaw: this.target.hx + ahx + this._drift.x * DRIFT_HEAD,
+      aimPitch: this.target.hy + ahy + this._drift.y * DRIFT_HEAD,
     };
   }
 }