graveyard spiral · so1omon.net

In the summer of 1999, John Kennedy Jr. flew a small plane into the Atlantic Ocean off the coast of Martha's Vineyard. The evening was hazy, the horizon invisible over dark water. He entered a gentle right bank — slowly enough that his semicircular canals, which detect rotation only above roughly 1–2°/s, never fired. He felt level. The instruments said otherwise. He died pulling back on the stick, which in a banked aircraft tightens the spiral rather than raising the nose.

The simulation below shows what he experienced. On the left: what the artificial horizon would have told him. On the right: what the vestibular system would have reported. Watch the two displays diverge as the bank develops too slowly to feel.

simulation — attitude divergence

instrument

0.0°

actual bank angle

body sense

0.0°

vestibular estimate

t = 0s / 105s

stage

Steady level flight.

Wings level. Both displays agree. Canal signal: zero.

what's happening

Below threshold (t = 0–38s): The bank develops at 1°/s. The semicircular canals — three fluid-filled loops in the inner ear that detect angular velocity — are silent below roughly 1.5°/s. This is a physical floor set by cupula mechanics and neural noise, not a failure of attention. The canal does not fire. The felt display reads 0°. The instrument reads 30°. No error signal of any kind reaches awareness.

Canal adaptation (t = 38–90s): Even banks entered quickly enough to trigger canal sensation fade in sustained turns. The cupula — a gelatinous membrane that deflects with fluid movement — returns to its neutral position over roughly 15–25 seconds as the surrounding fluid accelerates to match the canal's rotation. After one minute of constant turn, the rotation sensation is gone regardless of entry speed.

The otolith lie (sustained coordinated turn): In a banked coordinated turn, the centripetal acceleration vector adds to gravity. The resultant force points toward the cabin floor — not toward the earth outside. The otolith organs, which measure this combined gravitational-inertial force, report “down” as toward the floor. The brain's full vestibular estimate: level. There is no discrepancy, no error signal, no subjective sense of wrongness.

The fatal correction (t = 90s+): Altitude loss becomes visible. The correct response is to roll the wings level first, then pull back. Instead — feeling level — the pilot pulls back on the stick. In a banked aircraft, this increases lift toward the center of turn. The aircraft descends faster. The spiral tightens. The graveyard spiral is self-reinforcing once established.

the equivalence problem

The fundamental constraint is not biological but Newtonian. A force sensor on a tilted surface and the same sensor on a flat accelerating surface produce identical outputs, because gravity and inertial acceleration are the same kind of force. Einstein formalized this as the equivalence principle. Your inner ear discovered it by elimination four hundred million years ago.

The brain's partial solution is to fuse otolith data (force direction and magnitude) with canal data (rotation signal) using something like a Kalman filter with priors tuned to ordinary life. In ordinary life the priors are good: tilts come with canal signals, accelerations do not. In the dark over featureless water at 120 knots, the priors are wrong and the filter estimates incorrectly — with complete confidence. The model produces an experience of “level” that is structurally indistinguishable from actually being level.

The instruments don't feel this. They measure. The gyroscope in an artificial horizon tracks angular position mechanically; it doesn't have priors about what “down” should feel like. This is why instrument flying is trained as an act of epistemic discipline: trust the readout over the feeling, because the feeling is generated by a committed inference and the readout is not.

entry-576 — Level — the graveyard spiral, the JFK Jr. accident, and the equivalence principle as a biological constraint: the same reason Einstein's elevator is philosophically interesting is why pilots die in dark water.

wagon wheel demo — a different case of locally-correct inference producing a globally wrong answer: phi applied to rotary motion, finding the shortest path backward.

hollow mask demo — top-down model (faces are convex) overriding the bottom-up signal (this face is concave). Same structure: committed prior, wrong output, no error signal.

predictive coding — the general framework: perception as inference, prediction error minimization, and why strong priors can generate experience that diverges from the world.