GN Sim Fountain

A two-tier stone fountain whose twenty jets are integrated by a Simulation Zone, witnessing the geometry-nodes-python frame contract: a Simulation Zone only advances when the scene steps one frame at a time. The evaluated state at frame N is not a function of N. AI code that calls scene.frame_set(N) and reads positions gets a silently wrong answer.

Each droplet is a point with a vel attribute. Per step the zone applies the constant-gravity update that is exact for any dt:

p += v*dt - 0.5*g*dt^2 * z
v -= g*dt * z

so at frame f every point must lie on the closed-form ballistic arc p0 + v0*t - 0.5*g*t^2*z and carry v0 - g*t*z, with t = (f - frame_start) / fps. The check recomputes both from the launch data (8 crown jets thrown up and out, 12 rim jets arcing inward; g = 9.81, 24 fps, 24 steps = 1.0 s).

Measured behavior#

Identical on 4.5.11 LTS, 5.1.2 and 5.2.1 LTS. The Simulation Zone and its cache/bake operators have not diverged across the three, so there is no version branch; the probes that established this are the checks below.

Two silent traps, then. The jump is 1.68–2.42 m off the true frame-25 arc here. The tempting cache operator "succeeds" without raising and leaves interpolated garbage between cached frames (2.42 m / 9.40 m/s off).

Tolerance. TOL = 1e-4 m and m/s. The state is float32 accumulated over 24 steps; the measured worst case is 6.6e-7 m and 1.4e-6 m/s, so the band is ~150× the observed error and ~2000× below the smallest broken case (Euler, 0.204 m). The trap must also clear TRAP_GAP_FLOOR = 0.25 m from the true arc, so a rounding difference can never pass as the trap.

What failure each check would catch:

• exit 3 — the zone is unpaired or drops points: pair_with_output not called evaluates to 0 points (--unpair)
• exit 4 — the integrator is not the exact constant-gravity update: naive p += v*dt drifts by g*dt*t/2, 0.2044 m at t = 1 s (--euler)
• exit 5 — a direct frame jump did not reproduce the one-step trap — e.g. because a bake was already present (--prebake-trap). This is what keeps the trap claim above honest on every version CI runs
• exit 6 — random-access reads after a cache fill miss the arc, or the fill operator did not return FINISHED: calculate_to_frame headless (--calc-to-frame)

The render is the proof. Every droplet bead in the still is a recorded simulated position from the stepped run (check 4). The thin line each string of beads is threaded on is the closed-form arc, drawn from the launch data, not from the simulation. A drifting integrator would pull the late beads off their line — by 0.20 m at the pool for --euler, as check 4 measures (that flag exits before rendering, so this is computed, not a rendered still).

Run#

blender --background --python gn_sim_fountain.py --
blender --background --python gn_sim_fountain.py -- --output fountain.png

The --output render path measures framing via examples/gallery_framing.py (exit 10 on violation).

Falsifiers#

Each breaks one thing and exits the code of the check it targets, on all three versions.

Exit codes#

Per-script sequential checks. 10 is the shared framing helper.

The blender-smoke workflow runs the check on Blender 5.2 LTS and 4.5 LTS (5.1 on the weekly cron, the needs-5.1 PR label, or manual dispatch). Smoke does not pass --output. Its catalog falsifiers are the three in the table above: --euler (expects exit 4), --prebake-trap (expects exit 5) and --calc-to-frame (expects exit 6).