FJ9.7 — Artifact-Driven Animation

Date: 2026-08-23.

Animation is playback of recorded evidence, not a simulator run again.

STATUS

PASSED. FJ9.7a–9.7d closed. Three canonical single-solver animations and one paired animation export as MP4 from a fresh process, with ENV_MODEL_CALLS=0, PYTHON_MODULES=none and MCTS_LOADED=false measured rather than asserted by intent.

ENVIRONMENT

WSL Ubuntu-Baru, juliaup override 1.11.3
Pkg.test with ddm-ref + ddm-torch + MCTS.jl : 186 testsets, 148 645 assertions
render check (fresh process, CairoMakie only): RENDER_EXIT=0
                                               ENV_STEPPED=false
                                               ENV_MODEL_CALLS=0
                                               PYTHON_MODULES=none
                                               MCTS_LOADED=false

ENV_MODEL_CALLS is the strongest of these. The check builds an InstrumentedMDP, renders every frame and every video through it, and prints the counter. A single gen anywhere on the rendering path would make it non-zero.

THE RENDERING PATH

decisions.csv -> load_decision_log -> animation_sequence -> AnimationSequence
                                                          -> frame_scene
                                                          -> Makie -> MP4

The MDP is constructed for one purpose: the static track description. FJ8.0 certified map and stop_signs as the shared-by-design, never-written part of a world state, and the render check confirms empirically that two different reset seeds produce the same StaticWorld (STATIC_WORLD_SEED_INVARIANT=true). Nothing episode-specific can enter through it, and the reference state's ego pose, ducks, stop memory and RNG are all discarded.

Everything that moves comes from the log. Per-frame geometry is computed from the logged pose by the same functions the physics uses — get_agent_corners for the footprint, closest_curve_point for the lane frame the stop line is drawn in — applied to recorded values rather than to a re-simulated state.

FJ9.7a — THE DATA CONTRACT

AnimationFrame carries one logged decision: pose, lane projection, stop subsystem, duck subsystem (lane-relative), action, reward and running return, the events logged at that decision, planning cost, and episode status. AnimationSequence carries the frames, the outcome, the source fingerprint, and its own fingerprint over the ordered frame identities.

horizon is the episode's own length. An episode that terminated at 86 decisions has 86 frames; padding it to 150 would invent stationary frames that never happened.

The timeline is the decision index

FJ9.6a found no wall-clock timestamp in the artefact, so this is not real-time playback and the figure never says it is. The caption carries Decision-index playback, and framerate is documented as a display rate with no claim about elapsed world time.

model_time exists, as an exact identity only:

model_time(t) = (t - 1) * frame_skip * dt

It is labelled model time, NOT recorded wall-clock time, and neither frame_skip nor dt is in the log, so the caller must supply them from the configuration the protocol used — there is no default that could silently be wrong. planning_time is never used for pacing; it is computational latency, and a test asserts it is not reachable from the model-time label.

What is ABSENT

The duck is the one that would have been easy to fake. Its reset pose is seed-invariant and sitting right there in the reference state, so drawing it would have cost one line and produced a stationary duck for 150 frames that looks exactly like data. The world panel omits it; the lane-relative panel shows what was actually recorded. A test asserts the stated reason names the lane frame, so a later reader does not "fix" the omission.

FJ9.7b — SINGLE-EPISODE PLAYBACK

Layout: world view, a text panel of the current decision, and eight history axes — one per unit, following the FJ9.6 rule (lateral offset m, heading rad, speed m/s, angular rate rad/s, distance to stop line m, cumulative return, generative calls, planning time s).

Two invariants carry the gate, because both failure modes look perfectly convincing on screen:

trajectory at frame t = rows 1..t          never the whole episode
event marker at frame t                    only if logged at decision <= t

Both are enforced structurally: the panels can only reach the data through trajectory_through, series_through and events_through, all of which slice 1:t. Tests check the prefix property directly (series_through(seq, 33, nm) == series_through(seq, 150, nm)[1:33]) and walk every frame from 1 to 27 asserting the full_stop at decision 28 is invisible.

The camera is fixed to the map and the signs, never to the trajectory. A view fitted to the episode's full extent would hold the vehicle's future in the empty space around it, and a view refitted per frame would jitter.

Missing d_stop renders as a gap plus a grey tick, never as zero, in both the history axis and the numeric readout (MISSING (no candidate)).

FJ9.7c — THREE CANONICAL ANIMATIONS

Selected by stated rule, never by inspection. select_episode implements the rules and returns the justification, which is printed by the render check and written into the figure.

The median-return rule reproduces the FJ8.4b median column exactly from the per-decision log — 7.75 for mcts@1k and −211.59 for dpw@1k — which is a third independent path to those numbers and is pinned by test.

The TD3 animation is the one this gate exists for. It shows the FJ9.6 correction as behaviour rather than as a table: approach, full_stop at decision 28, sigma_stop from there to the end, d_stop → 0, speed collapsing to ~0.005 m/s, and the cumulative return falling linearly for 120 decisions under the stagnation penalty while the vehicle does not move.

FJ9.7d — PAIRED PLAYBACK

Same seed, side by side, on the absolute decision index. Decision 40 on the left is decision 40 on the right, and both panels share one x-span so it is also decision 40 at the same place on the page — the first version gave each panel its own span, which was numerically correct and visually misaligned.

When the shorter episode ends, its panel freezes on the terminal frame with a banner (FROZEN — TERMINATED at decision 116 (offroad)) and the header reports both states. It is never looped, restarted, or stretched.

Normalised progress is deliberately not offered for animation. It would align DPW's decision 40 of 80 with TD3's decision 75 of 150, which are not the same decision and were not taken under the same conditions. Aggregate plots (FJ9.6d) may use it; playback may not.

NEGATIVE CONTROLS

All four run against edited copies of the real log, not synthetic stand-ins.

Every one also changes the source fingerprint and the animation fingerprint.

FILES

MP4 is primary; the exporter falls back to GIF and reports which it produced, so a codec problem degrades the artefact format without touching the scientific content.

KNOWN DEVIATIONS

1. No duck in the world view. The log has no world-frame duck track. Reported ABSENT; the lane-relative pair is shown instead.
2. The paired layout shows four history axes per side, not eight. Sixteen axes plus two world views is unreadable at video resolution. The single animation carries all eight.
3. The MDP is constructed to obtain the static track description. It is never stepped, and ENV_MODEL_CALLS=0 is measured every run.

NEW DEFECTS

Three of my own, all in the renderer, all found by looking at output rather than at exit codes.

vlines! with an empty vector makes Makie reduce over an empty collection while computing data limits, and a panel legitimately has no events for most of an episode. Event markers are now line segments spanning the current data range. The same crash was reachable a second way: a panel whose values are all missing so far never set explicit y-limits and fell through to autolimits. It now gets a stated range and renders as an empty axis, which is also the honest picture — "no data yet", not "zero".

The paired figure was never initialised before record first laid it out, so an empty poly! observable raised a BoundsError; the single-episode path had been initialising via setframe!(1) all along. And the caption rounded planning_time to four decimal places, which printed a tabular policy's 4×10⁻⁵ s as 0.0 — a measured value displayed as if it were absent, which is the exact confusion FJ9.6 spent a gate preventing. It now rounds to significant digits.

ACCEPTANCE

[x] decisions.csv sole dynamic evidence source
[x] no solver/environment execution      ENV_MODEL_CALLS=0, measured
[x] one logged decision = one canonical frame
[x] trajectory grows only through current frame
[x] future events never shown early
[x] missing d_stop stays missing
[x] horizon vs environment termination visually distinct
[x] absolute decision indexing is canonical timeline
[x] unequal episode lengths handled without temporal stretching
[x] per-decision planning cost visible   (calls and seconds, separate axes)
[x] TD3 stagnation behaviour faithfully reproducible
[x] DPW compute collapse visible from logged values
[x] PNG/frame rendering works fresh-process
[x] GIF/MP4 headless export works        (MP4 produced; GIF fallback in place)
[x] no Python / no MCTS required

SCIENTIFIC INTERPRETATION

FJ9.6 established that TD3's failure was invisible in the aggregate. FJ9.7 makes it watchable, which is a different and useful thing: the stall is not a number that needs explaining but a vehicle that stops and stays stopped while the return counts down behind it.

The paired animation adds something neither gate had. On one seed, from one initial condition, two planners diverge into two distinct failure modes — TD3 halts permanently at the sign while DPW drifts off the road at decision 116 — and the pairing that FJ8.4b was built on is what makes that comparison mean anything. Freezing rather than stretching the shorter panel is what preserves it.

NEXT GATE

FJ9.8 — publication composite. The TD3 correction should reach the final figures.