The JSON output, field by field

The predictor writes several JSON documents per preset. This page describes each document and its fields, with real values from two presets: Various - Fire and Brimstone, a slow, fluid, flame-coloured feedback preset, and suksma - delicious carnal amorality, a fast, blue particle storm.

Document	Policy string	Size per preset	Produced by	Used for
Understanding summary	preset-understanding-v1	0.5-0.8 MB	preset_understanding.py	Input to everything else (Reading a preset)
Recipe	preset-visual-recipe-v1	80-250 KB	preset_recipe.py	The engine-neutral description; this page's main subject
Core recipe	same, graphs removed	30-60 KB	knowledge_pack.core_recipe	What the SDK and LLM readers receive
Mood evidence row	projectm-tv-static-mood-evidence	~400 bytes	static_evidence_export.py	The source of the shipped Mood table
Mood table line	attributes.tsv	~170 bytes	attribute_bundle.py --build-static	Read by the app
Prediction record	sdk-prediction-illustration-v1	~3 KB	prediction_illustration.py	Describes a prediction image
Self-comparison	checkpoint-self-comparison-v1	~2 KB	checkpoint review	Prediction vs render measurements

Conventions used everywhere

Every recipe repeats its conventions in a conventions block, so it can be read without this page:

Key	Meaning
coordinates	Per-pixel x/y in [0, 1], measured along the long side of the viewport, y = 0 at the top. The short axis spans a centred sub-range.
feedback	Each frame the previous image is resampled through the warp map, multiplied by decay, then layers are drawn on top; the composite displays the result.
warp_map	source = 0.5 + (p − 0.5) / zoom^(zoomexp^(2·rad − 1)), stretched and rotated about (cx, cy), minus (dx, dy). Content therefore moves right for positive dx and down for positive dy.
time	Seconds since the preset loaded. Audio inputs bass, mid and treb are about 1 on average; the _att variants are smoothed.
expressions	Serialised graphs {root, nodes} with projectm-eval semantics (integer conversion is floor; division by a denominator below 1e-5 gives 0).
fidelity	An approximate recipe for a similar effect, not a pixel-exact specification.

Common value shapes:

• Ranges are [low, high] lists of the values a quantity can take under the recorded premises.
• Estimates are objects with value, interval (the 10th to 90th percentile of out-of-fold errors) and evidence.
• evidence strings are JSON pointers into the understanding summary (/analysis/visual_description/...), or a label such as source-calibrated, source-estimate, mechanism-estimate, assumed or unknown.
• null means unknown or not applicable. It is never used in place of a zero.
• uses_rendered_images: false appears on every block that could in principle have used a render; it is always false in shipped output.

The mood evidence

The evidence file docs/superpowers/evidence/static-mood-table-2026-10-16/static-evidence.json.gz holds one row per bundled preset. Its top-level fields:

Field	Meaning
schema	projectm-tv-static-mood-evidence
version	1
producer.branch	The research branch the analysis ran on (feat/predictor-static-output-bounds)
producer.commit	The exact research commit
producer.calibration_model_sha256	SHA-256 of the calibration model used
producer.policy	static source analysis only; no ruler or simulation values
rows	One object per preset, sorted by name

A row, here for Fire and Brimstone:

{"name": "Various - Fire and Brimstone.milk",
 "sha256": "ddc3d588…cf914",
 "options": ["bright", "vivid", "warm", "liquid"],
 "palette": ["green", "red", "yellow"],
 "dominant": null,
 "values": {"luminance": 0.4910, "saturation": 0.8042, "coloured_share": 0.8259, "warm_cool": 0.4487,
            "flow": 0.0952, "frame_difference": 0.0342, "jumps": 0.0, "activity": 0.3092}}

Field	Type	Meaning
name	string	The preset file name, as in presets.idx
sha256	string	SHA-256 of the preset file the analysis read. CI fails if the bundled file no longer matches
options	list	Mood option ids decided by the static analysis, without colour ids: Energy (chill, normal, intense), Brightness (dark, balanced, bright), Colour (monochrome, pastel, vivid, warm, cool), Movement (still, drifting, fast), Style (kaleidoscope, tunnel, liquid, geometric, particles, fractal), Comfort (no_strobe)
palette	list	Hue families with at least 10% of the static display palette, in order of share. Empty for grey presets
dominant	string or null	The main colour family, or null when no colour dominates (rules)
values.luminance	0-1	Estimated mean brightness of the picture
values.saturation	0-1	Estimated mean saturation
values.coloured_share	0-1	Estimated fraction of the picture that is coloured rather than grey
values.warm_cool	−1 to +1, or null	Estimated balance of warm (red, orange, yellow) against cool (cyan, blue, purple) hues; null when there is no colour
values.flow	screen widths per second	Estimated motion speed
values.frame_difference	0-1	Estimated mean change between consecutive frames
values.jumps	per 10 s	Estimated whole-screen brightness jumps per 10 seconds
values.activity	0-1	Energy percentile against the frozen activity reference; 0.31 is calmer than 69% of the reference presets
error	string	Present only if the analysis failed; the preset then gets no options. The shipped export has none

The Mood table line

attributes.tsv turns each row into one tab-separated line:

Various - Fire and Brimstone.milk   bright,vivid,warm,red,yellow,green,liquid   0.4910  0.8042  0.8259  0.4487  0.0952  0.0342  0.0000  0.3092

The columns are name, options (definition order: Energy, Brightness, Colour, Palette, Main colour, Movement, Style, Comfort; Main colour ids are main_<family>), then luminance, saturation, coloured_share, warm_cool, flow, frame_difference, jumps, activity with four decimals (- for unknown). The header lists every group and option (#group, #option) and the columns (#columns); the app builds the Mood panel from the header, so new groups need no code change.

manifest.json records schema, version, source: static-analysis, static_evidence_sha256, producer (copied from the evidence), definition_sha256, presets_idx_sha256, attributes_sha256, preset_count, measured_count (rows with any value), analysis_errors and option_counts.

The recipe

The recipe is the engine-neutral description of one preset. Its top-level fields:

Field	Meaning
policy	preset-visual-recipe-v1
preset	name and sha256 of the preset
conventions	The conventions above
description	A generated prose summary of constructions and colour ingredients (about 3 KB)
render	How the picture is built: feedback, layers, composite, mechanisms, audio, randomness, mirrors, textures, palette
display_estimate	Calibrated estimates of what the engine shows, with intervals
facets	Mood options and values for this preset
intent	What the preset is for: named effects, audio reactivity, palette and energy character
mood	An older approximate feature vector, kept for reference (not used by the Mood table)
limitations	What the recipe does not capture
uses_rendered_images	Always false
program	Only with --include-program: the parsed equations and shaders for an exact port

render.feedback

Field	Meaning	Fire and Brimstone
decay	The decay applied each frame, when the engine applies it (null with a custom warp shader, which replaces decay)	null
stored_level_halving_steps	Frames for a stored 8-bit level to halve under decay	null
stored_level_floor	The highest stored level that never decays (the faint residue)	null
warp_patterns	Recognised motion patterns of the mesh warp, each with its control, value range and direction (see below)	radial zoom profile
warp_controls	Per control (zoom, zoomexp, rot, warp, cx, cy, dx, dy, sx, sy): uniform (same at every vertex), value_range, and terms (constant, time, audio or per-pixel parts)	zoom constant 0.99853
shader_advection	Sample offsets in the warp shader that move content by the image itself: stage, offset_range_uv (as fractions of the screen), evidence	±0.06 in u and v
ray_sampling	Set when the warp reads one source point per ray through a centre (every frame becomes converging rays)	null
warp_detail_response	The warp evaluated on grey multi-scale noise: mean_luma, mean_saturation, coloured_share, hue_shares, and its input (kind, level 0.5, amplitude 0.2, smoothing 1/4/16 px, 480×270, t = 5 s)	luma 0.51, coloured 0.09
warp_detail_gain	The warp on dim noise: gain above 1 means small structure grows each frame; flat_out is the output level for flat input	gain 1.08
flat_field	The state a flat canvas settles to under the warp, when known	null
warp_transfer	The warp evaluated on flat colours (see below)	channel-selective
evidence	Pointers to the summary facts used
steady_state	The feedback equilibrium (see below)	yellow

Each warp_patterns entry has a pattern name (radial_zoom_profile, uniform_rotation, spatially_varying_rotation, uniform_translation, translation_component, …), the control it comes from and its ranges. For a zoom: value_range, zoomexp_range, effective_zoom_range, content_motion ("contracts toward the screen centre" or "expands …"), percent_per_step_range (−0.149 to −0.145 for Fire and Brimstone: 0.15% inward per frame; −9.9 to −1.9 for the particle storm) and audio_inputs (bands that change it). For a rotation: value_range_radians_per_step (±0.315 rad per frame for the particle storm).

warp_transfer has these fields:

Field	Meaning
gain_rgb, offset_rgb	Per-channel gain and offset for uniform content per frame: [0.994, 0.97, 0.85] and [0.006, 0.03, 0.008]
linear	Whether output is proportional to input over all probed levels
halving_frames	Frames for uniform content to halve (null when it does not fade to half)
steady_level_from_full	The level a full-brightness canvas settles at
channel_selective	True when channels persist differently; read per_channel then
per_channel	r, g, b each with gain, linear, halving_frames, steady_level_from_full. Fire and Brimstone: blue halves in 4.3 frames and settles at 0.053, red and green settle at 1.0
wrap	Set when the warp takes a modulo: levels shift by shift_per_frame and wrap every cycle_frames (continuous colour cycling)
retention_by_level	For nonlinear transfers: the fraction kept at each level (above 1 means dim levels are lifted)
levels, rgb	The 11 probe levels and the RGB output for each
source	exported warp shader evaluated on flat colours

steady_state has these fields:

Field	Meaning
canvas_local_rgb	Equilibrium canvas colour where elements keep drawing
display_local_rgb	The same through the composite (what you would see there)
display_spread_rgb	Equilibrium if the drawn energy were spread over the whole screen
hue_family	The equilibrium's hue family (yellow for Fire and Brimstone)
composite_model	How the composite was applied (composite_transfer, legacy controls, …)
substitution_count	How many unknown element colours were replaced by neutral priors
level_note	Reminder that the level is a lower bound for bright presets; use display_estimate for brightness
source	feedback_equilibrium steady state (source facts, no frames stepped)

render.layers

One entry per stage or element, in drawing order: mesh_warp, builtin_wave, wave_0…wave_3, shape_0…shape_3, shader_warp, shader_composite.

Field	Meaning
id, stage	Layer id and stage (mesh_warp, drawing, warp, composite)
mechanisms	Mechanism names recognised in this layer
evidence	Pointers to the element facts
constants	Constant settings: for a shape sides, instances, x, y, rad, ang, inner colour r g b a, outer colour r2 g2 b2 a2, border colour, tex_zoom, tex_ang; for the mesh warp zoom, zoomexp, rot, sx, sy, warp
controls	Per animated control: control name, control_unit, curve_kind (constant, periodic, audio, …), constant_value, nominal_value_range, maximum_absolute_control_rate_per_second, and the expression graph
colour	mode_code, palette_diversity, constant_rgb and guaranteed_visible (whether the element is certain to show)
point_cloud	For custom waves: points_per_frame (512), sample_grid, colour_resolved, authored_rgba, hue_family_shares, mean_rgb, mean_alpha, mean_saturation, positions_closed_form, position_bounds, frame_inputs (per-frame state the points read) and the per-point programs

render.composite

Field	Meaning
flat_transfer	Displayed colour for each of 11 flat canvas levels, at three times (2, 5, 8 s): rgb (mean), rgb_spread (movement between times), time_varying, nonfinite_share (pixels with undefined output), output_undefined, flat_areas_only (relief composites only show edges)
detail_response	The composite on grey multi-scale noise: mean_rgb, mean_luma, mean_saturation, coloured_share, hue_shares and the coloured_rule used (saturation ≥ 0.25 and max channel ≥ 0.10)
colour_matrix	level (0.3), black_rgb and columns_rgb: displayed RGB ≈ black_rgb + Σ (input_c / level) · columns_rgb[c]. An identity-like matrix (Fire and Brimstone) passes colours through; off-diagonal entries mean channel mixing
per_load_random	Composite switches drawn once per load (rand_preset), each branch with probability_true
hue_shader	Set when the composite multiplies by projectM's hue shader

render.mechanisms, audio_bindings, random_redraws, mirrors, random_textures

• mechanisms lists every recognised construction with mechanism, stage, parameters and evidence. Examples: builtin_wave_primitive (mode, dots), polygon_shape_primitive, custom_wave_primitive, point_cloud_primitive, radial_feedback_transform, blur_difference_feedback, gradient_driven_advection, image_driven_advection, noise_driven_advection, gradient_relief, angular_mirror_fold, cartesian_mirror_fold, repeat_tiling, polar_radial_sampling, nonlinear_feedback_map, complex_quadratic_recurrence.
• audio_bindings records which element (layer) is driven by which input (bass, mid, treb, their _att variants, waveform) into which target (opacity, size, position, colour, …), with linear_gain when it is a simple multiple. The particle storm binds bass and mid to shape opacity.
• random_redraws lists per-frame state redrawn with rand when a condition fires: states, feeds (the q values that carry the draws to shaders and elements), event_times_s and period_s when known, and a note. Fire and Brimstone redraws q2 to q7 at unknown times, so colours that read them are a random draw.
• mirrors has operators, flip_directions, fold (number of mirror sectors), symmetry_score (0 to 1, an upper estimate) and conditions explaining why the symmetry is exact or not.
• random_textures lists the shader slots that sample a random library texture (sampler_rand00…), with the candidate textures.

render.display_palette

Field	Meaning	Fire and Brimstone
hue_shares	Pooled share per hue family	green 0.36, red 0.28, yellow 0.27, orange 0.05, pink 0.03
palette	Families with share ≥ 0.10	green, red, yellow
dominant_hue	Main colour, or null	null
monochrome	True when the Monochrome classifier admits the preset (palette then empty)	false
dominance	Share of the top family	0.49
sources	Which sources contributed, each with weight and leading family	calibrated hue (24, green), steady state (1.5, yellow)
random_tint	True when colours are drawn per load	false
method	The pooling rule and weights

display_estimate

These are the calibrated estimates (From estimates to moods).

Field	Meaning
policy	source-calibrated-display-estimate-v1
statistics	One entry per target, each {value, interval, oof_spearman, evidence, label_source}; classifier targets add threshold
activity	A display activity score (display-activity-v1): activity, activity_interval, bands, possible_bands (options within the interval), certain_bands, remapped. It is a separate, remapped score; the Mood table's activity value is the frozen-reference percentile of the calibrated frame difference, flow and jumps
reference	The ruler conditions the estimates are calibrated to
note	"Calibrated from static source facts; intervals are the 10th-90th percentile out-of-fold residuals; not a rendered measurement of this preset"

The 29 statistics:

Statistic	Meaning	Fire and Brimstone
mean_luminance	Mean brightness	0.49 (0.31-0.71)
mean_abs_frame_difference	Mean change between frames	0.034
mean_flow_screen_widths_per_s	Motion speed	0.095
global_brightness_jumps_per_10s	Whole-screen jumps per 10 s	0.0
mean_saturation	Mean saturation	0.80
warm_cool	Warm/cool balance	0.45
coloured_share	Coloured fraction	0.83
hue_share_red … hue_share_pink	Share of each family among coloured pixels	green 0.33, red 0.26, yellow 0.19
hue_top_share	Share of the top family (also sets the palette sharpening sharpening_power)	0.49 (power 1.5)
no_strobe_probability	Probability of fewer than 1 jump per 10 s	0.926 (just under the 0.93 gate)
facet_<option>_probability (12)	Probability per option for dark, bright, monochrome, pastel, warm, cool, still, chill, intense, balanced, drifting, normal, each with its threshold	warm 0.82 (threshold 0.51), chill 0.59 (0.69)

In the example, Fire and Brimstone's No strobe probability (0.926) falls just short of the 0.93 gate, so the preset is not admitted to No strobe even though its jumps estimate is 0. Its activity (0.31) satisfies the Chill rule (≤ 0.35), but the Chill classifier gives 0.59, below its threshold of 0.69. The classifier decides, so the preset gets no Energy option. Bright, Warm and Vivid come from the Bright and Warm classifiers (0.69 and 0.82, above their thresholds) and the Vivid rule (saturation 0.80 ≥ 0.5). The design puts precision first: the render does look calm, but the evidence was not strong enough to say so.

facets

facets holds the mood options for the preset, in these fields: policy (preset-facets-v1), definition_sha256, evidence and style_evidence (what the options rest on), options, possible_options (options whose rule changes inside the interval), unresolved_options (options that need an unknown value), style, values (the eight values, each with its evidence) and measurement.

For a bundled preset the recipe tool reports the options of the bundled table; the static export removes the preset's identity before calling it, so the evidence always holds the static decision.

intent

intent describes what a person or another engine should rebuild:

Field	Meaning
effects	Named effects with parameters (catalogue below)
reactivity	Per audio route: band, layer, target, kind (brightness, shape, motion, colour), smoothed, linear_gain, response (continuous or threshold_or_clamped)
palette	warm_cool, saturation, hue_variety_bins, hue_cycle_per_s, background_rgb from source facts
energy	intensity and smoothness scores (0-100) and style tags from the older mood layer
guidance	"Rebuild the listed effects at any resolution and frame rate in a modern engine; keep the reactivity map … and the palette/energy character. Exact pixels are not a goal."

The effect catalogue:

Effect	Parameters	Meaning
trails	half_life_s, decay_per_frame, faint_residue	Feedback persistence. Particle storm: half-life 1.8 s
zoom_flow	percent_per_frame, direction	Mesh zoom toward or away from the centre
rotation_flow	radians_per_frame	Mesh rotation
drift	axis, pixels_per_second	Translation
radial_streaks	direction, half_life_frames, spiral, condition	Detail smeared along rays by a strong one-way zoom
radial_rays	screen_share, centre_xy, centre_moves, condition	Every frame becomes rays converging on a centre
level_cycling	shift and cycle	Colours cycle through the whole range (modulo warp)
kaleidoscope, mirror, mirror_symmetry	sectors	Mirror folds
tiling	mirrored	Repeated tiles
polar_tunnel	radial_mapping (e.g. reciprocal_radius)	Tunnel by polar sampling
swirl, rotation, spatial_rotation, angular_warp, periodic_wobble		Shader-level twists and wobbles
self_advection, reaction_diffusion_flow, noise_turbulence	stage	Content moved by itself, by its gradients, or by noise
pattern_growth	channels, fine_levels, coarse_levels	Fine-minus-coarse blur that grows structure
relief	directions	Embossed edges
blur	taps and radii	Neighbourhood blur
recursive_copies	copies, scales	Several scaled copies of the previous frame
fractal_feedback, julia_fractal, mandelbox_fractal	construction	Fractal constructions
feedback_scale	scale	Shader-level zoom of the feedback
glow_spots, stripes		Periodic glows and bands
palette_map		The composite remaps levels to a palette
echo_double		Video echo: a second, scaled copy
frame_copy	layer, tex_zoom, tex_angle_rad, condition	A textured shape that redraws the previous frame
vector_grid		Motion vectors
audio_waveform	form, position, colour_cycles	The built-in waveform (two_channel_parallel_traces, stereo_xy_trace, …)
custom_waveform	points_per_frame, hue_family_shares, mean_alpha, position_bounds	A custom wave
parametric_curve		A custom wave recognised as a circle or spiral
shapes	sides, instances, radius, moving	Custom shapes
random_redraw	feeds, first_s, period_s	Values redrawn at random
per_load_variant	branches	A choice made once per load

mood (reference only)

An earlier research layer with 23 features (motion.speed_p95_vp_s, flash.coherent_transitions_hz, palette.effective_hue_bins, structure.symmetry, …), each with value, interval, evidence and evidence_paths, plus scores (intensity, smoothness, psychedelic, warm, cold) and suitability per listening profile (ambient-v1, techno-v1, gentle-home-tv-v1, …). Its own limitations say it is built on editable assumptions and not calibrated against viewers. The Mood table does not use it.

Expression graphs

Wherever the recipe carries an expression or programs entry, it is a dataflow graph:

{"root": 0,
 "nodes": [{"op": "constant", "dtype": "float", "args": [], "detail": {"value": 0.5}}],
 "resources_resolved": true, "complete": true}

nodes is a list of operations (constant, input, arithmetic, comparisons, if, function calls such as sin, pow, texture lookups); args are indices into nodes; root is the result. complete is false when part of the program could not be exported, and resources_resolved is false when it reads something the analysis could not resolve (a random texture, memory written elsewhere). The core recipe removes these graphs; the full recipe keeps them so a port can evaluate controls exactly.

The prediction record

Every prediction image has a prediction.json record next to it:

Field	Meaning
policy	sdk-prediction-illustration-v1
summary_sha256, illustrator_sha256, image_sha256	Hashes of the input summary, the illustrator code and the image
declared_time_seconds, history_steps	5.0 and 150
drawn_components	Elements drawn (e.g. wave_0, wave_1)
limitations	What the drawing omits or approximates (full list)
uses_rendered_images, reads_preset_source	Both false
evaluates_exported_closed_form_graphs, graph_evaluation_scope	Which exported programs were evaluated
declared_audio_sha256	Hash of the declared audio

The self-comparison record

Field	Meaning
checkpoint, preset	Checkpoint id and preset
frozen_before_capture	Hashes of the prediction, its record, the summary, the recipe and the look description, taken before the render
capture_result_sha256	Hash of the capture record
review_measurements_v1.prediction_luma, screenshot_luma	Single-frame mean brightness
mean_rgb_prediction, mean_rgb_screenshot	Single-frame mean colour
blurred_luma_correlation	Layout agreement of the two single frames
time_average_frames_60_149	luma, saturation (prediction, render) and the blurred_luma_correlation of the time-averaged pictures
mean_abs_frame_difference_60_149	How busy each is (prediction, render)
agreements, misses, verdict	The reviewer's findings