About

Short

RL Environments, Long-Horizon Evals, AI Training Gyms

Long

Idler builds reinforcement-learning environments for evaluating and training models. Take one off the shelf or have one built to spec. They cover coding, reasoning, and the knowledge work people are paid to do.

Thesis

Reality selects for intelligence: model the world, adapt as it changes, act without certainty. Idler builds the worlds where agents learn that, turning space, time, and what-if into environments that shape how a model reasons and recovers.

Mark

Seen down its body diagonal, the Shell flattens to a hexagon of three rhombi, and the three edges nearest the eye meet in a Y. The rhombi are its faces. The Y is the seam between them. Neither is drawn: both fall out of the geometry.

Horizontal · light
2000 × 389 SVG PNG
Horizontal · dark
2000 × 389 SVG PNG
Stacked · light
1200 × 941 SVG PNG
Stacked · dark
1200 × 941 SVG PNG
Mark · light
1024 × 933 SVG PNG
Mark · dark
1024 × 933 SVG PNG

Favicon

The symbol alone, in a square tile. The tile carries the silhouette, so the mark holds at 16px against light or dark browser chrome.

idler favicon, emerald
Color
idler favicon, dark
Dark
idler favicon, light
Light

Shell emoji

The Shell in a circular tile, for avatars, reactions and anywhere the square favicon would read as a chip. Transparent ground, so it sits on whatever it lands on.

Shell emoji, emerald
Color
PNG
Shell emoji, dark ground
Dark
PNG
Shell emoji, light ground
Light
PNG

Clearspace

One module, 1x, taken from the lockup itself: the gap between mark and logotype is 0.5x, so 1x is twice it. Keep 1x clear on every side. Nothing sets inside it: no type, no rule, no edge of an image.

1x 0.5x 1x
Horizontal · 1x outside, 0.5x between
1x 1x
Stacked · 1x on every side

Color

Five accents. One runs at a time.

Color logic

Color answers two questions. What is happening, which is state, and who this is, which is the model. It never answers how much: a magnitude is carried by the length of the bar it fills, and the color of that bar says whose bar it is.

A model takes its accent from where it sits on the board, and keeps it across every view. That is the whole reason to spend color on identity: the reading carries from the leaderboard to the skill table to the scatter without a legend, because the green row and the green column and the green point are the same model. Six models and five accents, so the sixth wraps and is marked striped rather than given a hue the identity does not have.

The ramp stays for the one case identity does not cover: a figure that belongs to no model. A share, a coverage, a count against a total takes a step of it, and the hue never changes.

One model, three geometries

The same four models, sorted the same way, in the three shapes the reader draws them in. Nothing carries a color that belongs to the measurement; the color belongs to the model, so a row, a column and a point can be read as one thing.

Table · value
Opus 598.1%
Fable 578.8%
Kimi K343.7%
Grok 4.617.8%
Matrix · state
Opus 5
Fable 5
Kimi K3
Grok 4.6
Graph · model
Opus 5
Fable 5
Kimi K3
Grok 4.6

Grayscale

Twelve steps, paper to ink. Structure comes from the middle.

Gradient

One gradient, and it is light, not pattern: white falling to n02 down the vertical. It sits behind a figure so the field reads as lit rather than filled. It never runs under text, never carries a control, never takes a hue. On the archive it mirrors, surface to ground, so anything built on it survives the theme flip.

0%FFFFFF100%E0E2E1
RaisedOn the page ground

-1 -1 1px #FFF .95
3 4 6px #1B1D1C .34

PressedOn the page ground

2 3 4px #FFF .95
-1 -2 4px #1B1D1C .30

Mark onlyOn the page ground

Raised pair, nothing beneath it

Opacity

Opacity does not make a new gray, it puts the same gray further back. Six steps, all of them ink on the paper plate. If a value here would land on a ramp step, use the ramp step: alpha is for things that sit over something you still need to see.

Typography

Faces

Two families, no serif, and the stack behind each one. A face is requested by its token, never by its name, so a swap happens in 08 and nowhere else.

PreviewType familyTokenStack
Aa Tekio Grotesk --institutional Tekio Grotesk, Helvetica Now Display, Helvetica Neue, Helvetica, Arial, sans-serif
Aa PX Grotesk --data PX Grotesk, Geist, Helvetica Now Display, Helvetica Neue, Helvetica, Arial, sans-serif

Sizes are given in rem and in pixels at a 16 px root. A step is named by the token that sets it; there are two faces and the token names which one.

Heading

Three steps of one face. Hero is the campaign line, Heading is every H1 through H3, Statement is the claim.

PreviewStepType familyTokenFont weight Font sizeLine heightTracking
Aa Hero Tekio Grotesk --institutional Regular 4.5 rem / 72 px 90% -1%
Aa Heading Tekio Grotesk --institutional Regular 3 rem / 48 px 90% -0.5%
Aa Statement Tekio Grotesk --institutional Regular 2.25 rem / 36 px 90% -0.5%

Body

One step. Reading text sets at a single size and is spaced by the measure it sits in, not by a second body style.

PreviewStepType familyTokenFont weight Font sizeLine heightTracking
Aa Body Tekio Grotesk --institutional Regular 1 rem / 16 px 100% -0.5%

Metric

Figures, pass@k and IDs. Always the data face, always uppercase, always tabular so a column of numbers lines up.

PreviewStepType familyTokenFont weight Font sizeLine heightTracking
Aa Numeral PX Grotesk --data Regular 1 rem / 16 px 90% 0%

Caption

Eyebrows, meta, column heads and button labels, set in the body ink rather than a faded one: a term list is stated, not whispered.

PreviewStepType familyTokenFont weight Font sizeLine heightTracking
Aa Caption PX Grotesk --data Regular 1 rem / 16 px 100% 1.5%

Atoms

The primitive is a substance. A halftone dot is ink, and these six units are the material the Shell is made of. So the category is Matter, and Matter carries two axes: constituents, which are the three forms, and states, which are solid or opened. Three by two is six. A seventh unit would have to declare what it was made of, and there is nothing left to declare.

Matter
Circle
Triangle
Square
Fill
Protoncircle
Electrontriangle
Neutronsquare
Hole
Aperturecircle-hole
Portaltriangle-hole
Socketsquare-hole
The six units

Proton +Y

A circle has no vertices and no orientation. Rotate it by any amount and it lands where it started, which is why no turn can leave it misaligned.

Electron +X

Three sides and one apex. Odd-sided, which nothing else here is, and the one unit that can point. Drawn 12% larger so its weight matches the rest.

Neutron +Z

Tiles at the grid’s own pitch and leaves no residue at the cell boundary. Every other unit sits inside its cell; this one becomes the cell.

Aperture −X

An annulus. The opening holds one distance from the boundary the whole way round, and admits without pointing anywhere.

Portal −Z

The hole sits at the triangle’s centroid, which falls below the centre of the bounding box. That offset gives the passage its direction.

Socket −Y

A square with the middle taken out. The frame is the whole unit, and the empty centre is where something else sits.

The Shell

Idler’s visual model of an environment: a closed, persistent structure in which agents move, act, and are evaluated.

Shell

The canonical build, and what /cube.html serves with no query at all. One form per face; the mark is the negative space between them, never a line drawn on top.

Transmissive

?v=trans. The three faces turned away are admitted at a fraction of strength, so the far side reads through the near side and the Y carries a trace. The far side tapers in size rather than fading, which keeps it reading as depth.

Closed field

?v=closed. The halftone at full coverage: every unit sized until neighbours overlap and seal, so the panels read as material rather than as dots. All six units remain; only their size changes.

Plate

?v=plate. The closed field is a screen, so a surface is sampled at each mark’s position and sets its value. Folds and sheen arrive as tone across the lattice.

The Shell Lab

The four builds above are the Lab’s own presets, rendered from it. A host selects one with ?v= and there are no forks besides these. Open it full screen to inspect those surface rules.

The Lab needs room for an editor beside a stage, so on a narrow screen it opens in its own tab rather than embedding.

Motion

Still first, alive second. The Shell moves in discrete turns: one layer or the whole body moves, lands, and holds before the next turn begins. Each frame below names the visible move.

Three faces of the Shell meeting around the central aperture2.1sResting cornerThree faces meet around the central aperture. No layer is moving.
The Shell's right outer layer rotating away from the body3.3sRight layer turns outThe right outer layer rotates 90° away from the body around its vertical axis.
The Shell closed again in a corner-facing pose5.7sCorner closesThe turned layer lands on the grid. Three faces meet again in a closed corner.
The Shell's upper-left outer layer rotating away from the body6.9sUpper-left layer turnsThe upper-left outer layer rotates 90° away from the body.
The complete Shell rotating toward a different corner7.5sWhole body rotatesThe complete Shell rotates toward a new corner-facing orientation. Its layers stay locked together.
One face of the Shell landing square to the viewer8.1sFace lands forwardThe whole-body rotation ends with one face square to the viewer.
The Shell's top outer layer rotating upward19.5sTop layer turnsThe top outer layer rotates 90° upward, opening a horizontal gap.
An upper horizontal slice of the Shell rotating above fixed lower layers29.7sUpper slice turnsAn upper horizontal slice rotates 90° while the lower layers stay fixed.

Applications

Buttons
BrowsePrimary ContactSecondary
Input
Tags
Long-Context Memory Efficiency MCP Tool-Use Reconciliation Instruction Following
Tabs · section strip
The panel the strip heads. The strip is its lid, so the two share an edge.
Tabs · metric switchers
Column filter
Table · leaderboard
Model Pass@1 Cost / rollout
Claude Opus 5 69.0% ± 0.6 pp $5.02
GPT-5.6 Sol 55.7% ± 1.9 pp $2.30
Claude Fable 5 54.7% ± 0.8 pp $17.04
Grok 4.6 39.4% ± 0.7 pp $3.61
Kimi K3 30.6% ± 0.9 pp $7.83
Gemini 3.6 Flash 16.1% ± 0.5 pp $2.29
States
Passed Running Pending Available Failed
Cards · environment and task
Warehouse operations
Brightwell5 Envs
Long-ContextReconciliation
0002
Reconcile two divergent inventory snapshots and file the discrepancies to the migration tracker.
Claude Fable 550%
12 Rollouts6 Passing0.58 avg score
Cards · collection, criterion and reference
Filed to the tracker67%
Cards · blog
Section nav
Model chip
Claude Opus 5 Claude Fable 5 8 runs
Table · skill
SkillTasksOpus 5highFable 5highKimi K3medium
Planning4278.8%60.8%28.7%
Unstated methods3140.3%38.0%21.9%
Numerical precision2667.2%50.5%43.5%
State management3868.5%69.4%29.3%
Data quality2254.3%39.2%17.8%
Tool use5798.1%88.5%94.9%
Table · matrix
EnvironmentC1C2C3C4C5
ShelfLife Enterprise
Corporate Law
Synthetic Enterprise
Table · document
PropertyWhat it records
CadenceThe work is currently done by a human at a recurring cadence.
LeverageThe failure modes of the task are high leverage and high risk.
DurationHours of human input to complete the workflow end to end.

Tokens

References

What the system is measured against. Hover a source image to see the six-atom reading it informed. The full argument is on the references page.

The lineage

Where the marks come from: systems run to exhaustion, and nothing used that is outside the vocabulary.

Manfred Mohr’s Cubic Limit II, a divided cube projected as black lines on white

Manfred Mohr

Cubic Limit II ↗

The cube as a finite alphabet, exhausted rather than drawn.

Sol LeWitt’s Incomplete Open Cubes installed as 122 white structures with drawings and photographs

Sol LeWitt

Incomplete Open Cubes ↗

Enumeration over composition: keep whatever the system honestly produced.

Cover of Karl Gerstner’s Designing Programmes

Karl Gerstner

Designing Programmes ↗

Design the programme; the poster is one of its outputs.

Vera Molnár’s computer-assisted drawing Des lignes, pas des carrés

Vera Molnár

Des lignes, pas des carrés ↗

Controlled disorder. A perfectly regular field goes dead on the page.

A floral wallpaper pattern rendered entirely in colored Ben-Day-style dots

The halftone screen

Ben-Day screen ↗

One mark at one size. Density carries the tone.

Otl Aicher’s athletics and football pictograms installed at the 1972 Munich Olympic stadium

Otl Aicher

Munich 1972 ↗

A whole sign system from one grid and a fixed set of angles.

A monochrome particle field from the Photons motion-reference channel

Photons

Photons channel ↗

Point fields: density does the shading, and motion resolves to one origin.

The environment

How a controlled world gets built, lit, and read from outside it. These set the room, not the marks in it.

James Turrell’s As Seen Below installation, a circular opening above a luminous green chamber

James Turrell

As Seen Below · ARoS ↗

Occlude the source and the structure appears. The aperture is the object, not an absence in one.

The interior of Rothko Chapel, with dark paintings surrounding a central bench

Rothko Chapel

Houston, 1971 ↗

The room is the work. You stand inside the system rather than in front of it.

Opening picnic frame from Charles and Ray Eames’s Powers of Ten

Powers of Ten

Charles & Ray Eames, 1977 ↗

One subject held at every scale. The zoom is the argument; the frame never cuts.

Technicians and semiconductor fabrication equipment inside a NASA cleanroom

The cleanroom

NASA semiconductor lab ↗

Precision belongs to the environment before it belongs to the part. Control the room and the yield follows.

The green and copper shell of a Chrysochroa baudoni jewel beetle, viewed from above

Beetle shell

Chrysochroa baudoni · Bernard Dupont ↗

The colour is structural. Nothing in the shell is pigment: stacked layers bend the light and the hue arrives from the geometry, shifting with the angle you stand at. This is where the iridescence comes from.

Bands of cyan, amber, violet, and white produced by thin-film interference on a soap bubble

Iridescent texture

Thin-film interference · Pksois23 ↗

Colour is generated by film thickness and angle. The surface changes because its optical structure changes the path of light.

Cover of Edward Tufte’s Envisioning Information

Edward Tufte

Envisioning Information ↗

Maximise data-ink. Every mark has to carry something a reader can verify.

A cybernetically modified figure shown in a luminous green anatomical cutaway in Cyberpunk: Edgerunners

Cyberpunk: Edgerunners

Cyberpunk: Edgerunners ↗

Two small marks against one enormous body. Scale delivered as feeling, with no number on screen.