How it works › Chart style guide
Chart style guide
One quiet, consistent style for every graph, so readers learn it once. The data is the only thing allowed to be loud. Every chart on the site is drawn by assets/charts.js, and the email and share-card images follow the same rules.
Palette
Each hypothesis has a fixed color everywhere: on the dashboard, in charts, on cards and in email. Colors follow the hypothesis, never its rank, and the order is always A, B, C, D, D-open. Both sets pass a colorblind-safety check (lightness band, chroma floor, deuteranopia and tritanopia separation, and contrast), and dark mode uses its own validated steps rather than an automatic flip.
| Series | Light | Dark | Used for |
|---|---|---|---|
| A | #B8700C | #C4861A | Hidden AGI |
| B | #00897B | #139A8C | Hidden self-improvement |
| C | #9150B8 | #A36ED0 | Covert AGI actor |
| D | #C8413A | #DC564A | Covert government influence |
| D-open | #3569D4 | #4F82DC | Open government influence |
| Accent | #B8700C | #D9A441 | The one thing a single-series chart is about: highlights, forecast bars |
| AGI part status | #747B83 | #9399A0 | Ink pips plus a word: Far, Partial, Close, Met. The pips show distance to a published test, not danger: never a status colour (green, amber, red), never the alarm's glyphs, never a hypothesis colour. Distance meters use --prog on --grid, linear on the measure's own scale with an ink tick at the bar; growth measures add a labelled log axis inside the drill-down only. Projections of a measured trend are dashed ink with an ink 95% band. |
| Index decomposition bar | A, B; --muted; hatched --ink | the same tokens | The one place the Index shows hypothesis colours, because its parts are hypotheses: A in A's colour, B outside A in B's colour; C or D outside A in neutral ink-grey, because it mixes two hypotheses and must not take either's identity. A hatched ink segment closes the gap to the published Index when its parts, rounded to 0.05, sum below it; the key says 'parts sum to 2.3; shown as 2.5 on our rounding grid'. Fixed 0–5% scale with '0' and '5%' printed in ink and an ink end tick. A run with no breakdown draws one neutral segment. |
Status colors are reserved for signals and the fire alarm and always come with an icon and a word. Signals (tripwires and Escape watch indicators): ○ Quiet (green), ◔ Open (amber), ● Confirmed (red). Fire alarm: ○ Normal (green), ◐ Watch (amber), ◉ Warning (red), ● Alarm (red). They never stand in for a hypothesis, and AGI parts never use them.
Text variants. Text on a background uses its own tokens so it passes WCAG AA contrast (4.5:1); the series colors stay for marks only.
| Token | Light | Dark | Used for |
|---|---|---|---|
--link | #2F62C8 | #6B96E6 | Links and the "credible report" tag |
--sA-text | #8A5608 | #C4861A | Amber text, such as the "expert opinion" tag |
--warn-text | #855C00 | #D4A72C | Amber status words, such as a met Watch condition |
--crit-bg | #B42318 | #B42318 | The alarm banner, behind white text |
--prog | #747B83 | #9399A0 | Distance-meter fill on the AGI part tracker (3.4:1 on --grid in light, 4.6:1 in dark) |
Rules
- Pick the form first. One number gets a stat tile, not a chart. Change over time gets lines. Durations get range bars. Probabilities get meters. Distance to a published test gets a meter on the measure's own scale.
- Thin marks. 2px lines with round joins. Dots at least 8px with a 2px ring in the surface color. Bars no thicker than 24px, with 4px rounded ends.
- Recessive chrome. Gridlines and axes are 1px solid hairlines one step off the surface, never dashed. Ticks fall on clean values.
- One axis. Never two y-scales on one chart. Different measures get separate charts.
- Identity is never color alone. Two or more series always have a legend, plus a few direct labels at line ends where they don't collide. Never a number on every point.
- Text stays ink. Labels and values use the text colors. The colored mark beside them carries identity.
- Emphasis over rainbow. When one thing matters (disclosed by outsiders, our forecast), it gets the accent and everything else goes gray.
- Every chart is interactive and readable without the picture. Hover or tap shows a tooltip, and every chart has a "Show as table" view.
- Forecasts vs projections. Our forecast is a solid line through the three numbers we actually state (today, end-2030, end-2035), with a shaded judgment band: not a statistical interval, but roughly how far our number could plausibly move as evidence arrives. A projection of a measured trend is dashed, with the 95% band of the trend line and a lighter band where a single new measurement should land. We never project our own daily probability line: the dated forecasts (end-2030, end-2035) shouldn't drift predictably; the line for today is expected to rise along our own forecast. Outside forecasts are hollow rings, and those that only measure an announcement are left off the chart.
- Definition changes break lines. When our definitions change, a line of readings over time gets a dashed hairline labelled with the new definitions, and no series joins across it.
- Log scales for growth. Exponential trends go on a log axis labelled in human units (a workday, a work-week), so a straight line means steady doubling.
- Type. Headings in Source Serif 4. Everything else, including big numbers, in Public Sans. Aligned columns use tabular figures.
Glyphs: one meaning each
| Glyphs | Meaning, and nothing else |
|---|---|
| ○ ◐ ◉ ● | Fire-alarm levels Normal, Watch, Warning, Alarm (from data/alarm.json; unchanged) |
| ○ ◔ ● | Signal status Quiet, Open, Confirmed, for tripwires and Escape watch indicators. ◔ marks Open so that ◐ always means Watch. |
| … plus a word | AGI part status Far, Partial, Close, Met |
| ✓ ≈ – ? | Parts of a test: Passed, Overall figure only (doesn't count as passed), Not yet, Not measured |
| hatched ink | The Index bar's rounding segment: the distance from the sum of its parts to the published, grid-rounded Index |
| borderline | A borderline alarm value ("borderline, doesn't count"); it replaces ◐ in the condition table |
| ● ○ ■ ◇ △ | In tracker lanes: public result, result from an unreleased model, new test, on-trend date, coming up |
AGI part status scale
Four words, each with a published meaning. A part's status is set by its test's own marks, never by our judgment of danger.
| Status | What it means |
|---|---|
| , status Far, step 1 of 4 | No part of the test is passed and the test's own 'Partial' mark is not reached. |
| , status Partial, step 2 of 4 | At least one part of the test is passed, or the test's own 'Partial' mark is reached, but not its 'Close' mark. A part passed on the overall figure only (results by occupation unpublished) does not count as passed. |
| , status Close, step 3 of 4 | The test's own 'Close' mark is reached, but not the bar. |
| , status Met, step 4 of 4 | A public system passes every part of the test, as published by the test's named source. |
Each part's test publishes its Partial, Close and Met marks (test.threshold). A part's status is the highest mark the best public evidence reaches. A part passed on the overall figure only does not count as passed. Unreleased-model results never set a status.
The eight-part tracker
One row per part: status in pips and a word, a distance meter on the part's own scale, and a timeline lane (● public result, ○ unreleased-model result, ■ new test, ◇ on-trend date, △ coming up). This example draws two parts from a sample frozen on Oct 5, 2026, so the guide never changes daily; the live tracker is on Is AGI here?
The example draws with JavaScript.
The Hidden AGI Index dial
Retired on live pages from redesign v2 (definitions v2.0); frozen reports keep it. An eye whose iris has 100 ticks, one per percentage point, lit in the accent color.
Lines: change over time
Range bars: how long something lasted
Meters: a probability, with market odds as a hollow ring
Status board: signals
Forecast: stated numbers, judgment band, outside forecasts
Projection: measured trend on a log scale
Dashed line: the fitted trend, projected. Darker band: the 95% band of the trend line. Lighter band: where a single new measurement should land. Faded dots: above METR's reliable range, left out of the fit.