The full draw is the oracle

The truth is drawing every point. Anything faster has to reproduce its pixels exactly. Anything the raster cannot say honestly is said another way. Never silently.

Why#

Ten million points do not fit a naive draw loop in interactive time, so every plotting library reduces. Most of them sample. Every nth point, or one per bucket, and hope. Sampling is a lie with good posture. The rendered chart is a chart of the data, not the chart, and the difference is exactly the points that matter. The dropout spike in a loss curve. The outage gap in a latency series. The one saturated cell in an attention map. A reduction chosen for speed has already decided what the analyst gets to see.

The same failure has quieter forms. Interpolating across missing data invents readings that were never taken. Smearing an out-of-range point onto the border fabricates an extreme. Rounding a tick label manufactures a value. Each is a small lie the reader cannot detect. That is what makes it a lie.

The idea#

Name the oracle. It is the raster produced by drawing every point. Every optimization is proven against it, pixel for pixel. Not benchmarked as "visually close." It is asserted equal.

For lines, M4 makes this achievable. Keep the first, last, minimum, and maximum point of every raster column, and that column's pixels come back exactly (Jugel et al., PVLDB 2014). The pipeline buckets by the column each point renders into, so the reduction is a theorem about the raster, not an approximation of it. Gaps are path topology. Finite runs summarize independently, so a break inside a column stays a break.

For a grid denser than the raster, sampling is replaced by ownership. Every screen bucket owns the cells whose centers fall inside it, and shows a declared reduction over all of them. The mean box filter is the default. Max is there when the sparse spikes are the point. The caller chooses, and the default is disclosed. Nothing is dropped because a sampler happened to step over it.

What the raster cannot represent, the chart says out loud. NaN renders as a visible gap, never interpolated. Out-of-range data clips rather than smears. Quantization is disclosed, not implied.

Consequences#

• A performance claim and an honesty claim are the same claim. The fast path is the exact path, so there is no fidelity knob to trade away.
• The oracle test is a permanent fixture. Raw raster versus reduced raster, asserted equal at several frame sizes. A reduction that breaks it is a different chart, not an optimization.
• Extremes always survive. A spike one sample wide renders at every zoom, because min and max per column are kept by construction.
• A gap in the data is a gap in the chart, at every reduction level.
• Reductions that are not pixel-exact — LTTB, smoothing — exist as explicit stats the caller applies, never as silent defaults.

Not this#

• No sampling, no striding, and no "one point per bucket" as an automatic reduction.
• Do not interpolate across NaN, and do not drop it silently.
• No fast path whose excuse is "indistinguishable at normal sizes."
• The oracle test is not a benchmark. It is an equality, not a budget.

See The axes are the product for the label half of honesty, and Vision rule 3.

Witness#

One hundred thousand points with three one-sample spikes, reduced through the auto-inserted M4 and spliced here by the doc generator. The spikes survive because per-column extremes are kept by construction. Byte-equality against the raw raster is asserted in the crate's oracle test:

             100,000 points, three one-sample spikes
 8 ┤           ⡇                   ⡇                   ⡇
   │           ⡇                   ⡇                   ⡇
   │           ⡇                  ⢀⡇                   ⡇
 4 ┤ ⢀⣀        ⡇                  ⢸⡇⢀⡀                 ⣧    ⣀
   │⣰⠋⠈⢧     ⣀⡀⡇            ⡼⠉⢧   ⢸⣷⠋⠙⣆     ⣀          ⣿  ⢀⡞⠉⢧
 0 ┤⠁  ⠘⣆  ⢀⡞⠁⠹⡇    ⣰⠋⢳⡀   ⡼⠁ ⠈⢧  ⣸⠃  ⠸⡄   ⡼⠉⠳⡄    ⡼⠉⠳⡄⣿  ⡞  ⠈⢧  ⣠
   │    ⠸⣄⢀⡞   ⠹⡄  ⣰⠃  ⢳⡀ ⣰⠃   ⠈⠳⠴⠃    ⢹⡀ ⡼⠁  ⢳⡀  ⣸⠁  ⠹⣿ ⡼⠁   ⠈⠳⠴⠃
   │     ⠈⠉     ⢳⡀⣰⠃    ⠳⠴⠃             ⠙⠚⠁    ⢳⡀⣰⠃    ⠙⠚⠁
-4 ┤             ⠉⠁                             ⠉⠁
   └┬───────────┬───────────┬────────────┬───────────┬───────────┬
    0          20k         40k          60k         80k       100k

Spelled today#

stat::m4 is the public reduction; resolution auto-inserts the column-mapped form for line layers past four points per column. Cells::reduce chooses the grid reduction through the shared stat::Reducer vocabulary. The oracle test is large_lines_downsample_pixel_exactly_against_the_raw_raster in src/plot/tests/plot_tests.rs, toggling the reduction against the raw raster at several frame sizes. stat::lttb and stat::ewma are the explicit, opt-in inexact transforms. This section may rot; the rest must not.