makepad/skills/makepad-game-director/references/evidence-manifest.md

Evidence manifest

What counts as proof. "It should work" is not on this list.

Levels of evidence, weakest to strongest

1. Compiles — cargo check -p <crate>. Necessary, never sufficient. Makepad resolves the splash DSL at runtime, so a widget or style error survives cargo check and appears on launch.
2. Unit-tested — cargo test -p <crate> for a Cx-free crate. Strong for rules: physics, generation, scoring, netcode, merge.
3. Driven — a makepad-test test that launches the real app, sends real input, and asserts on widget_snapshot() or a log line.
4. Seen — a screenshot() from that run, inspected by you, at the actual target aspect ratio.
5. Measured — frame time, draw calls, instance bytes, entity counts, against apps/arcade/BUDGETS.md.
6. Shipped — --release build succeeds and runs.

Required by change size

Change	Required
Constant, copy, colour	1
Rule change in a Cx-free crate	1, 2
New mechanic	1, 2, 3
Visual or UI change	1, 3, 4
Performance work	1, 5 (before and after)
Full game or premium pass	all six, plus the scorecard

Capturing each

Unit test — the Cx-free crates are the point of the architecture:

cargo test -p makepad-game-sim
cargo test -p makepad-game-gen
cargo test -p makepad-game-blocks

Driven test — makepad-test gives TestApp with type_text, press_key, touch_down, screenshot, widget_dump, widget_snapshot, wait_for_log_contains:

use makepad_test::{makepad_test, TestApp, KeyCode};

#[makepad_test]
fn player_scores_on_pickup(app: TestApp) {
    app.wait_for_log_contains("game: ready");
    app.press_key(KeyCode::ArrowUp);
    app.wait_for_log_contains("score: 1");
    let shot = app.screenshot();
    println!("evidence: {}", shot.display());
}

Headless probe — for render and perf claims without a window:

cargo run -p makepad-arcade --example bigworld_probe --release

Measurement — read from the renderer, do not estimate:

renderer.frame_p90_ms();      // Option<f32>
renderer.quality_level();     // adaptive quality actually applied
renderer.quality_reason();    // why it dropped

RenderStats::instance_floats is the authority on instance stream width — the BUDGETS numbers were taken from the compiled shader, not counted by hand. Do the same.

Asset resolution — prove the scene is not silently on primitives:

assert!(renderer.model_is_loaded("kenney/racing/vehicle-truck-yellow"));

Screenshot discipline

• Capture at the target aspect ratio, not the default window.
• Capture the states that exist: menu, playing, paused, win, lose. A premium claim with only one screenshot is unverified.
• Look at them. An agent that captures a screenshot and does not describe what is in it has produced a file, not evidence.
• Name them for what they show: artifacts/shots/playing-desert-1920x1080.png.

Reporting

Keep detailed evidence and the scorecard in artifacts/. Keep the final answer focused on what works and what the remaining risks are. State plainly what you could not run and why — an unrun check reported as passing is the one failure mode that destroys trust in everything else in the report.