Prove the checks can fail before paying to scale (source)

The decision

Before paid test runs scale up, each automated check must first reject a deliberately broken input. The receipt for that free demonstration is tied to the version of the checking code, so a changed check needs a new demonstration and an old green receipt cannot authorize spending.

That adds calibration and maintenance work. In return, a passing test means something stronger: the instrument has already shown it can catch its target failure.

From the source

Spend gate. Unsuccessful demonstrations and a changed helper both block progression:

  if (!parsed.allDemonstrated) {
    throw new Error(
      `Spec 184 D2 gate: the last can-fail demo did not demonstrate every probe red (${file}).`,
    );
  }
  if (parsed.probeHelperSha256 !== currentHelperSha256) {
    throw new Error(
      `Spec 184 D2 gate: render-seam-probe.ts changed since the can-fail demo ` +
        `(${parsed.probeHelperSha256.slice(0, 12)} → ${currentHelperSha256.slice(0, 12)}). ` +
        `Re-run the demo before spending.`,
    );
  }

Grayscale experiment, September 2026. The ungraded control’s chroma means already passed the tolerance of 2; within-frame chroma range separates it from the desaturated output. Only the fixture filename is redacted:

| reading | delivered `bw` render | ungraded CONTROL (`[REDACTED: synthetic control filename]`) |
|---|---|---|
| mean UAVG / VAVG | **128.000 / 128.000** (|Δ| 0.000) | 127.59 / 127.55 (|Δ| 0.41 / 0.45) |
| mean per-frame chroma range (UMAX−UMIN / VMAX−VMIN) | **0.00 / 0.00** | 29.75 / 29.02 |

Mutation run. Disabling both range comparisons made the targeted regression fail:

Test Suites: 1 failed, 1 total
Tests:       1 failed, 37 skipped, 4 passed, 42 total
Snapshots:   0 total
Time:        6.695 s, estimated 11 s

Reading the result

Check	Outcome	Why it matters
Colored control	Rejected by the range check despite passing the mean check	Exposes the original false positive
Same clip genuinely desaturated	Accepted	Protects against a check that rejects everything
Range comparisons disabled	1 failure, 4 passes, 37 skips	The targeted regression depends on the required behavior
Original comparisons restored	42/42 tests passed	The restored suite returned to green

What this shows

A check that tells the two cases apart in both directions, and a regression test that fails when the fix is removed. A separate 9 October receipt records 29 of 29 constructed-failure demonstrations bound to the helper hash. The pixels were read from encoder output after a delivery timeout, so they show the grade’s effect, not delivery timing.