Freeze the prompt
Record the exact input, model, settings, output length, and date. No moving goalposts.
THE TEST BENCH / ADVERSARIAL TESTING
Looks right.
Does it hold up?
An evidence-first lab for stress-testing AI outputs. Copy a controlled challenge, examine the original generation, and separate confirmed failures from suspected anomalies and measurement limits.
FIELD RECORD / 001
DOCUMENTED TRIALS03
TEST FAMILIES02
RECORDED TRIALS03 TESTS
The verdict must survive contact with the evidence.
A convincing result is not the same thing as a verified result.
01 / REAL EXPERIMENTS
Inspect the failure, not the hype.
Original AI-generated clips. Exact prompts. Traceable findings. No manufactured benchmark scores.
Under review Confirmed
Finding labels reflect the currently available observations. A suspected continuity problem is not proof of a model failure.
02 / EVIDENCE WORKSPACE
Select a trial
Choose any card above to open its evidence record.
ORIGINAL MODEL OUTPUTVIDEO EVIDENCE
Original video is not available from this deployment. The prompt and review notes remain accessible.
Local playback only; the selected file is not uploaded.
Select a video above.
TEST SPECIFICATIONREPRODUCIBLE
Exact challenge text is retained as supplied. Generated videos may not satisfy every constraint.
EVIDENCE LEDGEROBSERVED ≠ INFERRED
REVIEW CHECKLISTMANUAL REVIEW
Checking boxes here is a reviewer aid only. No automatic pass/fail verdict is generated.
03 / CONTROLLED CHALLENGE BUILDER
Generate a test with a target.
Deterministic prompt templates—no API key or automatic grading. Change one stressor at a time.
GENERATED CHALLENGE / V0.1
The output includes an observable pass condition and an uncertainty gate. Its answer key should be independently reviewed before benchmarking.
04 / EVALUATION METHOD
Don't let the grader become the failure.
Observe before labeling
Log frames and timestamps; distinguish observed changes from imagined or hidden actions.
Classify the layer
Separate generation defects, detector mistakes, interface limitations, and insufficient evidence.
Retest one variable
Change one stressor, measure the effect, and preserve the original trial for comparison.
05 / EXPAND THE LIBRARY
One more video.
One more testable claim.
Add another trial to this browser session using a local MP4. The video never uploads anywhere; it stays on your device. The local addition is not permanent storage.
For long-term preservation, place the prompt, metadata, and evidence notes in the private project source of truth. Publish only approved media.