Trade bits for memory with a measured error budget
one move: measure compression error
Open the code pane immediately, then use the Spec tab, optional design question, tests, and artifact when you want them.
Finish each move before carrying the artifact forward.
Quantization maps a real range to finite integer levels; calibration data chooses the range and the error budget decides whether the result is safe.
Deliverable: State the system invariant in one sentence.Calibrate a symmetric scale, quantize and dequantize values, and report max absolute error and bytes saved.
Deliverable: Type the deterministic core.Zero values stay near zero, the quantized range is bounded, and the report exposes the chosen scale and error.
Deliverable: Record a normal case and a boundary case.Add an outlier, change bit width, and compare percentile clipping with min/max calibration.
Deliverable: Name the failure signal and its guard.A quantization card with calibration set, scale, error budget, and quality decision.
Deliverable: Save the contract, evidence, and handoff note.Zero values stay near zero, the quantized range is bounded, and the report exposes the chosen scale and error.
Add an outlier, change bit width, and compare percentile clipping with min/max calibration.
A quantization card with calibration set, scale, error budget, and quality decision.
Deriva-authored extension beyond the supplied PDF. The five moves are Deriva’s implementation contract for this project.