E03: calibrate repetition counts and winner-reversal risk

[hanklin9188]

Parent

• Formal evidence program: execute E01–E30 by dependency and evidence level #5

What to deliver

A versioned repetition policy justified by measured confidence stability rather than an arbitrary fixed count.

Blocked by

• PR Establish EdgeFlow engineering and experiment governance #4 must merge.
• Final acceptance depends on E01 (E01: calibrate CUDA, wall-clock, and end-to-end timing boundaries #6) so the reference timing boundary is trusted.

Procedure

• Produce at least one stable 200-request reference block.
• Repeatedly subsample 10, 20, 30, and 50 measured rows.
• Evaluate median estimation error, confidence-interval width, and winner-reversal rate for a matched comparison.
• Calibrate separate classes when needed: kernel microbenchmark, single-request engine, online p95/p99.

Required artifacts

• raw 200-request block;
• subsampling seeds and code;
• stability and reversal distributions;
• repetition_policy.json or equivalent versioned artifact;
• validator/config update;
• report and verdict.

Acceptance criteria

• The selected engine minimum keeps winner-reversal risk below 5% for the calibrated comparison class.
• Kernel and online-tail minima are separately justified or explicitly inherited with limitations.
• The validator consumes the versioned policy rather than retaining unexplained hard-coded counts.
• Fewer-than-required fixtures cannot receive PASS.
• Increasing repetitions is not used to average over systematic thermal drift.

Permitted claim after completion

The chosen repetition minima meet the calibrated stability target for the tested workload and environment class.

Prohibited claim

The selected count is not automatically sufficient for every model, tail quantile, or future software stack.