Skip to content

Latest commit

 

History

History
49 lines (36 loc) · 2.54 KB

File metadata and controls

49 lines (36 loc) · 2.54 KB

Statistical Model

Fixed design

Confirmatory experiment uses a predeclared sample. The default bracketed schedule records control-before, mutation, and control-after for each round. A seeded balanced mode uses AB/BA blocks. The research assurance preset fixes 20 bracketed rounds, observation-scoped response-cookie state, fresh observation clients, disabled retries, standard privacy, and full retained bodies within the response bound. Redirects and retries within one observation share that observation's state; isolated mode resets it before the next control or mutation observation.

The design does not stop early for a positive or negative result. Early termination is limited to cancellation, policy/resource failure, or rejecting invalid/unstable controls. Such work is INCONCLUSIVE.

Pair classification

Pairs are CHANGED, UNCHANGED, or INDETERMINATE. Status, selected response headers, body evidence, redirect semantics, retry outcome/error sequences, and final origin can be decision bearing. Comparator failure, incomplete body evidence, or unsupported text semantics cannot be treated as unchanged.

Intervals and verdicts

MRMA uses two-sided 95% Wilson score intervals for mutation and repeated-control change rates.

  • Influence requires confidently stable controls and the mutation interval lower bound at or above min_reproducibility.
  • No influence observed requires confidently stable controls, zero indeterminate mutation pairs, and the mutation interval upper bound at or below no_influence_threshold.
  • Missing pairs, invalid controls, confident control instability, incomplete fixed sampling, or comparator limits yield INCONCLUSIVE.

The output includes operating characteristics for the planned rounds before networking. MRMA does not call these intervals Bayesian probability, severity, exploitability, or proof of absence.

Exploration and confirmation

Exploratory ranking changes the hypothesis-selection process and therefore cannot reuse confirmatory language. impact emits a digest-bound candidate manifest. Confirmation consumes one candidate in a new research run with a new run ID and seed, checks baseline and authorization digests, and uses the fixed design.

Independence limitations

Fresh observation clients reduce shared cookie and pool state but cannot guarantee independent upstream routing, rate limits, WAF scoring, or server state. Reuse/per-arm/per-round modes disclose weaker connection-independence dimensions. Timing/backoff differences are quantitative context, not single-sample binary influence signals.