Release v1.2.0

Added

• Add deterministic run-level BCa bootstrap confidence intervals for the
  direction-aware percentage ratio of median per-run statistics, with a
  reported percentile-bootstrap fallback for degenerate adjustments.
• Add Bonferroni matrix-wide multiplicity control, formal practical-equivalence
  decisions, per-cell inference diagnostics, and configurable confidence,
  resample count, random seed, and inference method.
• Add per-run IQR, coefficient-of-variation, and Tukey-outlier diagnostics,
  stronger round and raw-observation checks, and tail-latency-specific
  evidence requirements.
• Add comparison report schema version 2 with durable inference policy,
  confidence interval, estimand, family-size, method, seed, and adequacy
  fields while retaining strict schema version 1 loading.
• Add deterministic Williams-style matrix-cell ordering that balances ordinal
  position and, over a complete cycle, first-order carryover. Paired block
  members always receive the same scheduled cell order.
• Add manifest-backed paired collection with adjacent alternating AB/BA blocks,
  atomic complete-pair membership, separate baseline and candidate working
  directories, and resumable full-block retries that retain orphan results
  for audit without using them for inference. Automatic targets complete a
  joint supercycle in which every Williams row occurs under both AB and BA.
• Add explicit paired run-group comparison and paired run-level BCa intervals.
  Manifest-backed resampling preserves matched tuples and fixed AB/BA stratum
  counts while retaining the same direction-aware ratio-of-marginal-medians
  estimand as independent analysis.
• Add fixed-design precision planning for paired pilots using Student-t-scaled
  within-stratum paired log-ratio variability and Bonferroni-adjusted family
  confidence. Plans honor the evidence minimum, round to a complete design
  multiple, size a fresh future collection, and are explicitly not power
  analysis, sequential stopping, or permission to append to the pilot.

Changed

• Statistical-correctness exception: change the default minimum independent
  runs per side from two to five, require five rounds and retained raw
  round-duration observations per run, and require 100 observations with one
  iteration per round for tail-latency inference. These decision-default
  changes intentionally narrow the 1.x compatibility contract so weak
  evidence cannot produce a conclusive statistical claim.
• Make confidence-interval placement, rather than the observed Cartesian
  pairwise range, determine improved, regressed, unchanged, and
  inconclusive. unchanged now means the complete interval is inside the
  configured practical-equivalence region.
• Make single-run comparisons descriptive and inconclusive under the default
  inference method. Select legacy_consistency explicitly to retain the
  earlier non-inferential pairwise-range decision rule during migration.
• Apply configured CV and outlier limits to each process run. Pooled
  diagnostics remain available for description but no longer act as the
  primary quality gate.
• Write comparison and policy-inspection documents as schema version 3.
  Comparison reports continue to read and strictly validate schema versions
  1 and 2, and upgrade all three readable versions to the current typed model
  when loaded. The schema version 2 inference history above remains intact.
• Treat complete adjacent blocks, rather than successful commands, as the
  evidence target for paired collection. A failed or interrupted member makes
  the entire block ineligible until a fresh two-command block succeeds.

Deprecated

Removed

Fixed

Security