Sheaft v1.2.0

v1.2.0 turns Sheaft's stochastic-connectivity analysis into an operation-specific failure-tolerance release signal. It builds a fail-stop resilience curve, certifies the evaluated margin with confidence bounds, and can block releases that reduce that margin.

The release remains within the v1 semantic boundary documented in docs/v1-major-semantics.md.

Added

• analysis config schema 1.2 with endpoint-targeted sweeps
• independent replica failure-probability and exact failed replica-slot axes
• raw curve points, SLO crossing brackets, and exact k / total_replica_slots metadata
• two-sided Wilson confidence intervals at a configurable confidence level
• conservative certified_tolerance based on a contiguous prefix whose lower confidence bound meets SLO
• coupled per-trial fail-stop sampling to avoid avoidable Monte Carlo curve inversions
• explicit boundary gate rules for minimum certified tolerance
• compatible baseline boundary diffs and maximum regression budgets
• stable reasons: boundary_below_minimum, boundary_regressed, and boundary_indeterminate
• runnable configs/analysis.sweep.example.yaml in the default config pack

Product workflow

A release policy can now express:

Checkout must remain certified through 5% independent replica failure probability, and a new release may not regress more than 2 percentage points from the last-release baseline.

Raw baseline artifacts are evaluated with the same sweep definition and seed. Prior reports are compared only when their sweep fingerprints match. Missing, statistically insufficient, non-monotonic, or incompatible evidence is indeterminate and fails closed when mode: fail and default_action: fail are configured.

Compatibility

This release does not change accepted Bering contracts:

• [email protected]
• [email protected]

Analysis schemas 1.0 and 1.1 remain accepted. Schema 1.2 is additive and Sheaft-owned; it is unrelated to the retired Bering preview contract 1.2.0.

Boundaries

The reported margin is a modelled fail-stop boundary. v1.2.0 does not model traffic redistribution, per-replica capacity, queue saturation, or retry-generated load and therefore does not claim to predict an overload cascade or live runtime tipping point. Automatic probability calibration, arbitrary non-product P, live chaos execution, and rich temporal workflow models also remain outside the release claim.