ADR-0128: Retention and KSFLAG consumer experiment
Date: 2026-11-01
Status and decision provenance
Accepted for bounded experiment only; no production promotion. Venue: user/Codex workspace conversation,2026-11-01,America/Los_Angeles. Participants: user and Codex, with independent AI science/implementation reviewers. Decision owner: user authorized proceeding with the stated frost-enabled next step. Implementer: Codex. No institutional endorsement or additional attendees implied.
Decision and rationale
Cross the existing ADR-0127 retention preparations with soil KSFLAG0and1 for three fixed actual source horizons:84543623,84543461,85179901. Only experimental soil headers change; all other physical inputs remain paired-identical. Native KSFLAG1 enables frost and changes rangeland conductivity, so report within-setting retention contrasts and their interaction. A KS effect is not a pure frost effect.
The previous experiment skipped every frost opportunity and cannot support frozen consumer admission. A new observation-only probe will report actual executed conductivity/drainage domains; existing selected stage observations provide bounded accepted-flux accounting. Select windows from baseline-retention KSFLAG1 history before inspecting candidate differences. No candidate tuning or outcome thresholds.
Alternatives and limits
Reject using old dates from different profiles as presumed active events. Reject changing native frost enablement independently of the configuration for this phase: that would introduce another behavioral mechanism before the compound setting is understood. Reject global capacity multipliers and production fallback changes. Full-year physical closure, field accuracy, thaw generalization and source porous- fragment assumptions are not established by this bounded design.
Evidence and rollback
Work package contains exact declarations, source hashes, inputs, runtime, controls and reviews. All experiments use fresh private destinations; existing projects, production sources/defaults and release binaries are untouched. Failed or unsupported cases remain evidence and block promotion, rather than triggering silent repairs.
Executed decision
The comparator phase exposed an incomplete candidate KSFLAG1 history at the signed-withdrawal allocation boundary. Defer the new native observer campaign rather than weakening the complete-history gate or adding a tolerance. The retention inverse remains experimental. Compiled observer/reader controls are preparation evidence only; input/percolation closure remains unimplemented. The work-package execution amendment retains exact failure evidence and the next bounded signed-input provenance investigation.