# ADR: S4 final-target repetition and admission hold

Status: accepted
Type: adr
Updated: 2026-08-10
Next Action: preserve current A/B observations; review and correct exact-read scaling before explicitly reopening 129600 admission
Parent: work/checklists/checklist-hpt2-bounded-stop-recovery-and-regression-measurement.md
Supersedes: the v1 requirement for three observations in every 129600 region
Superseded By:

## Context

The original S4 contract required three observations for every target/region.
No confidence target or independent failure mechanism justified three complete
129600-record repetitions in each BEFORE, AT, and AFTER region. Live A/B data
also exposed occupancy-dependent exact-read cost, making the nine final-target
runs per controller dominate the campaign while testing the same boundaries.

## Decision

Retain three observations for ordinary targets and require one BEFORE, one AT,
and one AFTER observation at 129600 per controller. Hold every new 129600 run
before controller observation and Start creation. Current admitted runs and
their Stop paths continue. Treat lookup scaling as a system/method defect, not
as a card failure. Withhold downstream completion dates while the hold is
closed and show recent counter-based progress for the current observation.

## Consequences

- Existing evidence is preserved and remains append-only.
- The final target still covers all three boundary regions on both controller
  variants without six unsupported duplicate long runs per controller.
- The campaign cannot silently advance into 129600; reopening is a reviewed
  deployment change.
- Firmware release remains gated by the exact-read lookup correction or a
  separately reviewed bounded implementation and returned regression proof.
