ObliviOS Developer portal Project site

Home / Decisions / ADR-031

Accept ADR-025's emulator-baseline residual (6 of 25 images differing between platform-primary and platform-repro) as final without performing the confirming rebuild…

Accepted, recorded 4 August 2026.

Decision
ADR-031
Date
2026-08-04
Status
AcceptedAccepted
Mentions
ADR-018, ADR-025, ADR-029

1Decision

Accept ADR-025's emulator-baseline residual (6 of 25 images differing between platform-primary and platform-repro) as final without performing the confirming rebuild ADR-025 left open.

2Why and evidence

Re-verified on the build server before this decision: platform-repro no longer exists as a live tree (removed for disk space), but its final comparison output is preserved at <build-root>/evidence/repro-v3-images/ and primary-v3-images/, and <build-root>/evidence/final-diff-v3.txt still documents exactly the 6 residual images ADR-025 names (product, product-qemu, super, system, system-qemu, vbmeta).

platform-primary/frameworks/base still carries the same three uncommitted files ADR-025 identified (git status --porcelain re-run today, unchanged). platform-primary/out (the OUT_DIR=out used for that comparison) has since been cleaned to reclaim disk space (145 MB remaining, no product images) — a confirming rebuild would not be a cheap incremental repackage.

The closest empirical precedent on this same server is the very round that produced this residual: final-v3-2026071500.started/.finished timestamps show it ran from 2026-08-02 13:52 to 2026-08-03 05:54, roughly 16 hours combined for both a platform-primary and a platform-repro rebuild — implying on the order of 6-9 hours for a platform-primary-only re-run, for a result ADR-025 already explains with source-level evidence (git blob comparison, disassembly of the one affected binary) rather than merely inferring it.

3Consequences

The qualified Accepted baseline (qualified — see ADR-025) status in docs/reviews/phase-1-emulator-baseline-review.md is now the terminal Task 5 exit state, not a pending-confirmation one: no byte-for-byte re-run of the reverted checkout will be produced for this baseline. The three-file dirty-checkout explanation and the two unexplained .bpf differences remain documented, non-blocking residuals — this decision does not newly explain the .bpf pair, which stays open on its own terms. This mirrors the precedent ADR-029 already set for the lynx alpha candidate: a root-caused explanation accepted in place of a second multi-hour confirming rebuild.

4Review when

If platform-primary's three files are reverted anyway for unrelated reasons and a rebuild happens to occur, capture the resulting hashes and compare against the still-available repro-v3-images/ at near-zero incremental cost and record the result against this entry; or the two unexplained .bpf differences are later root-caused; or evidence of actual build-environment compromise surfaces, per ADR-018's original trigger.

Read from docs/governance/decision-log.md, entry ADR-031.