- Decision
- ADR-216
- Date
- 2026-11-04
- Status
- ProposedProposed (owner's requirement 2026-11-04, chat; built 2026-11-04, first used on
2026110204; the part that gave a row with a check script to a runner is superseded by ADR-220, and the runner, its self-check and the script were removed on 2026-11-05) - Mentions
- ADR-209, ADR-211, ADR-220
- Mentioned by
- ADR-218, ADR-220
1Decision
A validation result proves a code state, not an image. A round runs only the rows its build can have changed: the rows that a commit of freeholdos-platform or freeholdos-services between the two images names, the rows that cite a patch that changed, the rows whose own row paths changed, and the rows that the impact line of the build lists (written when the changes are reviewed, before the build).
A row in that set that has a script tools/round-checks/<ROW-ID>.sh is proved by tools/docs/round-run.py on the phone with no owner and no agent step; every other row keeps its result and is carried by tools/docs/round-gate.py. A result of class X (other device) or class O (owner step) carries until its cause ends. An image that is not built yet gets a preview of the rows to run after its flash, which never holds a build.
A release candidate runs round-gate.py --full: nothing carries except class X and class O. No row is dropped, and no row is run again without a change that can affect it.
2Why and evidence
Owner 2026-11-04 (chat): the same proven rows were done again after every image, which cost days of the owner's time, energy and tokens; the validation must stay complete and cost far less. Measured on 2026110204 before this decision: 109 open rows, after it: 31 to run, 1 by the runner, 78 carried, 0 unproven [measured: tools/docs/round-gate.py 2026110204 ACH BOOT CMP GRB HRV IDS IMG KBD LNK MEM MSK NET PRIV PWM].
Rejected: (1) every non-PASS row recorded again on every image (the practice until now): 109 rows per image for code that did not change. (2) Carry by the code paths of the prefix alone: 63 rows were then carried without any proof that the change spared them, which drops coverage; the owner rejected it.
(3) Retire the seven superseded IMG-* rows: they block nothing (recorded NOT_RUN with the reason) and ADR-209, ADR-211 and the status records cite them as the source of a measurement. (4) A script that runs on every image: it proves code again that did not change (CMP-034 ran twice on the image where it had passed).
(5) One runner for every [agent] step: the steps are prose; a script is written per row and only when it proves the whole expected result, so a row with an unprovable part is split first (CMP-034, CMP-049, CMP-050).
3Consequences
Every fix commit names the matrix rows it affects. Every built image gets an image line (platform and services commits, base tag) in tools/docs/round-areas.tsv, and a build gets an impact line from the review of its changes. The carried rows rest on those lines: a dependency that is missing carries a row that the change broke.
The impact list is a reviewer's judgement [assumption]; the host tests and mutation runs cover the code, the device covers the rows named. Amends the practice of the skill fz-hardware-validation and the build gate sentence of AGENTS.md; the rule there that nothing is called complete before an on-device check holds for every row at its last proof and for a release candidate in full.
Records the owner's wish of 2026-11-03 that no row is dropped: none is.
4Review when
Reopen if a defect is found on a carried row (add its dependency as a row line and prove the row), if the round of 2026110204 still costs the owner more than the PIN moments of the rows that need one, or before a release candidate (run --full).
Read from docs/governance/decision-log.md, entry ADR-216.