Open adjudications
Real conflicts between shipped members, recorded rather than resolved.
Resolving one of these changes released behaviour, and that is a decision with an owner. They are written down so a consumer knows what it is reading, not so someone quietly harmonises them.
Do not "fix" a member to match this page without an issue and a decision
entry. The vocabulary follows the tools; the tools do not silently follow
the vocabulary. Decisions live in each repo's docs/DECISIONS.md,
numbered, with the reasoning that produced them — do not relitigate a numbered
decision; if it is wrong, add a superseding entry.
Shared vocabulary
Statuses, per check
Common core, in every member that adjudicates: pass ·
fail · unsupported · skipped. Only
pass is green.
approximate is a domain-gated extension, for when a
measured error interval straddles the threshold. partspec has it; netspec
deliberately does not, because connectivity is discrete and importing interval
epistemics into an exact domain adds concepts without adding truth (D9, guarded by
test_report_carries_no_tolerance). Absence is a decision to
record, not an omission to fix.
Exit codes
Agreed across partspec and netspec today:
0 satisfied
1 violated
4 environment fault / error — not a verdict on the designThe adjudications
A1 — exit code 2 means two different things
resolved
Resolved: netspec adopted partspec's split —
2 is could not tell, usage moved to 64
(EX_USAGE). The two tools now agree on 0, 1, 2, 4 and 64. netspec
does not adopt partspec's EMPTY = 3: a contract asserting nothing is a
defect in the contract, and its synthetic result already says so.
partspec maps Verdict.INCOMPLETE to exit 2
(src/partspec/status.py:107) and puts usage at 64
(EX_USAGE). netspec uses EXIT_USAGE = 2
(src/kicad_netspec/cli.py:32), with no code for "could not evaluate".
A consumer branching on 2 across both tools is reading two different
facts.
Recommendation: adopt partspec's split.
A2 — netspec cannot report "could not tell" at the verdict level resolved
Resolved: Verdict is now
pass · fail · incomplete · error. A skipped or unsupported rule yields
incomplete at exit 2, and fail outranks it so
one unsupported rule cannot mask a real finding. Decided with A1, because exit
2 had to stop meaning usage first. Half of D10 stays
open: error is in the vocabulary and the exit map, but nothing
returns it — an environment fault raises before a netlist is read, so there is no
report to carry the field D10 asks for.
Verdict = Literal["pass", "fail"]
(src/kicad_netspec/check.py:39), green only when every rule is green —
so an unsupported rule collapses into fail, saying the
board is wrong when the truth is the tool could not check it. D10 also documents a
verdict: "error" the type does not contain; the environment fault
rides on the exit code, not on a report field as D10 requires.
Also case 2 in the record of the core defect.
A3 — gerberdiff has no third state resolved
has_changes: boolean (docs/schema.md), layer status
matched | added | removed. For a differ, too, silence must never read
as no difference.
Written up in full as case 1, which is now
resolved — gerberdiff#17 was fixed in d71f6ce
(2026-09-06), so an undefined aperture is an Error and exits
2 rather than dropping the flash in silence.
Resolved 2026-09-06 in 0685cde. The differ now has
its third state: a comparison reports
identical | different | indeterminate, and every operation the geometry
engine could not model is counted in summary.unrepresented and named on
stderr. different outranks indeterminate, so one
unmodellable stroke cannot mask a trace that moved. gerberdiff's exit codes now match
partspec and netspec — 0 · 1 · 2 · 4 · 64 — which was the other half of
A1 and left gerberdiff the last member out of step.
The instance that motivated it is
case 10: a trace stroked with a macro aperture read as
no change at exit 0 while the raster engine reported it. Note
the shape — has_changes was a boolean, so it was false for
two identical boards and for a board whose geometry could not be modelled.
Two verdicts could not carry three answers, exactly as in A2.
A4 — orlab's version fallback was log-only resolved
Resolved: profile_exact is now a public attribute on both
OpenRocketInstance and SimulationPool, so a caller can
branch on whether it is running against an exact profile or a nearest-older
fallback. Retained as case 3 in the record, where it is
the first instance found in the drive layer rather than the verify
layer.
The rule that produced them
Every adjudication here is a disagreement about how to say could not tell. That is not a coincidence — it is the one rule everything else follows from, showing up at the seams between tools that each solved it slightly differently.