02 — Specification rulings, 2026-09-05

Twelve questions the implementation could not settle for itself, ruled and written into the specification.

This document is an index, not a source. Every ruling lives at the requirement it changed, with its reasoning attached; what follows is a list of what moved and where to look. It carries no normative text of its own on purpose — a second copy of a rule does not merely risk drifting from the first, it hides the disagreement, because each copy reads correctly in isolation. That is the failure EC-290 records and that C1 below finally cleared, and repeating it here to save a reader one hop would be a poor trade.

Status: applied at commits 8426b98 → 4ea0508. If you rely on any requirement listed here, re-read it.

The specification disagreed with itself

The implementation read past the text

Recorded, and now paid

Ruled while authoring the roster

Two more, on 2026-09-05, found by the unit-set validator the moment it was pointed at the shipped tier-1 roster — which is the argument for building a validator before its data.

Where measuring changed the answer

Three of the twelve moved when the recommendation was tested before being applied. They are listed together because the pattern is the point: each looked settled on reading and was wrong on measurement.

• A1 — Duel. The recommendation was to amend the city count 8 → 7. Computing all five rows found a second wrong figure (Large 120 against a derived 119), and then that 450‰ is the lowest land ratio that both generates and lands Duel on exactly the 8 already printed. The smaller change made the table true instead of accurately describing a broken tier.
• C3 — the art collision. The recommendation was one renumbering pass. The collision is far worse than reported — 96 identifiers defined twice — but of 1,631 citations only 39 are genuinely ambiguous; 992 sit in the defining document and 197 resolve. Renumbering 96 to repair 39 would rewrite every reference in the log, the history and the art bundle's metadata with nothing to catch the breakage.
• C2 — the ownership measure. Ruled wrong twice. First as "the threshold stands, the palette is the defect"; the ceiling test disproved it, since black-on-white against white-on-black reaches only 0.155 against a threshold of 0.25, so no palette could satisfy it. The asset team, asked about the instrument rather than the colours, proposed the support-normalised CIEDE2000 score that is now normative. It was verified before adoption: a CIEDE2000 implementation conformant to the Sharma reference set (max error 4 × 10⁻⁵) reproduces their figures — foreground separation 41.5, background 13.1, weakest asset 32.7 against their reported 32.8.

What the rulings now require

1. Calibrate the two ownership thresholds in Phase 2, then choose the palette — in that order, since the palette would be chosen against the measure. Calibrate by showing real scenes to readers and finding where the mechanical score stops predicting correct identification; ART-610 is authoritative. No sprite needs re-tinting whatever is chosen: the shipped cells are tint-mask bases and seat colours are substituted at render time.
2. Carry works through the world increment — applied 2026-09-07: the Terrain Record carries the works level (admitted at fidelity 3, null otherwise, as of the observation turn), its canonical encoding and the fog refresh have the field, and CB-2310's predictions read it through the knowledge ClashView.
3. Narrow ReturnToPost to CM-380's three causes — applied 2026-09-07: the persisted displaced marker on the unit, hashed state, set by a Reflex's route, a retreat or rout, or an external effect, cleared on landing on the assigned post.
4. Qualify the 38 ambiguous art citations — the ratchet stops new ones; these need reading one at a time to say which document each meant.

Not ruled

Fifteen further readings are recorded in SESSION.md as interpretations of silence rather than conflicts. They need no ruling unless challenged, and are listed there rather than here so that this document stays a record of things that changed.

Owner scheduling ruling, 2026-09-19

The learning game against VC-960, 2026-09-24

What the victory readout publishes, 2026-09-25

Owner-delegated rulings, 2026-09-26

The owner asked for these to be decided rather than left open: R1–R3 in the morning, R4 when its measurements were in. Each lives at the place it changed.

Built from source 47fa2827eff0 · VERSION.json