A rule may say or, once

[WaylandYang]

Closes #476. Stacked on #483. Decision record: docs/decisions/0026-a-rule-may-say-or-once.md.

A criterion is rarely one pattern. A gas-bearing well is one whose total hydrocarbon is above 8 and whose interpretation reads as a gas anomaly — or one whose composite interpretation already says gas-bearing. Until now that was two rules, two names, the same conclusion, and nothing in the base saying they were one criterion: delete one and the other keeps firing.

The shape

A condition carries a group. Conditions in a group are joined with and; groups are joined with or. (A and B) or (C), one level, no nesting.

That is not a concession to keep the editor simple — it is the shape the evaluator already has. A hit does not report "true", it reports which facts made it true and the interval on which they all hold; getting there means walking the cartesian product of the per-condition matches and intersecting their intervals, which is a conjunction. Running that once per group keeps every property it has. An arbitrary and/or/not tree would have to answer "what are the premises of a disjunction", and neither answer is good: the fired branch's premises are what groups already give without the tree, and all of them would attach a conclusion to readings that had nothing to do with it.

What changed

• Migration 0039 adds group_seq; existing conditions default to group 0, so every rule keeps its exact meaning. The unique key becomes (rule_id, group_seq, seq). (Numbered 39, not 38 — another branch owns 38 in the shared dev database.)
• The evaluator runs group by group. The combination cap is per group: a group that blows past 64 is skipped and reported while the others still conclude, and the report counts a (rule, entity) pair once. Two groups true on the same interval are one conclusion, keeping the premises of the first group in group order so the proof is stable across runs.
• not_in joins the ops. It reads a value that is there and asks whether it falls outside a set, so it has a premise and an interval like any other condition. "This entity has no such attribute at all" is deliberately not here — it has no premise, so nothing could ever retire it, and the base is open-world: a missing reading means nobody wrote it down.
• The API carries group on a condition, defaulting to 0 — a caller that sends a flat list still sends one conjunction.
• The editor writes blocks: rows joined by and at the head of the row, blocks divided by a rule and the word or, one button per block to add a row and one at the bottom to add a block. Deleting a block's last row takes the block with it, so no block is ever empty except the first one before anything is written — an empty conjunction is true of everything and there is no state in which one can be saved. Blocks are not boxed; the connectives already say how far a block reaches.
• The table sentence reads A and B or C.

Tests

Four evaluator tests (either group fires and carries only its own premises; two groups on one interval are one hit; a capped group does not stop the other; not_in needs a reading to be true) and one database test that a rule with two groups names only the premises of the group that fired. Verified end to end against a running stack: a three-block rule written in the dialog stores 0,0,1,2 and reads back with both ors in the sentence.

Chaining — a rule reading what another rule concluded — is #477 and is not this.

🤖 Generated with Claude Code