Implements 0021 end to end, closing #277. Not ready to merge — the record's status line still has to move and a few polish items are listed at the bottom.

A mud-logging user wants a well marked gas-bearing when total hydrocarbon clears a threshold and the interpreter filed it as a gas anomaly. Both values were already dated attribute facts with their passage attached; the rule that reads them had nowhere to live.

What landed

Schema — derived_facts widens to match facts: nullable object_id, an object_value, and the same object_id IS NOT NULL OR object_value IS NOT NULL check. Rules live in attribute_rules with structured conditions. A derived row now comes from an axiom rule or a business rule, exactly one, enforced by a check.

Evaluator (utopia-reason::rules) — a second pass, not an extension of derive(): that one walks entity–entity edges under axioms, this one compares literals against thresholds. Evaluation runs over combinations — one satisfying fact per condition — because an entity can carry several readings of the same attribute, and the interval of each combination is the intersection of its premises. So a well whose 2023 reading fires and whose 2025 reading does not gets one interval, not a row that flips.

Materialisation — rule hits join the axioms' single reconciliation pass. They must: the stale check scans the whole table, so two passes each diffing separately would retire each other's rows every run.

Declaration and UI — CRUD with errors written for people, a "Run now" that reuses materialize rather than a second code path, and a Business rules page under Ontology where a rule reads as one sentence: Applies to Well (and its subclasses), when all of Total hydrocarbon is above 8 · Interpretation is one of gas anomaly, post-effect gas anomaly ⇒ concludes GasBearingWell.

Explanation — the entity panel's Derived tab shows the conclusion under the rule's own name with the readings that made it true.

MCP — two read-only tools (list_rules, rule_matches), and entity_facts now returns conclusions marked [rule: <name>]. Writing a rule is deliberately not exposed: 0002 and 0021 both hold that the reasoning engine's criteria are written by a person, and a tool call cannot tell "a person dictated this" from "the model invented it".

Five silent failures this turned up

Widening a table means re-reading every path that reads it. Each of these fails without an error:

• The identity index treated NULL object ids as distinct, so a literal conclusion inserted a fresh row every run. Rebuilt with NULLS NOT DISTINCT and object_value in the key.
• derived_for_entity inner-joined entities and rules, so rule conclusions vanished from the panel — in the database, absent from the UI, no error.
• The graph's edge count included literal conclusions, which are not drawn, so the status line claimed more edges than the canvas has.
• proof() had the same two inner joins: opening a conclusion's proof 404s.
• RDF export inner-joined rules, so no rule conclusion reached the file at all — and 0020's promise is that an auditor reads it without us. Fixed, with a test pinning that the object is a literal, the premises are attached, and the activity carries the rule's name.

One regression was mine and an existing test caught it: I overwrote DeriveReport.derived, whose meaning is "derivations computed this round" (blocked ones included), with "rows about to land". Rule hits now add to it instead.

Verified

Unit: 8 evaluator cases — the conjunction, a missing condition, two readings giving two intervals, premises that never overlapped, a number extracted as a string, raising the threshold, an empty rule concluding nothing, the combination cap being reported.

Database-backed: 5 cases — the premise chain, a rerun inserting nothing, retirement rather than deletion, disabling a rule, and the asserted type staying untouched.

Live, on a real instance: three wells separating a hit from two near-misses; then a subclass entity, two premises that never overlapped, a string-typed number, three disjoint intervals on one well, and both conclusion kinds coexisting on one entity. 48 test targets green, clippy clean with -D warnings.

Before this merges

• 0021's status line still says "planned · nothing built".
• A rule's conclusion cannot be edited, only deleted and rewritten.
• The rule card's entity count is not clickable — rule_matches is the backend for it, the UI is not built.
• capped is reported once in a toast with nothing persistent on the card.
• No canvas marker for a rule-classified entity, deliberately: a conclusion is a judgement about one entity rather than a connection, and asserted attributes have no marker either. Worth an open question rather than a fifth visual language on the canvas.

🤖 Generated with Claude Code