Markdown for LLMs
Facts and evidence: pitfalls
The source Markdown for this article. Copy it into your assistant or download it as a text file.
# Facts and evidence: pitfalls Wrong forms, silent outcomes, `LDC-E` diagnostics, typical mistakes of formalizers and AI agents, and how to detect them. Items marked “confirmed by runs” were reproduced during this research on the installed engine. ## 1. The fact exists but the rule stays silent Three different causes with the same Not established, not refuted: - the premise is Contradiction — a bare literal reads as established; - the fact is not accepted: off-mode `origin` (`ASSERTION_NOT_ACCEPTED`, warning) or the `valid` window missing the law date (`ASSERTION_OUTSIDE_VALID`, info); - the argument value mismatched (another region, another rate). Detection: the why-not explanation names the conjunct; a guard test on Not established, not refuted must record which of the three causes is meant. Confirmed by the `research.facts.polarity` teaching package: the empty case gives Not established, not refuted for exactly the third cause (the `licensed` premise is simply absent). ## 2. Two identical anonymous facts `LDC-E1331` in 0.2: both share one preimage (the normalized fact with `origin`, times, anchors), and the StableId would fork. Two assertions of one fact on different grounds need distinct `id`, `origin`, or `evidence`. The AI agent duplicates a fact “for reliability” — reliability here is a compiler rejection. ## 3. A `valid` window in the future with today’s law date (confirmed by runs) The fact is silently outside supports: Not established, not refuted and an info-issue. Reproduction: the `research.facts.valid_window` teaching package at law date 2026-03-01 answers Not established, not refuted on the same fact that at 2027-06-01 gives Established. A guard test is `expect issue(ASSERTION_OUTSIDE_VALID)`, otherwise window Not established, not refuted is indistinguishable from “no fact”. Remember half-openness: `[@2027-01-01, infinity)` includes the first day. ## 4. A typo in `origin` `LDC-E1374`, an error: the origin list is closed at six names (`case_input`, `source_asserted`, `external_snapshot`, `derived`, `adjudicated`, `assumed_for_simulation`). Assumptions are `assumed_for_simulation` in simulation mode, not disguised as `case_input`: faking provenance is not “test simplification” but a different fact meaning. ## 5. A rule in `given` A case cannot carry law: the program is presented separately, a test without it is rejected. The AI agent, used to “the test holds everything”, puts a rule in the case — and gets a rejection instead of a run. Fix: the rule goes in the package, `given` holds only facts, documents, and context. ## 6. An `evidence` declaration in the package `LDC-E1314` (warning): a document is a case input, not part of the program; without a `status` field it is already a rejection `LDC-E1208`. The core does not infer a document’s admissibility from a number — dedicated rules express it. A document goes in `case` or `given`, the `evidence Name;` reference in the fact. ## 7. A computation in a fact argument `100 + 20` is `LDC-E1325`, a non-canonical Rational spelling (`10 / 4`) is `LDC-E1324`: a fact is a value, not a term; write the result (`120`, `5 / 2`). Arguments match on canonical values; the predicate must be declared (`LDC-E2102`), arity (`LDC-E2103`) and types (`LDC-E2104`) must match. In a transport-JSON case the same check gives runtime-issue `TYPE_ERROR`. ## 8. Fact-group inclusion without context; swapped time order Including a missing or non-public group, or including without an import context, is `LDC-E1327`: a group materializes into the consumer’s compiled output only with a presented context. `observed` later than `recorded` is warning `LDC-E3101`: the file is accepted, but the order is broken — first observed, then recorded.