Contradiction is authored, never detected

Context¶

An agent reading a corpus it did not write asks two questions: does something here already say this, and does something here disagree with it. The first has an answer — the duplicate finding in docir lint --deep, two unlinked documents over 0.90 cosine. The second had none, and the symmetry is inviting: a second finding over the same vectors, flagging unlinked pairs in the band below duplicate as possible disagreements.

That check was designed, measured against this corpus, and refused.

Decision¶

docir computes no contradiction candidates, at any tier. contradicts stays an authored edge: a reader who finds a disagreement records it, and the kind's symmetric and successor properties make docir surface it from both sides on every later read. Detection is the reader's; amplification is docir's.

The corpus-wide view is delivered by pull, the way staleness is (Staleness as data (owner + verified + review cadence)) — a query over the authored edges, not a finding that arrives unasked.

What the vectors can and cannot see¶

Measured on this corpus at 255 document vectors and 32,385 pairs, under bge-small-en-v1.5:

• A claim and its own negation scores 0.771–0.983. "Tests live in a central tests/ tree, not inside each module", against its exact inverse, scores 0.983. That is the band in which agreement also lives, so cosine reports subject and never stance.
• The three contradicts edges this corpus actually carries score 0.726, 0.730 and 0.744 — ranks 6,922, 10,034 and 10,904 of 32,385.
• A threshold low enough to reach them admits 16,063 unlinked pairs. At 0.85 the candidate set is 80 pairs and contains none of the three.

No threshold is both readable and correct, because the signal is not in the vectors.

Why the structural variants fail the same way¶

Dropping similarity for co-governance does not rescue it. Unlinked pairs sharing at least one governed code: file: 801. Restricted to accepted decisions: 127, whose largest are pairs of ADRs that correctly govern one module together — a finding on correct usage, which is what A second relation graph, derived from the prose refused for the derived mention graph.

The zero-heuristic version fails too. "An authored contradicts whose two sides are both live" fires on 3 of 3 here, and all three are deliberate recorded deviations that are meant to stay live for as long as the deviation does.

Consequences¶

Disagreement is caught where a reader is already reading. The packaged skill tells an agent to run docir context over a description before docir add, and to record what it finds as an edge rather than as a memory.

The register is a query, and needs nothing built:

docir query --expr "length(related[?kind=='contradicts']) > `0`"
docir query --expr "length(related_by[?kind=='contradicts']) > `0`"

This reopens on evidence, not on symmetry (implements carries no property, and that is the decision): a model that scores stance rather than subject — which docir does not ship and does not intend to (Semantic embeddings on by default) — or a corpus whose maintainer reports disagreements found only by reading, at a volume a query cannot cover.