◉ Explain the mathematics
Claims and the discourse complex
Each sentence or line is a unit; from it the extractor reads typed claims: quantity (numbers with units), date (day, month or year intervals), order ("A before/after B", optionally "N days after"), relation ("X is N more than Y"), and proposition (the sentence's content with its polarity). Claims become vertices; two claims are joined when they speak about the same thing (content-word Jaccard ≥ 0.6, ≥ 0.8 for propositions, same structure and temporal scope). Modality is kept: asserted, reported (with source), hypothetical, question. Headings, reference lists, math and tables are not read as claims.
Γ — gluing, per layer
Γ = 1 if the asserted claims of that layer admit a global section (can all be true together), 0 if not, N/A if the text gives nothing comparable. It is the minimum over components, so adding unrelated text can never raise it. "Agreement" shows the share of comparable claim pairs that glue.
Linear layer — sheaf gluing and Hodge
Quantity claims about the same frame must agree slot by slot (relative tolerance 10% if hedged with "about", "nearly"; exact otherwise). Explicit relations ("A is 5 more than B") form an edge cochain on an entity graph; its Hodge decomposition y = grad + harm splits it into the part explained by consistent values and the harmonic part — a nonzero class in H¹, i.e. relations that cannot all hold around a cycle.
Ordinal layer — difference constraints
Dates pin events to day intervals; order relations add t_A < t_B. The system is consistent iff the constraint graph has no negative cycle (strict arcs counted lexicographically). Each failure is shrunk to a minimal conflicting set (deletion filter) and measured by ε*: the smallest uniform shift, in days, that would reconcile it (Karp's minimum mean cycle). Interval obstructions are not H¹ classes: a cycle basis does not suffice to certify them.
Propositional layer
Two units with the same content and opposite polarity (P and not-P) are unsatisfiable together. Contrastive negation ("X, not Y") and double negation are handled; deeper logic is out of scope.
κ classes
contradiction: all claims asserted by the text. conflicting accounts: at least one claim is reported speech from different sources — the text may be reporting a real disagreement accurately. inconsistent testimony: one reported source contradicts itself. tension with hypothesis: a hypothetical claim clashes with others.
Ψ — epistemic closure (prototype)
A property of the text's support structure, never of its author. Support edges come from markers ("therefore", "which proves"). ground: share of supported claims with no support path from an externally anchored claim (a citation, source, date). circ: groups of claims that justify only each other. seal: absence or denial of evidence used as evidence ("no documents exist, which proves the cover-up"). imp: asserted contradictions the text maintains. Pending validation on the LOCO conspiracy corpus.
Why v1 was retired
A preregistered audit (seed 201, 88 posts) showed v1's Γ, Ψ and K(t) did not react to injected numeric or negated contradictions (the tokenizer removed digits and "not"), rose when unrelated text was mixed in, and its "severed" count was unreachable (ρ ≥ 0.3475 > threshold 0.35). On the same protocol v2 detects 78% of injected negations (a naive near-duplicate check: 77%) and about two thirds of altered figures in sentences carrying a quantity or date (most misses are identifiers such as DRK numbers or citation years, which are not quantities), with zero false alarms on the 88 unmodified posts (the naive check: 25 posts) and full shuffle and segmentation invariance.