The cost curve.

Same workload. Two architectures.

Traditional LLM workflow

Q1→ LLM$
Q2→ LLM$
Q3→ LLM$
⋮
Qn→ LLM$

Cost scales linearly with query volume.

high $ / queryat scale

Truth Unit workflow

Q1→ LLMbuild Truth Unit · one-time$
Q2→ J(n) check · cache hit~0
Q3→ J(n) flip · re-score~math
⋮
Qn→ J(n) check · cache hit~0

Build cost amortizes across all queries that touch the unit.

3–5× cheaperat scale, stable domain

3–5× cost reduction on high-volume repeat queries (N ≥ 1k) within a single stable-precedent domain, once the Truth Unit cache is built and per-domain weights are calibrated.

Honest caveats: break-even ≈ N=770 queries (below that, LLM-only is cheaper). Per-query path adds J(n) traversal + settlement overhead, not ~0 — J(n) traversal IS the supersession traversal (UOM §4.2, clauses 2–3), same code, no new cost. Ratio compresses toward 2× in volatile-precedent domains (US constitutional case law). Asymptote ≈ 4.5× under realistic assumptions. See pressure-test-log.md Days 5, 10.