# Audit plan — Differentiable Conformal Training for LLM Reasoning Factuality

Audit all six exact live claims against the complete arXiv `2604.20098v1` PDF and 30-file source tree. Recompute both headline retention percentages and every Table-2 agreement percentage from its confusion counts. Construct a finite dependency graph that independently checks the coupled Theorem-3.1 and Theorem-3.2 temperature limits, plus an ancestor-removal destructive control.

## Scope controls

1. Render and visually inspect pages 3, 4, 6, 8, 12, 13 and 21.
2. Preserve Table 7's exact qualification: the 141.1% MATH retention gain at alpha 0.03 misses the 97% coverage target by 0.45 percentage points.
3. Preserve Table 8's exact qualification: the 61.3% FELM gain at alpha 0.01 meets its 99% coverage target.
4. State that Theorem 3.1 itself establishes nonconformity-score convergence; soft-quantile recovery is a separate Section-3.3 statement and Section-4.2 empirical validation.
5. Do not claim an LLM scorer, MATH/FELM training run, cross-validation rerun or dataset evaluation.

Run twice with warning escalation and deterministic hashing; require all 36 gates, byte-identical trees, packaged replay, a recursive manifest, target-bound validation and every static route.