cas-internal-inverse-quintic-degree-five-witness
- Kind
- witness-replay
- Status
- checked
Supports: the inverse-function certificate (the recomputed derivative, its recorded single sign, the Sturm monotonicity count, the shifted polynomial q = p - y, the named degree-5 witness and its isolating bracket, and the conclusion p(x) = y) is independently re-derived and accepted by a checker (verify_inverse_certificate) that never calls the producer and recomputes the derivative and the shift with checker-local implementations sharing no code with axeyum_ir::poly
cargo test -p axeyum-cas --lib inverse::tests::quintic_inverse_beyond_radicals -- --exact 2>/dev/null | grep -cE '^test inverse::tests::quintic_inverse_beyond_radicals \.\.\. ok$' Evidence notes
THIS EVIDENCE IS cas-internal, NOT kernel-reconstructed (ADR-0601 SS2): the checker_command only ever names the axeyum-cas package (never axeyum-lean-kernel), so scripts/validate-facts.py's classify_cas_certificate_checker classifies this fact's evidence as cas-internal. No Kernel::add_declaration term exists for this claim and none is claimed here. The checker's own falsifiability was MEASURED by guard deletion in a lane snapshot rather than asserted: of 15 checks, 12 survived deletion with every test still green, which refuted this module's first-draft claim of nine independently-falsifiable guards. One check that could never fail on its own was deleted outright, one adversarial fixture was rebuilt so that the monotonicity guard is killed by exactly one test, and the module doc now carries a measured table of which check backs up which. Current state: 4 of 14 checks are killed by exactly one test; the remaining ten are mutually backing, and the backup relation was itself measured by deleting a survivor together with its hypothesised backup. The adversarial fixture worth naming is verify_rejects_a_nonmonotone_bracket_that_every_other_check_accepts -- p = x^3 - 3x on [-3/2, 3/2] with y = 0, where p' is positive at BOTH endpoints (so the sign guard cannot reject), y is strictly interior to the range, and q has exactly one root in the bracket (so the uniqueness recount cannot reject either) -- yet p is not injective there, so the certificate is wrong and only the Sturm monotonicity count sees it.