nx_dr_refute.nx
buildroot/runtime/nx_dr_refute.nx
about
nx_dr_refute.nx -- SOVEREIGN adversarial REFUTATION aggregator (DR-2).
The measured gap (liar-killed h2h): our "verify" was source-AGREEMENT (confirms
popular-but-wrong claims) -> adversarial refutation 0-vs-9 vs Claude deep-research.
This organ is the Co-Scientist Reflection/Critic kill-stage: N independent verifiers
vote refute/abstain/support on a claim, and it is KILLED only with enough EFFECTIVE
INDEPENDENT refutations. It bakes in the Structured-Test-Time-Scaling theory
(xinmingtu.cn 2026): correlated retries buy almost nothing -- m_eff = m/(1+(m-1)*rho)
-- so verifiers must be DECORRELATED (different lenses/evidence/prompts) to count.
Integer, deterministic. Imports ONLY nx_syscalls = drift-immune. No hw writes (Rule 26).
module: nishi-core.research.dr_refute
depends: nx_syscalls.nx
genealogy_id: coscientist_reflection + structured_ttscaling_2026_decorrelation
dependencies 1 imports · 1 importers
imports: nx_syscalls.nx
imported by: nx_dr_refute_cli.nx
structs
| none |
consts
| none |
functions
| 17 | func rf_tally(votes: *i64, n: i64, out: *i64) -> i64 called by 1: main |
| 31 | func rf_effective_n(m: i64, rho: i64) -> i64 |
| 38 | func rf_effective_refutes(refutes: i64, m: i64, rho: i64) -> i64 |
| 46 | func rf_kill(refutes: i64, supports: i64, m: i64, rho: i64, bar_permille: i64) -> i64 |