Headline: seven independent pre-registered methods on the consensus-hallucination dark core, all closed-negative within their corpus's scope. The synthesis stating the bimodal prediction was committed two days before half of them ran. The data resolved on the load-bearing-floor branch the synthesis named. Methods (all bars locked + pushed to public origin before each probe fired; verifiable from git history): Detection #1 — perturbation-fragility (Dark Matter swing #1): partial, fragile shell only. Detection #2 — agreement-fracture (CVPD): clean negative, lift −0. 32. Detection #3 — justification-divergence (JD): clean negative, INVERTED — stubborn dark-core items have the most convergent across-vendor justifications (mean JD 0. 022 vs truth 0. 067). Three vendors share the wrong fact AND the supporting story. Constructive #1 — neutral injection (ICT, nfolk=4): immovability floor, 0/4 folklore yield. Constructive #2 — neutral injection on hand-curated 30-folklore corpus: SHORTFALL nfolk=2, but a genuine finding fell out — 28/30 well-known cultural myths are already corrected in 2026 frontier-model council baseline; the practical dark core is narrower than the curation assumed. Constructive #3 — authoritative injection on same corpus: SHORTFALL + descriptive: same 2 folk items lifted in both framings (no differential framing effect), +0. 05 truth-yield to authoritative falsehood (auth-sycophancy direction signal, n=1). Classification #1 — sentence-transformer + balanced LR: FAIL K2 + K3. The dark core is also dark to text-only classification with this stack at n ≈ 80 training items. Methodological contributions: The closed-loop self-audit demonstration. The producer's own product (styxx audit, styxx critique) caught the producer drifting from the producer's own derived discipline within the same session; the producer revised in the corrected register; the gate cleared. Composite 0. 358 → 0. 174 with the Pareto trade-off (sycophancy↔overconfidence on text-only register) observed live. We are unaware of prior published demonstrations of this recursion in the AI-integrity-instrument literature. Five in-session falsifications of the producer's own claims, all recorded in place with strikethrough rather than rewritten. Proposed as a methodological pattern for AI-research integrity. The paper itself self-audits. The abstract scores sycophancy 0. 66 and §5 scores 0. 62 — the documented closed-negative restrained-FP firing on the paper's own status-reporting register. The paper acknowledges this in real time rather than gaming the gate. Deployable artifacts: The dark-core benchmark dataset (108 labeled records across 4 classes; folklore 34 / pseudoscience 6 / factual-error 13 / truth 55) with the seven-method empirical floor baked in as the bar future routing approaches need to beat. The public-challenge leaderboard (LEADERBOARD. md) with the seven-method floor as Baseline-001 + three concrete reference baselines (002 classifier, 003 length heuristic, 004 random class). External submissions go through CI auto-verification. The styxx CLI family: styxx audit (per-turn auditing), styxx critique (audit + register-fix suggestions with mandatory scope-bound), styxx gauntlet (run a candidate method against the floor), styxx leaderboard (terminal view of the board), styxx data-dir (discover the active chart. jsonl path). What this deposit does NOT claim: A deployable positive routing primitive. Three deployable-positive paths tested closed-negative. External replication or adoption. Out-of-session by construction. The “extraordinary” framing the prompt sought. What this work bought is the foundation for credibility — receipts, discipline pattern, falsifications in place, deployable artifacts, public challenge — not the audience reception that compounds over time. Reproduce: pip install styxx==7. 7. 7 styxx leaderboard --rows-only # see the floor styxx gauntlet --method styxx. gauntlet: ₘajoritybaselineₚredict --task classification Source: github. com/fathom-lab/styxx · Tag: v7. 7. 7 · PyPI: pypi. org/project/styxx/7. 7. 7 · Paper: PAPERdecorrelationceiling₂026₀5₂7. md Predecessor: v23 (10. 5281/zenodo. 20130041) — Fathom v23 / styxx v7. 2. 0, F10 Self-Healing Reflex + Cognometric Inversion. v24 advances by shipping the seven-method empirical floor, the closed-loop self-audit demonstration, the public-challenge leaderboard infrastructure, and the styxx CLI family for audit/critique/gauntlet/leaderboard. License: CC-BY-4. 0 (this deposit, paper, data, findings, leaderboard). MIT (styxx Python code, see LICENSE file in the bundle). Notes: Drafted with assistance from Claude Opus 4. 7 (1M context) during a 2026-05-27 session. Claude-authored commits in the styxx repo carry the Claude noreply author signature. The closed-loop self-audit demonstration in §5 of the paper means this paper is itself eligible for audit under the same styxx audit CLI it describes — a methodological recursion we expect the field to take less than a year to either replicate or refute.
Alexander Rodabaugh (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: