Vibe Science

Vibe Science v5.0 — IUDEX (Claude Code Edition) An AI-native research engine for Claude Code that transforms LLM agents into rigorous scientific investigators through structural enforcement, not just prompting. Grounded in Huang et al. (ICLR 2024), who proved that LLMs cannot self-correct reasoning without external feedback, v5.0 introduces four architectural innovations that make adversarial review structurally unbypassable: - Seeded Fault Injection (SFI) — Injects known errors into claim sets before review. If Reviewer 2 misses the seeded faults, the entire review is invalid. Mutation testing applied to scientific claims. Gate: V0 (RMS ≥ 0.80). - Blind-First Pass (BFP) — Reviewer 2 assesses claims without seeing the researcher's justifications first, breaking anchoring bias. Independent assessment before full context. - Judge Agent (R3) — A meta-reviewer that scores Reviewer 2's work on a 6-dimension rubric (Specificity, Independence, Counter-Evidence, Depth, Constructiveness, Consistency). Reviews the review, not the claims. Gate: J0 (≥ 12/18). - Schema-Validated Gates (SVG) — 8 of the system's 27 quality gates enforce JSON Schema validation against 9 read-only schema files. Prose claims of completion ("confounder harness: DONE") are structurally impossible to accept. Core architecture: OTAE-Tree (Observe-Think-Act-Evaluate) loop embedded in a branching tree search over hypotheses, with 7 node types, 3 tree modes (LINEAR, BRANCHING, HYBRID), best-first selection, and automatic pruning. Reviewer 2 Ensemble: 4-reviewer adversarial co-pilot operating in 6 modes (BRAINSTORM, FORCED, BATCH, SHADOW, VETO, REDIRECT) with a 12-flag mandatory checklist, double-pass review (Fatal Hunt → Method Repair), and mandatory counter-evidence search. Confounder Harness (Law 9): Every quantitative claim must pass raw → conditioned → propensity-matched. Sign change = ARTIFACT (killed). Collapse >50% = CONFOUNDED (downgraded). Survives = ROBUST (promotable). Additional subsystems: - 10 Constitutional Laws — non-overridable by any agent, protocol, or user request - Evidence Engine with typed claims, geometric confidence formula, hard veto thresholds - Serendipity Engine with active radar scoring (0–20 scale), cross-branch pattern detection, escalation protocols - Brainstorm Engine (Phase 0) with Inversion Exercise, Collision-Zone Thinking, Productive Tensions - 5-Stage Experiment Manager with iteration limits per stage - Circuit Breaker — same R2 objection ×3 rounds with no state change → DISPUTED (frozen, not killed), Stage 5 poison pill - Agent Permission Model — separation of powers: R2 produces verdicts but never writes to the claim ledger; R3 scores reviews but never modifies them; schemas are read-only for all agents - Knowledge Base for cross-research-question persistence - Claude Code hooks for deterministic runtime enforcement Runtime modes: SOLO (all roles in one context window) and TEAM (Agent Teams with dedicated context windows per role, requires Claude Code experimental Agent Teams feature). Battle-tested over 21 sprints of CRISPR-Cas9 off-target research, where it caught four confounded claims that would have passed conventional AI review — including one with OR = 2.30 and p < 10⁻¹⁰⁰ whose sign reversed under propensity matching. Lineage: v3.5 TERTIUM DATUR → v4.0 ARBOR VITAE → v4.5 ARBOR VITAE (Pruned) → v5.0 IUDEX