The most widely accepted solution to the Hubble tension is also the worst-performing under cross-validation. MSCE proves it — and shows why peer review couldn't catch it.

When a physicist proposes a solution to the Hubble tension, they verify 1–2 observational conditions. But there are 8 independent verification conditions that must ALL hold simultaneously. MSCE checks them all at once — and reveals structural inconsistencies no single reviewer can detect.

Why didn't anyone find this before? Because peer review is serial. Reviewer A checks Condition 1. Reviewer B checks Condition 4. No one simultaneously checks all 8 — that's beyond human cognitive load. MSCE runs every claim against every condition in parallel. The conflicts were always there. They were just invisible to serial review.

MSCE is not an AI model. It is a multi-source verification system. It does not generate answers. It identifies condition inconsistencies across independent validation sources.

MSCE is to verification what a compiler is to code. A compiler doesn't write programs — it checks whether they can run. MSCE doesn't propose theories — it checks whether they can simultaneously satisfy all the verification conditions they claim to meet.

MSCE is not a research paper. It is verification infrastructure. Here is what it does for different people:

Scientists & Researchers Your theory satisfies conditions A and B. But there are 8 independent conditions that must ALL hold. Have you checked D, E, and F simultaneously? MSCE runs every claim against every known verification condition — in parallel. One command shows you where the conflicts are. → Example: Hubble tension

Peer Reviewers & Journal Editors A single reviewer typically checks 1–2 conditions per paper. No one person can hold all 8 in their head at once. MSCE flags cross-condition inconsistencies that serial review structurally misses. It does not replace reviewers — it gives them a tool to see what they collectively cannot.

Quantitative Finance & Risk Teams A trading strategy backtests well against 3 market regimes. Does it survive all 7 simultaneously — including the ones nobody thought to check? MSCE cross-validates strategies against a full matrix of independent risk conditions.

Security Auditors & Smart Contract Developers Your contract passed two audits. But have all known vulnerability categories been checked simultaneously? One audit covers reentrancy, another covers access control — who checks both at once? MSCE maps the protection gap.

Medical & Pharmaceutical Researchers Drug interaction studies typically verify 2–3 metabolic pathways. MSCE cross-validates claims against all known contraindication conditions — catching interactions that fall between specialist silos.

Legal & Compliance Teams Does your data policy simultaneously satisfy GDPR, CCPA, PIPL, and industry regulations? Each lawyer checks their jurisdiction. MSCE checks all of them at once — and finds where compliance in one region creates a violation in another.

Journalists & Fact-Checkers A claim cites two sources and looks solid. MSCE verifies it against all publicly available independent sources simultaneously. The contradiction is never in the sources you checked — it is in the ones you did not.

git clone https://github.com/sampson0826/msce.git
cd msce
pip install -e .
msce check hubble --quick

Output: A cross-validation matrix of 6 mainstream H₀ solutions × 8 independent verification conditions. All red.

The surprise is not that all 6 fail. It's which one fails hardest.

Early Dark Energy (EDE) — the most widely researched solution in the field, the one with the most papers, the most funding, the most citations — scores 0.076. Dead last. It simultaneously conflicts with CMB power spectrum, BAO scale, and S₈ large-scale structure.

If peer review worked the way people think it works, someone would have caught this. But no single reviewer simultaneously checks all three conditions. The conflict is spread across three different subfields, three different reviewer pools, three different sets of expertise. The contradiction is only visible when you look at all of them at once.

Even 2-factor combinations perform worse than single proposals. DDM + Local Void drops to 0.317 — below DDM alone at 0.358. The mechanisms interfere with each other. Fix one, break another. This challenges the foundational assumption that "combining solutions" will eventually resolve the tension.

→ Full analysis notebook

import msce

# Run the built-in Hubble tension analysis
result = msce.analyze("hubble_tension", quick=True)
print(f"Confidence: {result['confidence']:.3f}")
print(f"All proposals fail cross-validation: {result['all_fail']}")

# Check a custom theory (coming in v0.2.0)
result = msce.check(
    theory="My modified gravity model",
    conditions=["cmb_spectrum", "bao_scale", "sn_hubble", "bbn", "s8", "age", "gravity", "cross"],
    domain="cosmology"
)

Your claims ──→  [6-model ensemble]  ──→ Cross-validation matrix
                       │                          │
                 6 independent              N verification
                 LLMs vote on               conditions checked
                 each condition             simultaneously

1. Multi-Source Ensemble: 6 independent LLMs (GPT-5.5, Gemini 3.1, Grok 4.1, Kimi K2.5, GPT-5.1, o4-mini) vote on each verification condition independently.
2. 3-Layer Filter: Low-confidence judgments are discarded (L1), statistical outliers are flagged (L2), and collective blind-spot risk is detected (L3).
3. Cross-Validation Matrix: N claims × M conditions → every claim checked against every condition. Conflicts invisible to serial review become visible in parallel.
4. Deviation Diagnosis: The matrix is projected to a diagnostic space, identifying the deepest structural inconsistency — guiding where to fix first.

MSCE achieves 87.4% accuracy across 206 cross-domain verification tasks, compared to GPT-5.5's 74.8% — a +12.6 percentage point improvement.

MSCE excels in verification-dense domains. It falls slightly behind in open-ended creative tasks — and that's by design. A verification system should be conservative, not creative.

GPT-5.5 gave 40 high-confidence (>0.8) wrong answers in our 206-question benchmark. These are not edge cases — they are cases where a single model was extremely confident and completely wrong.

MSCE's average confidence is 0.49 — it achieves higher accuracy (87.4% vs 74.8%) while being more conservative. In high-stakes verification — science, finance, medicine — an honest "I don't know" is infinitely more valuable than a confident error. MSCE knows when it doesn't know.

git clone https://github.com/sampson0826/msce.git
cd msce
pip install -e .

Requirements: Python 3.10+. No GPU needed. For visualization features: pip install -e ".[notebook]"

• Quickstart
• Hubble Tension Analysis
• Benchmark Results
• API Reference (coming soon)

Is this AGI? No. It is a specialized verification system that uses 6 LLMs as independent voters, combined with a 3-layer filter and condition dependency analysis engine.

Can it check my paper? Custom claim checking is coming in v0.2.0. For now, the built-in Hubble tension analysis is available.

What domains does it support? Currently cosmology and general science. Finance, security, medicine, and engineering verification templates are on the roadmap.

Is the code fully open source? The CLI, visualization tools, and verification condition templates are MIT-licensed. The ensemble voting engine is available as a hosted API.

sampson1735937149@gmail.com

MIT — see LICENSE for details.

If you use MSCE in your research:

@software{msce2026,
  title={MSCE: Multi-Source Consistency Engine},
  author={Deng, Xinhang and MSCE Collaboration},
  year={2026},
  url={https://github.com/sampson0826/msce}
}