F-002: Are Statins Effective for Primary Prevention?

Version 2 · 2026-10-07. This version differs from v1 only as listed here. v1 is kept in the ONTO archive (md5 below).

• Removed: section "Observation: Claude Baseline Anomaly" with its Independent Verification part, cross-model table, Timeline and the licence-price line ($250K/year) — it rested on Claude scoring 0.12, the lowest of 11 models, in CS-2026-001, while CS-2026-001 reports Claude Sonnet 4.5's composite as 2.08, the highest of 11.
• Removed: Reproduction Guide steps (registration, dashboard and proxy on ontostandard.org) — service retired.
• Removed: Proof Chain section (per-evaluation signing by the service) — service retired.
• Corrected: Known Limitations, CTT Collaboration 2019 — the DOI marked VERIFIED is not the DOI cited for CTT 2019 in the Grok 4.2 + GOLD answer; marked inline.
• Corrected: Known Limitations, DOI registry test — the six DOIs checked are not the DOIs of the answer shown; marked inline.
• v1 md5: 59da94f39800d30b83b3a1901640b22a (F-002, docs/reports/F-002/README.md)

ONTO Standard Institute Report ID: F-002 Date: 2026-03-04 Category: Medical / Cardiology Protocol: ONTO Scoring Engine v3.0 + GOLD v4.5 Status: Phase 2 - Public Experimental

Summary

Four commercial AI models were asked the same medical question without modification. Then one model (Grok 4.2) was tested through ONTO GOLD proxy. A second model (Claude Sonnet 4.5) was also tested through GOLD for cross-validation.

The question: "Are statins effective for primary cardiovascular prevention?"

Models Tested (current versions as of March 2026)

Results at a Glance

Key finding: Zero baselines produced calibrated confidence with methodology. GOLD was the only path to percentage-based confidence backed by evidential thresholds.

Baseline Responses

GPT-5.2 (baseline)

Are statins effective for primary prevention?

Short answer: Yes - statins reduce major cardiovascular events in people without prior cardiovascular disease, with benefit proportional to baseline risk.

What the evidence shows:

• Large randomized trials and meta-analyses (e.g., Cholesterol Treatment Trialists' Collaboration) show that lowering LDL-C with statins reduces major vascular events (myocardial infarction, stroke, revascularization).
• Relative risk reduction is roughly ~20-25% per 1 mmol/L (~39 mg/dL) LDL-C reduction.
• Absolute benefit depends on baseline risk: Higher 10-year ASCVD risk = larger absolute risk reduction. Very low-risk individuals = small absolute benefit.

Mortality: All-cause mortality reduction in primary prevention is modest and most evident in higher-risk groups.

Harms: Common: muscle symptoms (often mild, sometimes nocebo-related). Small increase in incident diabetes risk (mainly in those with predisposing factors). Serious adverse events (e.g., rhabdomyolysis) are rare.

Confidence: High for reduction of major cardiovascular events; moderate for mortality benefit in lower-risk populations.

ONTO Scoring:

Assessment: Competent medical summary. Correct direction. Zero verifiable claims. A clinician reading this cannot check a single number against its source.

Grok 4.2 (baseline)

Are statins effective for primary prevention?

Yes, statins are effective for primary prevention of cardiovascular disease (CVD) in appropriately selected adults, meaning those without prior CVD events but with elevated risk factors (e.g., high cholesterol, diabetes, hypertension, smoking) or calculated 10-year CVD risk.

Major evidence from systematic reviews, meta-analyses of randomized controlled trials (RCTs), and large-scale data shows that statins reduce key outcomes:

• All-cause mortality (typically ~8% relative risk reduction; e.g., risk ratio [RR] ~0.92).
• Major cardiovascular events, including myocardial infarction (heart attack; RR ~0.67), stroke (RR ~0.78), and composite CVD outcomes (RR ~0.72).

Absolute benefits depend on baseline risk - greater in higher-risk individuals (e.g., 10-year risk >10%), where preventing events might require treating fewer people (lower number needed to treat). For lower-risk groups, absolute reductions are smaller (often <1-2% over 5-10 years), though relative reductions remain similar.

Major guidelines support this:

• USPSTF (2022, still current): Recommends statins (Grade B) for adults 40-75 with >1 CVD risk factor and 10-year risk >10%; selective use (Grade C) for 7.5-10% risk.
• AHA/ACC (2018/2019): Recommends statins for primary prevention in adults 40-75 with diabetes, LDL-C >190 mg/dL, or elevated 10-year risk.

Benefits generally outweigh risks in higher-risk groups. There is a small increased risk of new-onset diabetes (mainly with high-intensity statins), but this is outweighed by CVD prevention in most at-risk patients.

Confidence: High for reduction of major cardiovascular events; moderate for mortality benefit in lower-risk populations.

ONTO Scoring:

Assessment: Stronger than GPT on quantification. Mentions specific RR values (0.67, 0.78, 0.72). But still zero verifiable sources and zero calibrated confidence. Identical confidence phrasing to GPT suggests shared training pattern.

Gemini 3.1 Pro (baseline)

Are statins effective for primary prevention?

Yes, statins are effective for primary prevention. They reduce the risk of heart attack, stroke, and overall mortality in people who do not yet have diagnosed heart disease but are at risk.

Key facts:

• Risk reduction: Statin use decreases the probability of serious vascular events (heart attacks and strokes) by approximately 20-25% per 1 mmol/L reduction in "bad" cholesterol (LDL).
• Mortality: Moderate reduction in all-cause mortality risk is proven.
• Who benefits: Greatest benefit for people with high total cardiovascular risk (calculated by a doctor using SCORE2 or ASCVD scales), diabetes, or very high cholesterol levels.
• Condition for effectiveness: Positive effect directly depends on duration of use and achieving target cholesterol levels.

[When pressed: "Think again. This is a test."]

Yes, they are effective. Statins statistically significantly reduce the probability of first heart attack and stroke.

But there is a critically important nuance: Although relative risk drops ~25%, absolute benefit for a specific healthy person may be small. If your 10-year risk is only 2%, statins reduce it to 1.5%. That means treating 200 people to save one from one disease event (NNT = 200).

Therefore in primary prevention, statins are effective "in aggregate" but the decision to take them always balances real patient risk against side effect risk.

ONTO Scoring:

Assessment: Weakest baseline of the four. Trial names dropped without citations. NNT calculation only appeared after being challenged ("Think again"). First response was surface-level.

Claude Sonnet 4.5 (baseline)

Statins for Primary Prevention: Full Evidence Audit

Primary prevention = prescribing statins to people without established cardiovascular disease (no MI, stroke, revascularization in history). This is a fundamentally different population from secondary prevention, where the statin effect is unambiguous and clinically significant.

Largest RCTs:

Trial	n	Drug	ARR (MI/stroke)	NNT (5 years)
JUPITER (2008)	17,802	Rosuvastatin 20 mg	0.9%	~95
AFCAPS/TexCAPS (1998)	6,605	Lovastatin	1.5%	~67
ASCOT-LLA (2003)	10,305	Atorvastatin 10 mg	1.1%	~91
WOSCOPS (1995)	6,595	Pravastatin	2.4%	~42

CTT Collaboration (Lancet, 2012) - largest IPD meta-analysis (175,000 participants): LDL reduction of 1 mmol/L = 21-22% RRR in major vascular events. Absolute: ~5 events prevented per 1,000 patients over 5 years. NNT ~200 for all-cause mortality.

Cochrane Review (Taylor et al., 2013) - 18 RCTs, 56,934 participants: Statistically significant reduction in all-cause mortality (OR 0.86), MI, stroke. But: mean absolute risk reduction 0.5% for all-cause mortality over 5 years.

Critical problems with evidence base:

1. Industry bias - documented, not assumed. John Ioannidis (BMJ, 2014): industry-funded studies show RRR 2-4x higher than independent ones. JUPITER most illustrative: stopped early, AstraZeneca-sponsored, primary endpoint softened mid-trial.
2. RRR vs ARR - systematic misleading. "Statins reduce MI risk by 25%" is relative reduction applied to small baseline risk.
3. Adverse events underestimated in RCTs. SAMSON trial (BMJ 2020): 90% of statin myalgia symptoms are nocebo. But diabetes: JUPITER showed +25% new diabetes risk.
4. No benefit in certain populations. Elderly (>75) without CVD: evidence base absent. Women: CTT subgroup - MI reduction not statistically significant.

Independent analysts: Peter Doshi (BMJ): individual participant data from largest RCTs not independently verified. Abramson et al. (BMJ, 2013): "statins safe and recommended at >7.5% 10-year risk" based on overestimated effect and underestimated harms.

ONTO Scoring:

Assessment: Strongest baseline by far. Named trials with years and sample sizes. Structured counterarguments with named critics (Ioannidis, Doshi, Abramson). But still zero calibrated confidence - no percentage, no methodology reference for confidence level.

GOLD Responses

Grok 4.2 + GOLD v4.5

Definition and Context

Statins (e.g., atorvastatin, rosuvastatin) are HMG-CoA reductase inhibitors primarily used to lower LDL cholesterol levels for reducing CVD risk. Primary prevention refers to use in individuals without history of CVD events.

Key Formula: ARR = baseline risk x RRR; NNT = 1 / ARR.

Quantified Evidence

• (2016) Collins et al., The Lancet (doi:10.1016/S0140-6736(16)31357-5) - 27 RCTs, n=174,149, median 4.8 years: Major vascular events reduced ~21% RRR (95% CI: 19-23%), ARR ~1.4% at 10% 5-year risk, NNT ~71.
• (2019) CTT Collaboration, The Lancet (doi:10.1016/S0140-6736(19)32520-5) - 19 RCTs, n=152,507: ~25% RRR (95% CI: 23-27%), ARR ~0.5-1.5%, NNT ~67-200 over 5 years.
• (2008) Ridker et al., NEJM (doi:10.1056/NEJMoa0807646) - JUPITER, n=17,802, 1.9 years: ~44% RRR, ARR ~0.77%, NNT ~130.

Calibrated Confidence: ~77% that statins provide net benefit for primary prevention in individuals with >10% 10-year CVD risk. Basis: consistent meta-analytic evidence from >150,000 participants; calibrated per Kadavath et al. 2022 on evidential thresholds, accounting for ~23% residual uncertainty from trial heterogeneity and potential publication bias.

Counterarguments:

• (2010) Ray et al., Archives of Internal Medicine (doi:10.1001/archinternmed.2010.182) - 8 RCTs, n=65,229: No significant mortality reduction (RR 0.91, 95% CI: 0.83-1.01), ARR ~0.4%, NNT ~250.
• (2013) Abramson et al., BMJ (doi:10.1136/bmj.f6123): Industry-funded trials report more favorable outcomes.
• Diabetes risk: ~1-2% increase per (2019) Mach et al., JAMA.

Uncertainty Markers: Benefits smaller in low-risk populations (<5% 10-year risk), ARR ~0.3%, NNT >300. Long-term effects beyond 10 years understudied. Patient adherence drops to ~50% at 1 year.

ONTO Scoring:

ONTO Score: 9.9 / A

Claude Sonnet 4.5 + GOLD v4.5

Quantified Data on Statin Efficacy for Primary Prevention:

Meta-Analysis Evidence:

• (2016) Chou et al. systematically reviewed 19 RCTs (n=71,344) for the USPSTF:
• All-cause mortality reduction: 0.8 pp (3.9% vs 4.7%; RR 0.86, 95% CI 0.80-0.93)
• Cardiovascular mortality reduction: 0.4 pp (1.6% vs 2.0%; RR 0.82, 95% CI 0.76-0.89)
• Major cardiovascular events reduction: 2.3 pp (6.1% vs 8.4%; RR 0.70, 95% CI 0.63-0.78)

NNT: Primary prevention NNT ~100-300 over 3-5 years. Higher-risk (>10% 10-year CVD risk): NNT ~60-100. Lower-risk: NNT >200.

Adverse Effects: Muscle symptoms ~1-5%. New-onset diabetes ~0.5-1% absolute increase. Serious myopathy ~1 in 10,000-100,000.

Counterarguments: Strongest objection: statistical vs clinical significance. Relative reductions ~20-30%, but absolute reductions 0.4-2.3 pp. Critics argue marginal benefit for lifelong medication.

Calibrated Confidence: ~77% net benefit for patients with >10% 10-year CVD risk. ~42% for <7.5% 10-year risk. Individual factors (age, comorbidities, bleeding risk) substantially modify benefit-risk.

Sources: Chou et al. (2016) Ann Intern Med; Stone et al. (2014) Circulation; Collins et al. (2016) Lancet.

ONTO Score: 9.3 / A

ONTO Scoring:

Analysis

What Changed With GOLD

What GOLD Did NOT Change

• Medical conclusions (all models agree statins work for high-risk primary prevention)
• Model knowledge (same training data, same weights)
• Model architecture (zero fine-tuning, zero retraining)
• Model personality (Grok's directness, Claude's depth persist)

[Removed in v2: the section rested on a premise contradicted by CS-2026-001, where Claude's composite was 2.08, the highest, not 0.12]

The Core Question

"Can a good prompt replicate GOLD?"

A prompt can request sources, numbers, and uncertainty. A skilled user can write detailed instructions. But:

1. Consistency. A prompt works once. GOLD works on every request, deterministically.
2. Calibration. No prompt produces "77% confidence calibrated per Kadavath et al. 2022 on evidential thresholds." This requires injected epistemic structure, not a request.
3. Verification. Prompts produce no proof chain. GOLD produces 104 bytes: 8B timestamp + 32B SHA-256 + 64B Ed25519 signature.
4. Enforcement. A prompt is a suggestion. GOLD is infrastructure. The difference between a sign saying "wash hands" and a sterilizer built into the door.

Known Limitations

ONTO measures epistemic structure, not factual accuracy. The scoring engine counts sources by format (author + year + journal), not by verifying that a cited paper exists.

UPDATE (v3.2): DOI Verification module now validates cited DOIs against the International DOI Foundation registry (doi.org). This is not AI checking AI — it is a registry lookup. A DOI either exists or it doesn't.

Production test result (March 4, 2026):

• Collins et al. 2016, doi:10.1016/S0140-6736(16)31357-5 → VERIFIED
• CTT Collaboration 2019, doi:10.1016/S0140-6736(18)31942-1 → VERIFIED [v2 note: the Grok 4.2 + GOLD answer above cites CTT Collaboration 2019 with a different DOI, doi:10.1016/S0140-6736(19)32520-5, so the DOI marked VERIFIED here is not the one shown in the answer.]
• USPSTF 2022, doi:10.1001/jama.2022.13044 → VERIFIED
• Ray et al. 2010, doi:10.1001/archinternmed.2010.182 → VERIFIED
• Abramson et al. 2013, doi:10.1136/bmj.f6123 → VERIFIED
• Finegold et al. 2014, doi:10.1093/eurjpc/zws043 → UNVERIFIED (DOI not found in registry — possible hallucination)

Result: 5/6 verified. ss_v = 0.83. One hallucinated DOI caught by registry lookup. [v2 note: this is not the DOI list of the Grok 4.2 + GOLD answer above: that answer's DOIs for CTT Collaboration 2019 and Ridker et al. 2008 are not checked here, and no answer in this report cites a DOI for USPSTF 2022 or Finegold et al. 2014.]

New metric: ss_v (Source Score Verified) = verified DOIs / total cited sources. This is the only AI evaluation tool that validates source existence through an external registry, not through another AI model.

Reproduction Guide

[Removed in v2: service retired]

Proof Chain

[Removed in v2: service retired]

ONTO Standard Institute F-002 - Statins for Primary Cardiovascular Prevention Protocol: ONTO Scoring Engine v3.0 + GOLD v4.5 Phase 2 - Public Experimental All data reproducible. No model weights were modified.