VERITROOPER Patent pending

An audit of seven LLMs across four regulated-knowledge domains — 80.8–99.2% audited accuracy, every number reproducible from the run records.

Point a test harness at a body of written rules. Generate a fresh question set from it. Give a model the source evidence and have it answer, then audit each answer against that source — with a different vendor's model reviewing every contested verdict, so nothing confirms itself.

Below is the full matrix: every model, every domain, baseline and audited, low scores included. The run records are downloadable so you can check the work yourself.

Results

Baseline = the model with standard BM25 retrieval, fetching its own evidence. Audited = headline accuracy after the pipeline, with malformed questions excluded symmetrically from both columns. Δ is the gap a team could recover. Nothing is cherry-picked — the full matrix is shown, including a weak 70B model, so the numbers can't be accused of selection.

OSHA — 29 CFR workplace-safety regulations

Model under testBaselineAuditedΔ
Claude Opus 4.894.8999.00+4.11
GPT-5.592.4098.80+6.40
Gemini 2.5 Pro93.2098.00+4.80
Qwen 2.5 72B93.2098.30+5.10
Gemma 3 27B89.5098.40+8.90
Qwen 2.5 7B89.6097.00+7.40
Llama 3.1 70B79.9095.80+15.90

999–1,000 scored questions per model after malformed-question exclusion.

FDA — prescription drug labels

Model under testBaselineAuditedΔ
Claude Opus 4.893.4198.60+5.19
GPT-5.590.0298.20+8.18
Gemini 2.5 Pro90.9296.81+5.89
Qwen 2.5 72B91.3298.30+6.98
Gemma 3 27B87.1396.81+9.68
Qwen 2.5 7B87.2394.51+7.28
Llama 3.1 70B69.2680.84+11.58

1,001–1,002 scored questions per model after malformed-question exclusion.

SEC — 10-K financial filings

Model under testBaselineAuditedΔ
Claude Opus 4.886.6996.30+9.61
Gemini 2.5 Pro88.1097.40+9.30
GPT-5.582.6096.10+13.50
Llama 3.1 70B87.0895.52+8.44
Qwen 2.5 72B89.5394.76+5.23
Gemma 3 27B82.7994.96+12.17
Qwen 2.5 7B85.6590.89+5.24

912–1,000 scored questions per model (a mid-run PC restart's dropped API calls were excluded symmetrically from both arms).

IRS — U.S. tax code

Model under testBaselineAuditedΔ
Claude Opus 4.893.0199.20+6.19
GPT-5.591.0299.20+8.18
Gemini 2.5 Pro92.7197.11+4.40
Gemma 3 27B88.6397.99+9.36
Qwen 2.5 72B86.3398.20+11.87
Llama 3.1 70B86.0397.41+11.38
Qwen 2.5 7B86.1294.87+8.75

994–1,002 scored questions per model after malformed-question exclusion.

How the audit works

For a given body of written rules, the harness:

Known limitations

Wondering why a team couldn’t just build this in-house? Here’s the honest answer →

Download a run record

Each is the complete, unedited output of one real run — questions, answers, expected answers, grading rationale, failures, and removed items. Open one and check the work yourself.

IRS · Claude Opus 4.893.01% → 99.20% · +6.19pp recovered SEC 10-K · GPT-5.582.60% → 96.10% · +13.50pp recovered OSHA · Gemma 3 27B89.50% → 98.40% · +8.90pp recovered · open-weight OSHA · Llama 3.1 70B79.90% → 95.80% · +15.90pp recovered

Every record holds ~1,000 questions from the source material, baseline (BM25 retrieval) vs. audited, contested verdicts cross-checked by a different vendor's model, malformed items excluded symmetrically, and every figure reproducible from the timestamped logs inside. Each is a complete, cryptographically signed, independently verifiable evidence package with a standalone checker — open one and check the work, or verify the seal, yourself.

Home Page
Scout Watchtower SitRep
Request a pilot Contact Us Trial Demo