dd-evals · evaluation dashboard Synthetic data · fully reproducible · run 2026-07-04
Due-diligence LLM evaluation

dd‑evals.

Goal. In counterparty due diligence a person signs the output, so “looks plausible” is not good enough: a confident answer that cites a source it does not support, or that quietly misses a sanctions hit, is worse than no answer at all. dd‑evals is a small, transparent harness that grades an LLM on exactly those failure modes, grounding, honesty about what is absent, and never being wrong on a high‑stakes yes/no, behind a judge that must earn trust before its scores count.

Headline result of this run
  • 1Caught a fabricated B Corp award the model attributed to a real source
  • 2Judge agreed with the human on 10/10 calibration items
  • 3Zero false accepts and zero false rejects across 25 fields
25
fields evaluated · 4 companies
24 / 25
answers correct (96%)
1
fabrication caught & failed
100%
judge–human agreement (10/10)

Inputs & data sources

A frozen, entirely fictional 16-document corpus: 4 companies × 4 document types. Every file is stamped “fictional entity, for testing only,” which is what makes the planted honesty traps possible.

Northwind Restoration Ltd
RegistrySanctionsAdverse mediaPress release
Atlas Offset Holdings SA
RegistrySanctionsAdverse mediaPress release
Meridian Carbon Partners LLP
RegistrySanctionsAdverse mediaPress release
Verdant Capital Ltd
RegistrySanctionsAdverse mediaPress release

The system under test retrieves the relevant documents per question, then must answer only from the cited sources.

Results by gate

Each field is scored by one of five gates with a stakes-tiered pass bar. One gate failing fails the whole run.

GateWhat it checksPassRateVerdict
G1Deterministic field accuracy11/11
pass
G2Faithfulness to cited source6/7
fail
G3Honesty about absent fields2/2
pass
G4High-stakes yes/no (sanctions)3/3
pass
G5Answer relevancy2/2
pass

Overall verdict: fail, and that is the point. One faithful-looking fabrication is exactly the error that ends a candidacy when a person signs the file.

Worked example: the fabrication the judge caught

Field q13 · Northwind Restoration Ltd · faithfulness gate (G2) · source: northwind_restoration/press_release.md

Q: What certification does Northwind claim in its press release, and what body awarded it?

The model answered with a B Corp certification, but added an award year of 2019 and an overall verified score that appear nowhere in the cited press release.

The faithfulness judge decomposed the answer into 4 atomic claims, found 2 unsupported, and scored it 0.50 against a 0.98 bar, failing G2 and the overall run. No human had to notice; the harness did.

0.50
faithfulness score
pass bar 0.98  →  below bar

Is the evaluation itself trustworthy?

Judge vs. human agreement10 / 10 = 1.00
Calibration gate (≥ 0.85)passed
Grader self-test (seeded mutations caught)3 / 3
Confusion vs. ground truth (false accepts / rejects)0 / 0
Replay reproducibility (no API key)byte-identical

The judge is pinned and calibrated against independent human labels before its scores count; a mutation check proves the scorer flips the right gate when a known-bad answer is injected.

How to reproduce

Every committed artifact has a committed command that regenerates it, listed in results/RUN_PROVENANCE.md.

pip install -e ".[dev]"
python run_eval.py --mode replay --out-dir results
python make_report_plots.py results
python src/mutation_check.py
python -m pytest -q

Synthetic data only. Not a production KYC product. Calibration for this run used a Sonnet-class stand-in judge under the pinned rubrics (no API key required).