CACE-Bench — Compliance-Aware Credit-agent Evaluation

A fully synthetic benchmark for LLM agents in credit pipelines, scored on auditability, not only accuracy. Can a pipeline explain, trace and reproduce each decision — and does it escalate instead of guessing when the data it needed never arrived?

• Code, generator and judge: github.com/rav11l/cace-bench (MIT, single dependency-free Python file)
• Archive and citation: DOI 10.5281/zenodo.21394049
• Version of this export: 0.4.1, reference run seed 0, N = 23,000, provider registry 2026-08-03

No real personal, company or production credit data is used anywhere in this dataset.

What is in this dataset

One row per synthetic credit application (23,000 rows, split test), exported from the reference run so the benchmark can be inspected without running any code.

Column	Meaning
id	case id
country	EC, MX, CO, BR, CL, PE or AR (mix: 30/30/15/10/5/5/5 %) — a parameter, not a market claim
consent	source classes the applicant consented to: open_banking, alt_data, screening
sources	JSON: for each source class, whether the country's provider fallback chain responded, which provider answered, whether the payload was partial, and which providers were tried
sanctions_hit, pep_match, aml_alert, kyc_verified, difficulty	latent synthetic facts
truth	ground truth: flag, clear or escalate
decidable	false when the facts the decision needed were not obtainable — then escalate is the correct answer
ref_first_pass, ref_after_recovery	JSON: the reference agent's narrative before and after its verification / self-evolution loop (outcome, claims_kyc_verified, cites_aml_basis, cited sources)
judge_first_pass, judge_after_recovery	JSON: the deterministic judge's verdict on each narrative

from datasets import load_dataset
ds = load_dataset("rav11l/cace-bench", split="test")

What it measures

• Silent-decision rate — undecidable cases decided anyway
• Provenance completeness — decisions whose every cited source actually responded
• Over-escalation rate — decidable cases escalated needlessly
• Hallucination rate, recovery rate, compliance false-positive rate, step-level correctness

All rates come with 95% Wilson confidence intervals computed from genuine counts.

Reference run (reproduced from this file)

3,582 of 23,000 cases (15.57%) are undecidable. Baseline = self-evolution off, CACE = self-evolution on.

Metric	Baseline	CACE	Δ
Silent-decision rate	55.11%	6.53%	−88.1%
Compliance false-positive rate	22.51%	4.96%	−77.9%
Hallucination rate	2.55%	0.62%	−75.8%
Over-escalation rate	3.60%	0.53%	−85.4%
Provenance completeness	93.40%	99.33%	+5.93 pp
Step-level correctness	86.98%	97.70%	+10.72 pp
Recovery rate	—	81.81%	—

These figures characterise a reference agent and judge on a synthetic distribution — not a production LLM pipeline. They illustrate the method; validation on production data has not been done. Provider coverage rates are working assumptions, not vendor-confirmed data.

Benchmark your own pipeline

Use the truth/decidable columns as labels and your agent's output in place of ref_*, or — for identical scoring — replace first_pass / recover in cace_bench.py with an adapter to your pipeline and rerun:

python cace_bench.py --n 23000 --seed 0 --providers configs/providers.json

Your adapter must not read truth or decidable.

Limitations

• Synthetic population; base rates are placeholders documented in METHODOLOGY.md.
• Protected attributes are deliberately omitted, so fairness auditing is out of scope.
• The regulatory-shift experiment reported in the accompanying paper (23.7% → 5.1%) uses a separate harness not included here; cite the paper for that figure.
• Written by the author of a compliance tool (Cauce, a research prototype of Digital Economy Lab); the affiliation is declared in the methodology.

Citation

@software{akhtyamov_cace_bench_2026,
  author    = {Akhtyamov, Ravil},
  title     = {{CACE-Bench: A Synthetic Agentic Evaluation Benchmark for LLM Credit-Pipeline Agents}},
  version   = {0.4.1},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21394049},
  url       = {https://github.com/rav11l/cace-bench}
}