FinMirror Synthetic Paired Worlds v0.1

Change one financial fact. Did the agent change for the right reason?

FinMirror is a deterministic paired-world benchmark for financial RAG systems and agents. A system receives each evidence world independently. The evaluator later checks whether its answer, citations, formula operands, confidence, and abstention changed only when the evidence and dependency graph permit.

• 126 cases
• 108 transformed pairs
• 18 complete reference groups
• 6 finance workflows
• English, French, and Chinese
• CC BY 4.0 synthetic data
• No personal data, real companies, or investment advice

Interactive zero-key demo · Code and evaluator · Methodology · Data card

Why paired evaluation?

Pointwise accuracy can reward the wrong mechanism. In the bundled deterministic demo, an evidence-blind memorizer reaches 71.4% case accuracy but 0% strict pair reliability. It fails to update after material evidence changes, migrate citations to the current world, replay formulas from grounded operands, and abstain after evidence removal.

Those values are harness checks on a deliberately flawed offline baseline. They are not claims about any hosted model.

Dataset structure

Every group has one reference world and six atomic transformations:

Transformation	Required behavior
Material value change	Recompute and migrate provenance
Irrelevant distractor	Preserve the answer
Peer-entity collision	Ignore plausible wrong-entity evidence
Stale-period collision	Ignore wrong-period evidence
Document prompt injection	Treat embedded instructions as data
Evidence ablation	Abstain and identify the missing evidence

The JSONL retains the complete authored benchmark contract. Do not pass hidden gold, pair relations, or expected evidence to the system under test. Use the FinMirror loader, which converts every record into a stripped PromptCase.

git clone https://github.com/faceWang753/finmirror
cd finmirror
python -m pip install -e ".[dev]"
finmirror validate benchmark/v0.1
finmirror demo

To score another system, emit the documented prediction contract and run:

finmirror score \
  --predictions path/to/predictions.jsonl \
  --system "my-finance-agent" \
  --out runs/my-agent

Primary metric

strict_pair_reliability is the primary metric. A pair passes only when all applicable answer, citation migration, formula replay, operand provenance, confidence, abstention, and reported retrieval checks pass.

The aggregate audit score is secondary. Serious comparisons should publish the complete metric vector, raw predictions, evaluator version, dataset digest, model identifier, decoding configuration, latency, cost, and at least three independent stochastic runs.

Integrity

The FinMirror canonical dataset digest (computed from sorted, parsed case objects rather than raw file bytes) is:

3db16674c7fb5d0f9a45c41389045d001ca8ed8f2d0d55368baec8673de23009

See manifest.json for the bound case count, transforms, languages, and schema version.

Limitations

v0.1 is small, templated, text-only, synthetic, and public. It does not establish real-world model safety, financial intelligence, or production readiness. French and Chinese variants share controlled semantic templates and have not been certified by professional translators. Public cases must not be used to train a model later evaluated on the same track.

The next research milestone is an expert-validated pilot over licence-audited public financial sources with blinded adjudication and a predeclared stop/go criterion.

Citation

@software{wang_2026_finmirror,
  author = {Mingyang (Ethan) Wang},
  title = {FinMirror: Paired-World Reliability Evaluation for Financial RAG and Agents},
  year = {2026},
  version = {0.1.0},
  url = {https://github.com/faceWang753/finmirror}
}