ACCURACY METHODOLOGY
What "0% false merges" means, and how we measure it.
Entity resolution is the safety-critical step of screening: merge two different parties and every downstream verdict is wrong. TrueParty publishes the figure, so it also publishes the method.
Figures from the evaluation of 22 June 2026 on matcher v4, the version in production. Re-run on every matcher change.
The design: three verdicts, and abstaining is a feature
Resolution runs in two stages: a high-recall retrieval pass (fuzzy, transliteration-aware) surfaces candidates, then a fine-tuned matcher issues one of three verdicts on each candidate, with a plain-English reason and a confidence:
SAME PARTY
Committed only when the evidence supports identity, not name similarity.
DIFFERENT PARTY
Cleared with the reason stated: a conflicting identifier or birth year is decisive.
UNCLEAR → REVIEW
Genuinely ambiguous cases go to a human review queue. The designed default is to abstain, never to guess.
The numbers
0
false merges across the 1,000-pair benchmark and the adversarial stress set
99.66%
committed accuracy on the benchmark (accuracy of the verdicts the matcher commits to)
99.3%
three-way accuracy including the review verdict
- Main benchmark: 1,000 pairs (464 same-party, 431 different-party, 105 review-gold): real pairs from Companies House, sanctions lists, SEC EDGAR, Wikidata and GLEIF records, plus constructed adversarial pairs (look-alike companies, identifier-backed positives, bare-name traps, person look-alikes).
- Adversarial stress set: 45 hand-built pairs, each constructed to attack a specific rule: near-duplicate names, identical names with different identifiers, shared agent addresses across different companies, transliteration variants, shared surnames, generic names.
- Targeted probes for the failure classes we have seen in the wild: identifier-backed positives, bare-name queries, common-name traps.
The guardrails
Each rule below can only downgrade a match or send it to review. None can invent a hit:
- A name alone is never enough to auto-confirm.
- A matching registration number or LEI is decisive; a conflicting one clears a look-alike, whatever the name similarity.
- A conflicting date-of-birth year clears a person match.
- A common name without an identifier is held for human review.
- A longer or extended name goes to review rather than being assumed identical.
- PEP status is treated as enhanced due diligence, never presented as a sanctions alert.
The regression gate
Every change to the matcher or retrieval re-runs the fixed benchmark, the stress set and the full API and browser regression suite. A change is rejected if retrieval recall drops on any query type, or if it produces a single false merge. Recall improvements are made under that constraint: for example, dropped-patronymic name queries improved from roughly 60% recall into the 90s while the zero-false-merge bar held.
WHAT THIS CLAIM IS, AND IS NOT
It is a measured result on a fixed, adversarially constructed evaluation, re-run on every change to the system, with the bar set at zero.
It is not a guarantee over every possible input. No matcher can promise that. The design response to uncertainty is abstention: when the evidence is thin or contradictory, TrueParty routes the case to a human with its reasoning stated, rather than committing either way. That behaviour is itself part of the evaluation.
