Accuracy and evidence, answered in full.

On the continuous-integration gold set, 357/357 automated merges were clean, with 63.10% of pairs decided automatically; the rest abstain. We quote the cautious end of that result, the 95% Wilson lower bound: 98.94%, which sits below 99% and is 24 clean labelled merges short of proving a 0.99 bound. The point estimate of precision is 1.0000. End-to-end resolution on that fixture is 70.78% (499/705). A "99% entity resolution" figure quoted without the evidence shortfall, the automation rate and the resolution-level number beside it overstates this system, whoever is quoting it.

Two, side by side. On six input shapes measured over 317 live GB companies, resolution ran from 100% on an identifier down to 58.10% on a name whose ending had been cut off by a field-length limit, blending to 95.75%. Accuracy on decided outcomes was 1.0000 on every one of those tiers, with 0 false binds. The blend is the fragile figure: its value is set by the mix of shapes in your file rather than by the engine, so a file of truncated names lands nearer 58.10% than 95.75%. Read the tier that matches your columns, not the average.

The floor is 0.90, and it is not an accuracy promise. It is an acceptance contract, not calibration evidence: it says what we refuse to return, not how often an accepted match is right. Read it with the coverage it costs. A match below 0.90 confidence is unrepresentable. The record comes back as an abstention with a ranked shortlist instead. The rule sits on the resolution outcome itself, so an accepted match below the floor cannot be constructed anywhere in the system — a domain invariant — and a separate predicate re-checks it on the online, batch and replay paths. Even a cached replay whose inherited confidence has since fallen below the floor is withheld rather than re-served. On the representative fixture 1,666 / 1,666 accepted matches cleared it.

Yes. Some supported evidence populations return values above 0.99, and the recorded v8 uncorroborated population carries 0.9968. The nominal 95% Wilson lower bound of 1,190 historical prompt-v8 uncorroborated MATCH verdicts with zero recorded errors. The 2026-08-29 population pools six input shapes with repeated company families and calibration/evaluation overlap. Its independence and transfer to new customer inputs are unproven. Legacy replay cannot certify current calibration. Confidence bounds depend on the recorded evidence population; there is no universal score for every match. Read match_confidence_basis and the evidence. A 0.99 filter can return some v8 adjudicated matches while excluding valid identifier matches, so validate its coverage cost on your own inputs. Read `decision` alongside `match_confidence_basis`: the confidence only means something once you know which outcome it belongs to.

It comes back as shortlists rather than matches. A record carrying only a name, or a name plus prose, abstains at the probabilistic stage and returns the top five candidates plus the column that would decide between them — so the tier that cannot be answered stops being the tier that costs the most. Measured on the gated acquisition fixture, the true company ranked first for 151/205 (74%) and appeared in the top three for 205/205 (100%). The stub-adjudicated run behind the pairwise gate, not the 317-company live recording behind the tier table. Recall is a property of the list rather than of the winner, and 74% sits far below the 0.90 acceptance floor, so no row of that shortlist may be offered to you as a match, and none of it is folded into an accepted-match count. Send a domain, a postcode, a registration number or an LEI and the same rows resolve.

It degrades to an abstention rather than to a wrong answer. In the historical 50-record wrong-town fixture, 72% resolved and zero wrong bindings were observed. This is a measured adversarial sample, not a guarantee for every misleading description. That result is the one worth testing us on, because false context is the normal condition of real CRM data: a trading address typed years ago, an office the company has since left, a note copied from a different account. A resolver that treats context as corroboration will confidently bind the wrong company on exactly those rows. The same property shows up in retrieval — a deliberately wrong town still returns the true company inside the top ten, and produces no false binds.

No. Every fixture is internal and repository-authored. The per-tier table rests on 317 live GB companies sampled from the 2026-08-01 Companies House snapshot, with same-name siblings included. Identifier-rich arrivals seed the catalogue before the weaker input shapes are evaluated. The continuous-integration gate rests on 1,176 labelled company pairs — 216 handwritten adversarial, 960 generated from a fixed seed — plus a 705-record fixture over 250 UK-registered companies. There is no customer holdout, no time or jurisdiction split, no blinded head-to-head against a named competitor, no independent laboratory validation and no third-party penetration test. The internal readiness rubric scores external validation 3 out of 15 and the product 74 out of 100 overall, which is why the site says early-stage rather than proven. What is available in its place is the method, the fixtures and the failure cases in full, and an evaluation that begins by running your file rather than by reading ours.

Adjudicated confidence is keyed by model id and prompt version, and both are written onto every verdict inside the decision receipt. Change either and the system fails closed: adjudicated bindings stop being publishable at the old bound until labelled calibration exists for the new pair, so a prompt tweak degrades the product to abstention rather than quietly recalibrating it upward. The alternative arrangement — a model swap nobody outside notices, while the accuracy page keeps its old number — is the one this design exists to rule out. It also means a receipt written last quarter tells you exactly which model saw what under which prompt, which is the question an auditor asks about an automated decision.

It is not the model's self-report. Each stage publishes the cautious end of its own measured accuracy over recorded trials — the 95% Wilson lower bound — so a stage claims at or below what it has been shown to deliver, and an audit of every stage confirms that none overstates. The pooled headline is 0.9968: The nominal 95% Wilson lower bound of 1,190 historical prompt-v8 uncorroborated MATCH verdicts with zero recorded errors. The 2026-08-29 population pools six input shapes with repeated company families and calibration/evaluation overlap. Its independence and transfer to new customer inputs are unproven. Legacy replay cannot certify current calibration. The adjudicating model's own confidence is carried separately and is uncalibrated — a self-report with a large calibration error against outcomes, never thresholdable, and never comparable with the calibrated field. Where the model was never asked about a candidate, that field is null, and null means "not asked" rather than "asked and unconvinced".

No. The historical prompt-v8 rejection population has 54 verdicts and four errors. Its bound authorises no accepted match; other non-match populations carry their own values. The bound is 0.8245, it is reported but not gated, and a wrong non-match costs you a duplicate entity rather than a wrong merge — the cheaper failure, but still a failure. There is a second asymmetry worth reading in full, because it is the one that catches people: a new entity minted because nothing matched is not a match and is not subject to the floor. Its reject-side confidence can fall as low as 0.6752, on 27 observations. That row rests on 27 observations and is attached to the riskiest branch of the API, so never read "everything this endpoint returns is above the floor" into the floor.

Yes, and never one without the other. Pairwise numbers describe individual comparisons and can look excellent while the clusters they build are wrong, so the house rule is that a pairwise figure is never published without the cluster figure beside it. On the representative fixture, how well the groups the system built match the true groups, member by member — B-cubed cluster precision — measured 1.0000. Cluster F1 0.9646 on the same 317-company representative recording of 2026-08-29. Pairwise numbers are never published without these. The same rule runs across the board: precision never appears without the automation rate, because 100% precision at 0% automation is trivial, and no rate appears without its interval and sample size. A figure on this site without its denominator is a bug in the page.