Measured September 2026 · sample 2026-09b
How often is a Fundz signal on the right company?
Every funding round, hire, filing and acquisition we show is attached to a company. We check that attachment on a blind random sample, publish the result with its margin of error, and show you how to check it yourself. Most data vendors don’t publish this number.
Right company, exact legal entity
90.6%
95% CI 84.4%–96.9%
Right company family
(parent, subsidiary or sibling counts)
(parent, subsidiary or sibling counts)
96.8%
95% CI 92.4%–100.0%
Weighted by how many links each data lane produces. Based on 873 decided links across 12 data lanes, drawn with seed e94c0ff9bc780226f1d53df8931de7a3. Excluding job postings, which are most of the volume: exact 89.6% (95% CI 86.7%–92.5%; n=797).
By data lane
Dot: exact legal entity. Ring: company family, shown where it differs. Bar: 95% confidence interval.
| Data lane | Links in window | Exact company | Company family |
|---|---|---|---|
| Form D filings | 1,207 | 100.0% (95% CI 95.2%–100.0%; 77 of 77) | 100.0% (95% CI 95.2%–100.0%; 77 of 77) |
| SEC 8-K filings | 846 | 96.0% (95% CI 89.0%–98.6%; 73 of 76) | 97.4% (95% CI 90.9%–99.3%; 74 of 76) |
| Executive hires | 2,132 | 96.0% (95% CI 88.9%–98.6%; 72 of 75) | 97.3% (95% CI 90.8%–99.3%; 73 of 75) |
| Funding news | 1,706 | 93.5% (95% CI 85.7%–97.2%; 72 of 77) | 93.5% (95% CI 85.7%–97.2%; 72 of 77) |
| Acquisitions: buyer | 2,851 | 93.5% (95% CI 85.7%–97.2%; 72 of 77) | 93.5% (95% CI 85.7%–97.2%; 72 of 77) |
| Contracts | 2,418 | 93.4% (95% CI 85.5%–97.2%; 71 of 76) | 94.7% (95% CI 87.2%–97.9%; 72 of 76) |
| Job postings | 145,948 | 90.8% (95% CI 82.2%–95.5%; 69 of 76) | 97.4% (95% CI 90.9%–99.3%; 74 of 76) |
| WARN layoff notices | 40 | 87.2% (95% CI 73.3%–94.4%; 34 of 39) | 97.4% (95% CI 86.8%–99.6%; 38 of 39) |
| Office permits | 107 | 87.5% (95% CI 77.9%–93.3%; 63 of 72) | 91.7% (95% CI 83.0%–96.1%; 66 of 72) |
| Product launches | 1,377 | 85.7% (95% CI 76.2%–91.8%; 66 of 77) | 85.7% (95% CI 76.2%–91.8%; 66 of 77) |
| Trademarks | 5,122 | 82.9% (95% CI 72.9%–89.7%; 63 of 76) | 92.1% (95% CI 83.8%–96.3%; 70 of 76) |
| Acquisitions: target | 2,762 | 82.7% (95% CI 72.6%–89.6%; 62 of 75) | 84.0% (95% CI 74.1%–90.6%; 63 of 75) |
Filings keyed by a government identifier (SEC CIK) are near-perfect. The misses cluster where a record names a company only by name: two different companies with the same name.
What the audit found, plainly
- Two precisions. Weighted by link volume, 90.6% of event links point at the exact legal entity (strict, 95% CI 84.4–96.9%) and 96.8% at the right company family (CI 92.4–100%); unweighted, 794 of 873 and 817 of 873. Excluding job postings, which are most of the volume, strict is 89.6% (86.7–92.5%) and family 92.4% (90.0–94.9%).
- Against the first September sample (2026-09, 77.2% strict). Most of the difference is a larger sample of the same job-posting data (19 of 25 then, 69 of 76 now), not a measured improvement: the intervals overlap, and events created after the 2026-09-27 ingest fixes are too few in this window to measure. The October sample will measure them.
- Links keyed by an identifier are near-perfect. Form D 77 of 77, 8-K filings 73 of 76, executive hires 72 of 75, funding news 72 of 77, acquirers 72 of 77, contracts 71 of 76.
- Name-matched lanes are where the errors are, and the error is almost always a different company with the same name. Trademarks 63 of 76 strict and 70 of 76 family (after 10,828 loosely matched trademark links were detached on 2026-09-27), permits 63 of 72, product launches 66 of 77 (for example a startup called Brain linked to Brain Corp), WARN notices 34 of 39 strict and 38 of 39 family.
- Acquired companies. 57 of 75 were right before 524 placeholder links ("Not specified", "N/A", blank) were removed on 2026-09-27, and 62 of 75 after (the stratum re-drawn with the same seed). The remaining misses are mostly a buyer or seller filed on the wrong side of the deal, and same-name companies.
- 8-K completeness depends on how filings are resolved. Through the parser-company path this sampler measures, 39 of 40 unresolved filings had a company with the exact CIK. The product now resolves 8-Ks by CIK first (deployed 2026-09-27): across all 1,896 September filings, 1,682 (88.7%) resolve to exactly one company and 210 to none. The precision of the CIK path is not yet measured.
- Missed links elsewhere are rare. Of unlinked records with a candidate shown, 5 of 26 WARN notices, 1 of 25 permits and 3 of 22 trademarks should have matched; 3 of 42 newly created companies duplicated an existing one. Records whose candidate search returned nothing are excluded, so these are upper bounds.
- Duplicates. 14 of 74 randomly drawn companies that had any candidate had a duplicate row (18.9%, CI 11.6–29.3%), a lower bound. Candidate pairs sharing a parser company id were the same company 20 of 20 times and a normalised name 16 of 19; a shared LinkedIn page 14 of 20, CIK 12 of 18 and website domain 11 of 17, which is why none of those is a merge key on its own.
- Investor links. In this blind sample 27 of 34 decidable links were right, but 43 of 77 could not be decided from stored evidence, so that figure stays below the publication rule; the separate source-article check (93.7%) is published beside it.
- Raters. Across 1,384 rows the primary rater and the model second rater agreed 93.4% of the time (Cohen's kappa 0.89). Both are Claude models, and so is the adjudicator of the 91 disagreements; no row has a human label.
What we miss, and duplicates
| Question | Result |
|---|---|
| Unlinked WARN notices that should have matched a company | 19.2% (95% CI 8.5%–37.9%; 5 of 26) |
| Unlinked office permits that should have matched a company | 4.0% (95% CI 0.7%–19.5%; 1 of 25) |
| Unlinked trademarks that should have matched a company | 13.6% (95% CI 4.8%–33.3%; 3 of 22) |
| 8-K filings with no matched company where one existed | 97.5% (95% CI 87.1%–99.6%; 39 of 40) Measured on the parser-company path (sampler v1). Since 2026-09-27 the product resolves 8-K filings by CIK first, which this figure does not measure; it describes the gap that change closes. |
| New companies that duplicated an existing one when created | 7.1% (95% CI 2.5%–19.0%; 3 of 42) |
| Companies with a duplicate row (random sample) | 18.9% (95% CI 11.6%–29.3%; 14 of 74) lower bound: only generator-surfaced candidates are checked |
| Candidate pairs sharing a normalised name that are really one company | 84.2% (95% CI 62.4%–94.5%; 16 of 19) |
| Candidate pairs sharing a website domain that are really one company | 64.7% (95% CI 41.3%–82.7%; 11 of 17) |
| Candidate pairs sharing a LinkedIn company page that are really one company | 70.0% (95% CI 48.1%–85.4%; 14 of 20) |
| Candidate pairs sharing a parser company id that are really one company | 100.0% (95% CI 83.9%–100.0%; 20 of 20) |
| Candidate pairs sharing a CIK after zero-padding that are really one company | 66.7% (95% CI 43.8%–83.7%; 12 of 18) |
Investors
Blind audit from stored evidence: 79.4% (95% CI 63.2%–89.6%; 27 of 34) 43 of 77 sampled links could not be decided from stored evidence
Checked against each round’s source article (2026-09-28, 160 of 320 sampled links readable): 93.7% (95% CI 88.8%–96.6%; 149 of 159).
Only 160 of the 320 sampled links could be measured. Links whose round source is BusinessWire, FinSMEs or a Google News redirect were excluded (the page could not be fetched or decoded into article text), as were links with no stored source URL or an unreadable page. The measured half therefore over-represents rounds reported by sources that serve plain article text, and the excluded sources may have a different precision. It is a second, independent measurement of investor-link precision, not a replacement for the blind audit's p_investor_link stratum (which could decide only 16 of 30 from stored evidence).
How we measure
- Sampling
- Every month a stratified random sample is drawn from the links Fundz made in that month. The seed is recorded before any row is labelled, and the order of rows within each stratum is md5(seed | stratum | row id), so anyone with the seed can re-draw exactly the same rows. The evidence each rater sees is frozen at draw time with a SHA-256 checksum, so a label can be re-checked even after the underlying data changes.
- What is sampled
- Precision strata sample event-to-company links per data lane (funding news, Form D, executive hires, contracts, acquisitions by role, product launches, job postings, WARN notices, permits, 8-K filings, trademarks) and investor-to-round links. Completeness strata sample records that were NOT linked and companies that were newly created, and ask whether an existing company was missed. Pair strata sample the current output of each duplicate-candidate generator. A duplicate stratum samples companies uniformly at random and asks whether each has a duplicate row.
- Labelling
- Each row is labelled by a primary rater from the frozen evidence. A second rater, a language model (disclosed by model name), labels the same rows blind to the primary label. Agreement is reported per stratum as raw agreement and Cohen's kappa. Every disagreement is re-read and adjudicated with a written note; the adjudicated label is final. Labels are append-only.
- Who labelled
- The primary rater for the first samples is Claude (an AI model) working in supervised sessions, and the second rater is also a model. No human labeller has labelled these samples. For that reason every audited row, its evidence and both labels are published, and a buyer is invited to draw and label a fresh sample of their own with the published seed procedure. Across 1,384 rows the two raters agreed 93.4% of the time (Cohen’s kappa 0.8939).
- Statistics
- Proportions are published with 95% Wilson score intervals. The cross-lane figure is a population-weighted stratified estimate with a finite-population correction. Rows a rater could not decide are reported as unknown and excluded from the denominator, never counted as correct.
- What runs automatically
- An automated analyst runs daily checks over new links and the duplicate backlog. A check may change data without a person only when its precision on a frozen, versioned gold set is at least 0.97 over at least 30 labelled rows, the change is reversible, it touches no company on an active customer's watchlist or lead list, and it is within a per-run cap. Everything else goes to a human review queue.
- Publishing rule
- A metric is published only with its sample size, its 95% interval and its measurement month, and only when its sample is at least 30 decided rows. Anything else is shown as not yet measured, with the reason. Limitations are always published beside the numbers.
Limitations
- Completeness is measured only against the candidate companies the evidence generator surfaces (name similarity and shared identifiers). A missed match with no shared name or identifier is invisible to it, so completeness is an upper bound and the duplicate rate a lower bound.
- Two different companies with the same name are the dominant precision failure; a name match with no second identifier is the weakest link method and is labelled as such.
- WARN notices, permits and trademarks mostly concern employers that are not in the company spine at all; their low link rate is a coverage fact, not a matching defect, and their precision is measured separately.
- Investor identity has no registry identifier (no CIK/LEI for most funds), and only a short summary of the source article is stored, so many investor-to-round links cannot be decided from stored evidence and are reported as unknown.
- The raters are AI models; agreement between two models is not the same as agreement with a human expert.
- Single-lane samples of 30 to 80 rows give intervals roughly 6 to 15 points wide; the cross-lane estimate is tighter, and single-lane figures are indicative. Differences between months within overlapping intervals are not evidence of change.
- Prediction calibration is not measurable until each prediction's horizon (90 or 180 days) has closed after snapshots began.
- The second rater (claude-sonnet-5, prompt rater-v2) was run through Claude Code sessions with the exact published prompt, not through the metered API call the monthly job uses, so its spend is not in the usage ledger. The model and prompt are the same.
- The adjudicator (Claude Opus 5.5) is the same model family as the primary rater and sided with it on most disagreements; agreement between the primary and the adjudicated label is therefore not an independent check.
- The acquired-company stratum was re-drawn with the same seed after 524 placeholder links were removed; the pre-cleanup result is published beside it in the sample notes.
Predictions
We freeze every prediction on the day it is made and grade it once its horizon has passed. None of the current model’s horizons has closed yet, so its calibration is not yet measured:
- Likely to raise next (180 days): first measurable 2027-06-24
- Likely to be acquired (90 days): first measurable 2027-03-26
- M&A brief: Acquires N% in 6 mo (180 days): first measurable 2027-06-24
- company page next-event panel (90 days): first measurable 2027-03-26
Check it yourself
- The numbers on this page are version-controlled JSON: api.fundz.net/v1/data_quality.
- The sample is reproducible. Seed
e94c0ff9bc780226f1d53df8931de7a3, sampler v1, window 2026-09-01 to 2026-09-27. ORDER BY md5('<seed>|<stratum>|' || <row id>) within each stratum's population filter; first n rows. - Every audited row, its frozen evidence and both labels are in the data-room CSV for this month. Evaluating Fundz? We will draw a fresh sample with a seed you choose and show you every row — data@fundz.net.