Five data files describe the same companies. Two of them copy one upstream file, so when they agree, that is one opinion, not two.
This run finds that pair from the files, merges companies and contacts across all five, picks the right value for each field against a hand-checked sample, measures how old each record is, and keeps the source and date of every value it outputs.
- Files B and D likely share an upstream file. On the checked sample, 70% of their errors are the same wrong value in both files (next highest pair: 34%), and 72% of the companies they share carry the identical verified date.
- 531 company records became 159 companies. 700 contact records were matched inside them. 13 unclear matches wait for review instead of merging.
- Where the files disagree, the resolver picked the right value 78 times out of 94. One vote per file got 67. Scored on 50 hand-checked companies that did not set the weights.
- File B is the oldest: half its records were last verified over 495 days ago. Intent topics go out of date fastest, about 5.6% a month.
- All 1191 output fields carry a source file and a verified date.
The five files
Each file holds companies (name, domain, city, size, industry, tech in use, intent topic) and contacts (name, title, sometimes an email), each with a last-verified date. Today is fixed at 2026-11-01.
1. Overlap, and the pair that shares an upstream file
Two files that buy the same upstream file agree with each other, and they are wrong together. Agreement is not evidence on its own, because two accurate files agree too. The run looks for what copying leaves behind: the same wrong values and the same verified dates.
| Pair | Companies in both | Plain agreement | Shared mistakes | Same verified date | Reading |
|---|---|---|---|---|---|
| B D | 81 (69% of either) | 90.5% | 70% of 116 errors | 72% | Likely one upstream file: count once |
| B E | 55 (40% of either) | 60.8% | 34% of 98 errors | 0% | Independent |
| D E | 46 (35% of either) | 59.4% | 27% of 64 errors | 0% | Independent |
| C E | 64 (44% of either) | 62.3% | 17% of 83 errors | 0% | Independent |
| A C | 93 (62% of either) | 89.8% | 16% of 49 errors | 0% | Agree often, mistakes independent |
| C D | 71 (49% of either) | 73.3% | 15% of 75 errors | 0% | Independent |
| B C | 84 (57% of either) | 63.8% | 14% of 121 errors | 0% | Independent |
| A D | 69 (49% of either) | 72.7% | 11% of 72 errors | 1% | Independent |
| A B | 81 (57% of either) | 62.8% | 8% of 120 errors | 0% | Independent |
| A E | 58 (40% of either) | 60.9% | 7% of 82 errors | 0% | Independent |
Shared mistakes are counted on all 100 hand-checked companies. Rule: a pair is flagged when at least 40% of its checked errors are shared, that share is at least twice the median pair, and at least 30% of its shared companies carry the identical verified date.
2. One company and one contact, found across the files
Matching goes by domain first (after removing https, www, slashes and hyphens), then by name and city. A record with no domain merges only on an exact name in the same city; anything close waits for review. Two different domains never merge, even when the names look alike.
Orrinth Freight Co., as each file wrote it
| File | Name as written | Domain as written | Verified | How it matched |
|---|---|---|---|---|
| A | Orrinth Freight Co. | orrinth-freight.example | 2026-10-26 | first record: starts the company |
| B | ORRINTH FREIGHT, LLC | https://www.orrinthfreight.example | 2026-01-13 | same domain (orrinthfreight.example) |
| C | Orrinth Freight | orrinthfreight.example | 2026-07-02 | same domain (orrinthfreight.example) |
| D | Orrinth Freight, LLC | https://www.orrinthfreight.example | 2026-01-13 | same domain (orrinthfreight.example) |
| E | Orrinth Freight Inc. | www.orrinthfreight.example | 2026-01-24 | same domain (orrinthfreight.example) |
One contact inside it
| File | Name as written | Title as written | Verified | |
|---|---|---|---|---|
| A | Kenji Delacroix | none | RevOps Manager | 2026-10-26 |
| B | Kenji Delacroix | none | RevOps Manager | 2026-01-13 |
| C | Kenji Delacroix | on file | RevOps Manager | 2026-07-02 |
| D | Kenji Delacroix | none | RevOps Manager | 2026-01-13 |
| E | Kenji Dealcroix | none | RevOps Manager | 2026-01-24 |
Contacts merge on the same email, on the same name after nicknames are read as the full name (Bob as Robert), or on a close name with the same job title (VP Sales read as VP of Sales). Any other close name waits for review.
Waiting for review (5 companies, 8 contacts)
- C “Quillon Group, LLC”, no domain, close to “Qiullon Group” (name 0.79, same city)
- C “Sorvik Industreis Co.”, no domain, close to “Sorvik Industries, LLC” (name 0.83, same city)
- E “Wexomor Works”, no domain, close to “Wexmoro Works, LLC” (name 0.86, same city)
- E “Larkwlel Holdings”, no domain, close to “Larkwell Holdings Inc.” (name 0.89, same city)
- E “Quillon Freigthline”, no domain, close to “Quillon Freightline” (name 0.85, same city)
- C contact “Kenji Pterov”, close to an existing contact (name 0.77)
Look-alikes kept apart (4)
- A “Tamberlyn Partnering, LLC” (tamberlynpartnering.example) is not “Tamberlyn Partners”: different domain
- A “Quillon Freightline” (quillonfreightline.example) is not “Quillon Freight Co.”: different domain
- B “OSTERY LABS” (osterylabs.example) is not “Ostery Lab”: different domain
- D “Norvane Systems” (norvanesystems.example) is not “Corvane Systems Inc”: different domain
Against the generator’s answer key, which a real run would not have: 100.0% of automatic company merges are correct (702 of 702 merged pairs), and 100.0% of contact merges.
3. Conflicts, settled against a hand-checked sample
100 merged companies form the hand-checked sample: a person checks size, industry, contact titles and intent against the company itself (here the generator plays that person). The first 50set each file’s accuracy. The other 50 only score the result.
Accuracy per file, on the first 50
| File | Company size | Industry | Job title | Intent topic | All four |
|---|---|---|---|---|---|
| A | 100% of 40 | 95% of 40 | 90% of 51 | 88% of 40 | 93% |
| B | 75% of 32 | 88% of 32 | 50% of 38 | 41% of 32 | 63% |
| C | 90% of 42 | 95% of 42 | 78% of 51 | 95% of 42 | 89% |
| D | 93% of 28 | 89% of 28 | 74% of 34 | 61% of 28 | 79% |
| E | 76% of 25 | 76% of 25 | 60% of 35 | 36% of 25 | 62% |
Right value on the other 50, where the files disagree
| Field | Disagreements | One vote per file | Resolver |
|---|---|---|---|
| Company size | 21 | 17 | 18 |
| Industry | 9 | 9 | 9 |
| Job title | 33 | 21 | 26 |
| Intent topic | 31 | 20 | 25 |
| All four | 94 | 67 (71%) | 78 (83%) |
The resolver weighs each file by its measured accuracy on that field, lowers the weight of old records by the decay measured below, and counts the flagged pair as one file. A close vote is marked for review in the output.
4. Freshness: how old each record is, and how fast it decays
| File | Median age | Older than 12 months | Titles and intent out of date now (estimated) | In 6 months |
|---|---|---|---|---|
| A | 57 days | 0% | 8% | 30% |
| B | 495 days | 76% | 50% | 61% |
| C | 99 days | 0% | 13% | 33% |
| D | 389 days | 53% | 36% | 51% |
| E | 463 days | 61% | 46% | 58% |
Titles and intent topics still right, by record age
- 0 to 3 months89% of 279
- 3 to 6 months82% of 175
- 6 to 12 months49% of 85
- 12 to 24 months44% of 230
- 24+ months25% of 40
Measured change per month (on the first 50)
- Intent topic5.6%
- Job title3.4%
- Company size1.4%
- Industry1.2%
Same dates, same staleness: files B and D age together, so a refresh of one does not refresh the other.
5. Provenance: every output field keeps its source and date
| Field | Value | From | Verified | Why this value |
|---|---|---|---|---|
| Company name | Orrinth Freight Co. | A | 2026-10-26 | most recent record |
| Domain | orrinthfreight.example | A | 2026-10-26 | most recent record |
| Company size | 5000+ | A with D, C | 2026-10-26 | weighted vote |
| Industry | Manufacturing | A with C, D | 2026-10-26 | weighted vote |
| Tech in use | cloud CRM | A with C, E | 2026-10-26 | weighted vote |
| Intent topic | data enrichment | C with D, E | 2026-07-02 | weighted vote |
| Job title (Kenji Delacroix) | RevOps Manager | A with C, D, E | 2026-10-26 | all 5 sources agree |
| Job title (Selin Achterberg) | Head of Marketing | A with C, E | 2026-10-26 | all 3 sources agree |
When a value turns out wrong, its source and date point to the one file that supplied it.
Checks on this run
Recomputed in your browser on every sample. The same checks run at /api/sources-selftest.
- Pass: All five files load with rows 118 / 106 / 127 / 92 / 88
- Pass: Exactly one pair is flagged as sharing an upstream file 1 flagged
- Pass: The flagged pair is the pair the generator planted B-D
- Pass: Plain agreement alone does not single out that pair top raw agreement B-D 90.5%; B-D 90.5%; within 6 points: A-C
- Pass: Identical verified dates appear only in the flagged pair B-D 72%, max other 1%
- Pass: On the checked sample the pair shares more of its mistakes than any other pair 70% of its errors shared vs max other 34%
- Pass: Automatic company merges are at least 99% precise 100.00% of 702 merged pairs
- Pass: No look-alike company is merged into another 0 mixed entities
- Pass: Unclear company matches wait for review 5 in review
- Pass: Contact merges are at least 98% precise 100.00% of 797 pairs
- Pass: Accuracy is set on one half of the sample and scored on the other 50 + 50, overlap 0
- Pass: The resolver beats one-source-one-vote on the held-out half 78 vs 67 of 94
- Pass: Intent topics go out of date faster than titles, titles faster than industry per month: intent 0.056, title 0.034, industry 0.012
- Pass: Every output field carries a source and a date 1191 fields, 0 without
- Pass: Every cited source really holds that value on that date 0 of 1191 do not
- Pass: Every domain is an invented .example domain 481 domains
- Pass: The same sample gives the same result twice fca4331d
What this run does not show
- The data is invented and the shared upstream file is planted. Real vendor files are larger and messier, and their sourcing is often not published.
- The hand-checked sample is played by the generator. On real data a person checks those companies, and the sample size sets how sure each accuracy figure is.
- Matching rules are tuned to these files. Real files need their own review of the unclear matches before the rules are trusted.
- It runs on about 159 companies in a browser, not on millions of records.