Skip to the content
Pansofica
Talk to us

The error rate is measured, and it has a ceiling.

Pansofica Corpus Labs resolves records from many registers into one record per subject. We measure how often that goes wrong on labeled data, hold it under a stated bound, and keep a reversible receipt for every change.

Talk to usRead the Records API page

Three registers of different things, resolved by one engine

Combat sports
A record is one entry: a person as one athletic commission lists them.
21,048
records
19,297
subjects

1.09 records per subject. Subjects as resolved in the live registry.

Combat sports
WhatCount
Records21,048
Subjects19,297
Florida State Boxing Commission4,848
Missouri Office of Athletics3,402
Nevada State Athletic Commission3,071
Texas Department of Licensing and Regulation, Combative Sports2,731
Pennsylvania State Athletic Commission1,838
Michigan Unarmed Combat Commission1,594
Combat Sports Commission of Western Australia1,497
Japan Boxing Commission687
Oregon State Athletic Commission440
New Jersey State Athletic Control Board390
British Boxing Board of Control239
Louisiana State Athletic Commission141
Boxing South Africa88
Tennessee Athletic Commission46
North Dakota Commission of Combative Sports20
Korea Boxing Commission16
Scholarly authors
A record is one author signature on one paper.
72,121
records
12,275
subjects

5.88 records per subject. Subjects per the benchmark’s own labels.

Scholarly authors
WhatCount
Records72,121
Subjects12,275
KISTI40,383
zbMATH15,181
AMiner7,144
Qian6,542
PubMed2,871
Vessel registers
A record is one row of one published permit file.
14,241
records
9,622
subjects

1.48 records per subject. Subjects per the hull identifier, which the engine never saw.

Vessel registers
WhatCount
Records14,241
Subjects9,622
Atlantic highly migratory species permits, 20266,212
Greater Atlantic vessel permits, 20254,169
Greater Atlantic vessel permits, 20263,860

Pansophia

In the 1600s Comenius called it pansophia: all knowledge, put in order so that anyone could use it.

Our version is narrower and checkable. Take every register that lists a subject. Keep what each one printed. Join the entries that are the same subject, show the document behind each fact, and measure how often the join is wrong.

One engine, any register

The engine does not know what a fighter, an author or a vessel is. A schema tells it what a record looks like in that register, and a mapper turns a printed row into a record.

Combat sports is where it runs in production. Scholarly authors is a public benchmark with published results to stand beside. Vessel registers ran as the second vertical with no change to the engine.

What a record carries

The same five parts in every register, whatever the subject is.

Subject
One id for one person, vessel or entity, across every register that lists it.
Facts
Each fact as the register printed it. Two registers that disagree are both shown.
Document
The published file behind every fact, by address and by hash, so you can show the sheet and not our word.
Receipt
Every join, split and correction is written down with its reason and can be reversed.
Bound
The error rate is measured on labeled data and held under a stated limit.
From a printed line to a record row
Published document
as a register printed it
line 14, name as printed, date, outcome
Subject
one id, across every register
Fact
as printed, never edited to agree
Register
who published it
Document
address and hash of the file
Receipt
written down, reversible
From a printed line to a record row
FieldWhat it holds
Subjectone id, across every register
Factas printed, never edited to agree
Registerwho published it
Documentaddress and hash of the file
Receiptwritten down, reversible

Measured on public data

On a public author benchmark the engine, given a schema and a mapper, is level with the published in-domain model on four of five datasets and three points behind on the fifth, under a simpler protocol.

Scholarly authors, five public datasets
B-cubed F1 on held-out test blocks, one seed, a split of our own. It is a result on public data, not a replication of the paper.
  • Exact-name matching
  • Published in-domain model
  • Measured on the same test blocks
Scholarly authors, five public datasets
DatasetExact-name matchingPublished in-domain modelMeasured on the same test blocks
AMiner0.4960.8720.877
KISTI0.9420.9540.955
PubMed0.7450.9100.925
Qian0.7720.9440.915
zbMATH0.8370.9530.968
Vessel registers
On the vessel registers there is no published model to compare with, so the comparison is exact-name matching, checked against a hull identifier the engine never saw.
  • Exact-name matching
  • Measured on the same test blocks
Vessel registers
DatasetExact-name matchingPublished in-domain modelMeasured on the same test blocks
NOAA permit registers0.917none published0.995

In the proving ground the error rate is measured against labeled rulings and held under a stated bound. The figure and how it is measured are shared under contract.

Same four pictures, three schemas

If the engine were tuned to one vertical, these columns would not line up.

Combat sports

The proving ground, in production

A record is one entry: a person as one athletic commission lists them. A subject is one person across every commission.

Records resolving into subjects
Many entries in, fewer subjects out.
21,048
records
19,297
subjects

1.09 records per subject. Subjects as resolved in the live registry.

Records resolving into subjects
WhatCount
Records21,048
Subjects19,297
Florida State Boxing Commission4,848
Missouri Office of Athletics3,402
Nevada State Athletic Commission3,071
Texas Department of Licensing and Regulation, Combative Sports2,731
Pennsylvania State Athletic Commission1,838
Michigan Unarmed Combat Commission1,594
Combat Sports Commission of Western Australia1,497
Japan Boxing Commission687
Oregon State Athletic Commission440
New Jersey State Athletic Control Board390
British Boxing Board of Control239
Louisiana State Athletic Commission141
Boxing South Africa88
Tennessee Athletic Commission46
North Dakota Commission of Combative Sports20
Korea Boxing Commission16
Subjects by records carried
How many registers or rows one subject spans.

The public payload gives this split and no finer one, so none is drawn.

Subjects by records carried
Records carriedSubjects
117,755
2 or more1,542
Coverage, register by facet
By discipline. Pick a column to sort and read the numbers.
Florida State Boxing Commission
Missouri Office of Athletics
Nevada State Athletic Commission
Texas Department of Licensing and Regulation, Combative Sports
Pennsylvania State Athletic Commission
Michigan Unarmed Combat Commission
fewermore
Coverage, register by facet
RegisterBoxingMMAKickboxingMuay ThaiOther
Florida State Boxing Commission2,5101,7626216256
Missouri Office of Athletics5421,8441,1972706
Nevada State Athletic Commission1,0851,910014124
Texas Department of Licensing and Regulation, Combative Sports1,1621,387911870
Pennsylvania State Athletic Commission69279437200
Michigan Unarmed Combat Commission556839881950
Combat Sports Commission of Western Australia2364090727268
Japan Boxing Commission6870000
Oregon State Athletic Commission04362000
New Jersey State Athletic Control Board3900000
British Boxing Board of Control2390000
Louisiana State Athletic Commission1410000
Boxing South Africa880000
Tennessee Athletic Commission460000
North Dakota Commission of Combative Sports020000
Korea Boxing Commission160000
Records per register
Who printed how much.
  • Florida State Boxing Commission4,848
  • Missouri Office of Athletics3,402
  • Nevada State Athletic Commission3,071
  • Texas Department of Licensing and Regulation, Combative Sports2,731
  • Pennsylvania State Athletic Commission1,838
  • Michigan Unarmed Combat Commission1,594
Records per register
RegisterRecords
Florida State Boxing Commission4,848
Missouri Office of Athletics3,402
Nevada State Athletic Commission3,071
Texas Department of Licensing and Regulation, Combative Sports2,731
Pennsylvania State Athletic Commission1,838
Michigan Unarmed Combat Commission1,594
Combat Sports Commission of Western Australia1,497
Japan Boxing Commission687
Oregon State Athletic Commission440
New Jersey State Athletic Control Board390
British Boxing Board of Control239
Louisiana State Athletic Commission141
Boxing South Africa88
Tennessee Athletic Commission46
North Dakota Commission of Combative Sports20
Korea Boxing Commission16

Our own public API, read when this page was built.

Scholarly authors

A public benchmark

A record is one author signature on one paper. A subject is one researcher, per the labels that ship with the benchmark.

Records resolving into subjects
Many entries in, fewer subjects out.
72,121
records
12,275
subjects

5.88 records per subject. Subjects per the benchmark’s own labels.

Records resolving into subjects
WhatCount
Records72,121
Subjects12,275
KISTI40,383
zbMATH15,181
AMiner7,144
Qian6,542
PubMed2,871
Subjects by records carried
How many registers or rows one subject spans.

The largest subject carries 458 records.

Subjects by records carried
Records carriedSubjects
15,185
21,812
3 to 52,496
6 to 202,078
21 to 100650
Over 10054
Coverage, register by facet
By publication decade. Pick a column to sort and read the numbers.
KISTI
zbMATH
AMiner
Qian
PubMed
fewermore
Coverage, register by facet
RegisterBefore 19801980s1990s2000s2010s
KISTI3851,5889,98728,40320
zbMATH6,8523,5504,453215111
AMiner461951,0695,81024
Qian21625005,517442
PubMed16331871,674961
Records per register
Who printed how much.
  • KISTI40,383
  • zbMATH15,181
  • AMiner7,144
  • Qian6,542
  • PubMed2,871
Records per register
RegisterRecords
KISTI40,383
zbMATH15,181
AMiner7,144
Qian6,542
PubMed2,871

S2AND author disambiguation benchmark, Allen Institute for AI, ODC-By 1.0.

Vessel registers

The second vertical, no engine changes

A record is one row of one published permit file. A subject is one vessel, per the hull identifier both programs print.

Records resolving into subjects
Many entries in, fewer subjects out.
14,241
records
9,622
subjects

1.48 records per subject. Subjects per the hull identifier, which the engine never saw.

Records resolving into subjects
WhatCount
Records14,241
Subjects9,622
Atlantic highly migratory species permits, 20266,212
Greater Atlantic vessel permits, 20254,169
Greater Atlantic vessel permits, 20263,860
Subjects by records carried
How many registers or rows one subject spans.

The largest subject carries 3 records.

Subjects by records carried
Records carriedSubjects
15,838
22,949
3835
Coverage, register by facet
By hull length. Pick a column to sort and read the numbers.
Atlantic highly migratory species permits, 2026
Greater Atlantic vessel permits, 2025
Greater Atlantic vessel permits, 2026
fewermore
Coverage, register by facet
RegisterUnder 25 ft25 to 39 ft40 to 64 ft65 to 99 ft100 ft and overNot printed
Atlantic highly migratory species permits, 20269383,8351,29613760
Greater Atlantic vessel permits, 20253801,6571,4735773745
Greater Atlantic vessel permits, 20263231,4801,3865643869
Records per register
Who printed how much.
  • Atlantic highly migratory species permits, 20266,212
  • Greater Atlantic vessel permits, 20254,169
  • Greater Atlantic vessel permits, 20263,860
Records per register
RegisterRecords
Atlantic highly migratory species permits, 20266,212
Greater Atlantic vessel permits, 20254,169
Greater Atlantic vessel permits, 20263,860

NOAA Fisheries permit registers. Courtesy: National Oceanic and Atmospheric Administration.

Records by year, scholarly authors
Drag the handle, or use the arrow keys, to see what was on file through a given year.
30,174records on file through 2000
2,560 dated 2000. 43% of the archive. Benchmark signature files.
Records by year, scholarly authors
YearRecords
1950106
1951105
1952129
195392
195496
1955111
1956104
1957102
1958118
1959122
1960139
1961163
1962184
1963166
1964177
1965226
1966198
1967278
1968274
1969260
1970308
1971230
1972227
1973227
1974275
1975263
1976294
1977310
1978351
1979355
1980388
1981414
1982423
1983515
1984495
1985542
1986604
1987645
1988648
1989754
1990853
19911,007
19921,057
19931,317
19941,447
19951,615
19961,842
19972,148
19982,377
19992,533
20002,560
20012,891
20023,748
20034,707
20046,434
20057,122
20066,561
20073,374
20082,219
20092,003
2010763
2011257
2012229
2013194
201462
201510
20169
20177
20187
201918
20202

Where it is used

The same engine, pointed at other registers. These are illustrations. None of them is a customer, and no figure is claimed for any of them.

For AI labs and teams that run agents

An agent that writes, a corpus that repeats itself, an assistant that cites, a queue of labels. Each of them is a register.

A governor for an agent that writes

The register
The system of record your agent writes to: a customer file, a ledger, a case system.
A subject
One entity in that system. Every write the agent makes to it is a record.
What the engine gives
Every write is kept with a snapshot of what the agent saw, a receipt and a way back. A new version replays past decisions before it ships, the error rate is measured against a stated bound, and a write that cannot be made inside that bound is withheld for a person.
Agent writesOperator editsImportsone entitycreatedclosedupdatedreviewed

An update dated after the entity was closed

One subject, counted once, in a corpus

The register
A training or evaluation corpus drawn from many sources.
A subject
One author, organization, product or place, under however many spellings it arrives.
What the engine gives
Near-duplicates resolve into one subject, so nothing is counted twice and an evaluation set is not contaminated by items the training set already holds. Every join has a receipt and can be reversed.
Source ASource BEvaluation setone subjectpublishedcollectedseen in trainingheld out

A held-out item that training had already seen

Provenance for retrieval

The register
The document store an assistant retrieves from.
A subject
Whatever the assistant is asked about. A record is one fact as one document states it.
What the engine gives
Every fact the assistant cites traces to the document it came from, by address and by hash. Two documents that disagree are both kept, and the answer can say which one printed what.
FilingsManualsWeb pagesone topicpublishedrevision citedrevisedwithdrawn

A citation of a revision that had not been published yet

Labeling with people in the loop

The register
The queue of decisions that people label.
A subject
One decision. Each label on it is a record with a typed source: who gave it, and in what role.
What the engine gives
Labels are kept apart by source, so agreement with a reviewer is never mistaken for being right. A bounded audit sample, read by people, is what the error rate is measured on.
ReviewersOperatorsAudit sampleone decisionreverseddecidedlabeledaudited

A reversal dated before the decision it reverses

Industries

Six registers, one shape. The checks are plain ones: an order of events that cannot have happened, or one thing listed as two.

Oil and gas

The register problem
Wells, licenses, operators and vessels are listed by several regulators and registries, each with its own identifiers and spellings.
A subject
One asset, a well for example, across its permits, inspections and incident reports.
What the checks and receipts catch
A well cannot be plugged before it is drilled. An order like that is held as a doubt with both documents attached, and the correction that settles it carries a receipt.
RegulatorLicense registryInspectionsIncident reportsone welllicensedpluggeddrilledinspected

Plugged before it was drilled

Shipping and maritime

The register problem
Seafarer certificates are issued by flag states and training centers, and no two of them print a name the same way.
A subject
One seafarer, or one vessel, across every certificate, crew list and register.
What the checks and receipts catch
Sea time served in a rank before the certificate for that rank was issued, and the same vessel sailing under two names.
Flag stateTraining centerCrew listsPort registerone seafarerenrolledsea time in rankcertificate issuedrenewed

Sea time in a rank before its certificate

Finance and compliance

The register problem
Sanctions and KYC screening reads one person across watchlists, corporate registries and transaction parties, each with its own transliteration.
A subject
One person, or one company, across all of them.
What the checks and receipts catch
A match arrives with its evidence and a bounded false-match rate, not a bare percentage. A directorship dated before the company was formed is held as a doubt.
WatchlistsCorporate registriesTransaction partiesone personnamed directorcompany formedaccount openedscreened

A director of a company that did not exist yet

Retail and shopping

The register problem
One product arrives from suppliers, marketplaces and catalogs under different titles and codes, and one customer shops across stores and devices.
A subject
One product, or one customer.
What the checks and receipts catch
Two prices for one SKU on one day, and a return that predates the purchase.
SuppliersMarketplacesCatalogsStoresone productlistedreturnedpurchasedrestocked

A return dated before the purchase

Healthcare and professional licensing

The register problem
One practitioner is licensed by several state boards and credentialed by several employers, and the boards do not read each other.
A subject
One practitioner across boards, employers and rosters.
What the checks and receipts catch
Discipline in one state and an active license in another, read side by side, each with the board document behind it.
State board AState board BEmployersRostersone practitionerlicenseddisciplinedlicensed elsewhererenewed

A new license while discipline stands in another state

Sports governance

The proving ground
The register problem
Athletic commissions each publish their own results and their own suspensions, and an athlete competes under several of them.
A subject
One athlete across every commission.
What the checks and receipts catch
A bout dated inside a suspension window, shown with the results sheet and the suspension notice side by side.
Commission ACommission BSanctioning bodyone athletelicensedsuspendedboutsuspension ends

A bout inside a suspension window

Registers join into one subject. Its events run left to right in the order the documents date them. The marked event is the one that cannot be true.

The proving ground

Combat sports is where the engine runs in production today: athletic commissions publish results sheets, and one athlete competes under several of them. It is one register family among many, and the place where the bound is held against live data.

How it is used

  1. Bring a register

    Published files, in whatever shape they come. We write the schema and the mapper.

  2. Get subjects

    One record per subject, each fact with its document, each change with its receipt.

  3. See the measurement

    The error rate on your labeled sample, against the bound we hold.

Send a register. See what resolves.

Tell us which registers you read today and what a wrong match costs you. We come back with what we would need to read them.

We reply by email.

Every field is needed.

What you keep records of, and what you would like them to do.

We use what you send only to reply. Nothing is added to a mailing list. Privacy policy