Our data · accuracy audit month September 2026 · counts as of 16 Nov 2026

Every signal we ship is measured.

Fundz collects records from SEC filings, government registries, press wires and the news, reads them with AI, ties each one to the right company and scores it. This page shows how much we hold, how each step works, where AI and machine learning are used, and how often the result is right, measured on blind random samples you can redraw yourself.

47.2M
records held
4.2M
added in the last 30 days
239K
companies in the spine
90.6%
of events on the exact right company95% CI 84.4%–96.9%

01 · What we hold

47.2M records across 21 families

Counted live from the production database on 16 Nov 2026. These are records held, not distinct companies: some families carry amendments and quarterly re-filings. Where checks hold rows back, the smaller figure is what customers see.

27M
13F investment holdings
SEC EDGAR · +2.9M in 30 days
14.9M
Benefit plan filings (Form 5500)
US Department of Labor · +962K in 30 days
2M
New business formations
State business registries · +131K in 30 days
1.2M
Trademark filings
USPTO
712K
Job postings
Company applicant-tracking boards · +148K in 30 days
432K
H-1B labor applications
US Department of Labor
197K
Form D offerings
SEC EDGAR · +4,107 in 30 days
88K visible to customers
131K
UCC lien filings
State secretaries of state · +6,700 in 30 days
121K
Investor-to-round links
Derived from rounds · +5,416 in 30 days
64K
Executive changes
Press wires, news, SEC 8-K · +2,477 in 30 days
62K visible to customers
60K
SEC 8-K filings
SEC EDGAR · +3,469 in 30 days
56K
Funding rounds from news
Press wires and news · +2,026 in 30 days
50K visible to customers
55K
Acquisitions
Press wires, news, SEC · +4,549 in 30 days
39K visible to customers
50K
Contracts and agreements
Press wires, news, SEC · +2,614 in 30 days
47K visible to customers
47K
Major company events
SEC 8-K and news · +3,230 in 30 days
33K
Product launches
Press wires and news · +1,369 in 30 days
32K visible to customers
29K
Office-lease permits
City open-data portals · +3,679 in 30 days
18K
Crowdfunding offerings
SEC EDGAR (Form C) · +131 in 30 days
14K
Data-breach notices
State attorneys general · +220 in 30 days
7,957
WARN layoff notices
State labor departments · +220 in 30 days
6,986
10-K and 10-Q filings
SEC EDGAR · +766 in 30 days
239K
canonical companies, each event tied to one
54K
investors with their rounds linked
104K
rows held back by a check, with a reason, never deleted

02 · How accurate

Measured at two levels

First, is the event on the right company? Then, is the whole record right: the company, what happened, when, and the source?

Level 1

Right company

Monthly blind audit of 873 event-to-company links across 12 data lanes, drawn with seed e94c0ff9bc780226f1d53df8931de7a3.

90.6%
exact legal entity95% CI 84.4%–96.9%
96.8%
right company family (parent, subsidiary or sibling counts)95% CI 92.4%–100.0%

Weighted by how many links each lane produces. Excluding job postings, which are most of the volume: exact 89.6% (95% CI 86.7%–92.5%; n=797).

Dot: exact legal entity. Ring: company family, where it differs. Bar: 95% confidence interval.

50%60%70%80%90%100%Form D filingsForm D filings: 100.0%100%SEC 8-K filingsSEC 8-K filings: 96.0%96%Executive hiresexecutive hires: 96.0%96%Funding newsfunding rounds (news): 93.5%94%Acquisitions: buyeracquisitions: acquirer: 93.5%94%Contractscontracts / agreements: 93.4%93%Job postingsjob postings (ATS): 90.8%91%WARN layoff noticesWARN layoff notices: 87.2%87%Office permitsoffice lease permits: 87.5%88%Product launchesproduct launches: 85.7%86%Trademarkstrademarks (USPTO): 82.9%83%Acquisitions: targetacquisitions: acquired company: 82.7%83%
Lane table
Data laneLinks in windowExact companyCompany family
Form D filings1,207100.0% (95% CI 95.2%–100.0%; 77 of 77)100.0% (95% CI 95.2%–100.0%; 77 of 77)
SEC 8-K filings84696.0% (95% CI 89.0%–98.6%; 73 of 76)97.4% (95% CI 90.9%–99.3%; 74 of 76)
Executive hires2,13296.0% (95% CI 88.9%–98.6%; 72 of 75)97.3% (95% CI 90.8%–99.3%; 73 of 75)
Funding news1,70693.5% (95% CI 85.7%–97.2%; 72 of 77)93.5% (95% CI 85.7%–97.2%; 72 of 77)
Acquisitions: buyer2,85193.5% (95% CI 85.7%–97.2%; 72 of 77)93.5% (95% CI 85.7%–97.2%; 72 of 77)
Contracts2,41893.4% (95% CI 85.5%–97.2%; 71 of 76)94.7% (95% CI 87.2%–97.9%; 72 of 76)
Job postings145,94890.8% (95% CI 82.2%–95.5%; 69 of 76)97.4% (95% CI 90.9%–99.3%; 74 of 76)
WARN layoff notices4087.2% (95% CI 73.3%–94.4%; 34 of 39)97.4% (95% CI 86.8%–99.6%; 38 of 39)
Office permits10787.5% (95% CI 77.9%–93.3%; 63 of 72)91.7% (95% CI 83.0%–96.1%; 66 of 72)
Product launches1,37785.7% (95% CI 76.2%–91.8%; 66 of 77)85.7% (95% CI 76.2%–91.8%; 66 of 77)
Trademarks5,12282.9% (95% CI 72.9%–89.7%; 63 of 76)92.1% (95% CI 83.8%–96.3%; 70 of 76)
Acquisitions: target2,76282.7% (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, duplicates and investors
QuestionResult
Unlinked WARN notices that should have matched a company19.2% (95% CI 8.5%–37.9%; 5 of 26)
Unlinked office permits that should have matched a company4.0% (95% CI 0.7%–19.5%; 1 of 25)
Unlinked trademarks that should have matched a company13.6% (95% CI 4.8%–33.3%; 3 of 22)
8-K filings with no matched company where one existed97.5% (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 created7.1% (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 company84.2% (95% CI 62.4%–94.5%; 16 of 19)
Candidate pairs sharing a website domain that are really one company64.7% (95% CI 41.3%–82.7%; 11 of 17)
Candidate pairs sharing a LinkedIn company page that are really one company70.0% (95% CI 48.1%–85.4%; 14 of 20)
Candidate pairs sharing a parser company id that are really one company100.0% (95% CI 83.9%–100.0%; 20 of 20)
Candidate pairs sharing a CIK after zero-padding that are really one company66.7% (95% CI 43.8%–83.7%; 12 of 18)
Investor-to-round links, blind audit from stored evidence79.4% (95% CI 63.2%–89.6%; 27 of 34)
43 of 77 sampled links could not be decided from stored evidence
Investor-to-round links, checked against each round’s source article (2026-09-28)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).
Level 2

Whole record right

A fresh random draw per lane, judged field by field against the original filing, article or company site. 7 of 8 lanes are above their bar today. A lane below its bar shows no figure until a fresh draw clears it.

Funding rounds

Above our bar
97%no material error

95% CI 83.3%–99.4% · 29 of 30 rows · strict, every arguable call counted wrong: 70%

Funding rounds as served by the Fundz API, event in the last 90 days. Measured 20 Oct 2026, seed yr-today-2026-10-20.

SEC 8-K events

Above our bar
97%no material error

95% CI 83.3%–99.4% · 29 of 30 rows · strict, every arguable call counted wrong: 93%

8-K events filed in the last 90 days, checked against the EDGAR primary document. Measured 20 Oct 2026, seed yr-today-2026-10-20.

Executive changes

Above our bar
90%no material error

95% CI 74.4%–96.5% · 27 of 30 rows · strict, every arguable call counted wrong: 87%

Visible executive appointments and departures. Measured 20 Oct 2026, seed yr-today-2026-10-20.

Form D filings

Above our bar
90%no material error

95% CI 74.4%–96.5% · 27 of 30 rows · strict, every arguable call counted wrong: 73%

Form D offerings, judged on issuer, amount sold and first sale date against the filing. Measured 20 Oct 2026, seed yr-formd-2026-10-20-after-nonop-split.

Company profiles

Above our bar
83%no material error

95% CI 66.4%–92.7% · 25 of 30 rows · strict, every arguable call counted wrong: 80%

Canonical company profiles attached to a recent event. Measured 20 Oct 2026, seed yr-today-2026-10-20.

Product launches

Above our bar
77%no material error

95% CI 59.1%–88.2% · 23 of 30 rows · strict, every arguable call counted wrong: 67%

Visible launches created in the last 365 days. Measured 20 Oct 2026, seed yr-today-2026-10-20.

Contracts

Above our bar
77%no material error

95% CI 59.1%–88.2% · 23 of 30 rows · strict, every arguable call counted wrong: 77%

Visible contracts created in the last 365 days. Measured 20 Oct 2026, seed yr-lanes-2026-10-20-contracts-after-nonct.

Acquisitions

Below our bar
Measuredfigure held until it clears 80%

Last measured 20 Oct 2026 on a fresh draw of 30 rows. Fixes are in progress; next re-measure 1 Nov 2026.

Visible acquisitions created in the last 365 days.

AI reviewers (Claude) reading each row against its original source, with web checks; seeds committed before judging. With 30 rows a figure can move by about 15 points either way, which is why each card shows its interval.

03 · Who uses it

Built for the people who act on a change

Built in-house

We collect from primary sources: SEC EDGAR, state and federal registries, press wires and the news.

Same-day

Filings and announcements are read as they appear, so a round or a hire can reach you before the press picks it up.

Measured in public

Accuracy is audited on blind random samples and published here, with intervals and limits.

Sales teams

Reach a company in the week it raises, hires a leader or signs a contract.

PE and VC firms

Source deals and watch portfolios from filings, not just headlines.

Recruiters

Follow executive moves and newly funded companies that will hire.

Consultants

Find companies in transition: new money, new leaders, layoffs, leases.

Platforms

License Fundz data inside your own product, with this audit behind it.

04 · How it flows

Six steps, each with its own guard

Each step says what does the work, rules, AI or machine learning, and what stops a bad record at that step.

  1. 1

    Collect

    Rules

    Filings, registries, press wires, news and company job boards are pulled on fixed schedules, each source on its own lane.

    Guard A signal-health check runs every day and asserts fresh output on every advertised lane. A lane that goes quiet raises an alert the same day.

  2. 2

    Extract

    AI

    Language models read each article or filing and return the companies, amounts, round types, deal types and dates in a fixed structure.

    Guard Intake guards hold back what should not ship: stories that are not a funding round or a deal, late re-reports of old news, implausible amounts and foreign-currency slips.

  3. 3

    Match

    Rules

    Every event is attached to a company. Government identifiers come first (SEC CIK, EIN, LEI), then the web domain, then the name with a second signal.

    Guard Each link records how it was made and how confident it is, so a weak link is labelled as weak and can be audited later.

  4. 4

    Check

    Rules

    An automated analyst sweeps new links and the duplicate backlog every day. What it cannot settle on its own goes to a review queue.

    Guard A check may fix data on its own only when it scores at least 97% on a frozen gold set. Anything it removes is hidden with a reason, never deleted, and can be undone.

  5. 5

    Score

    Machine learning

    Machine-learned models estimate how likely each company is to raise, hire, buy or be bought, and fuse each signal with the customer’s own market.

    Guard Every prediction is frozen the day it is made, committed with a cryptographic hash anchored in the Bitcoin blockchain, and graded once its horizon closes.

  6. 6

    Audit

    Rules + AI

    A blind random sample is drawn every month with a seed recorded before any row is read, and every row is judged against its source.

    Guard Results are published here with sample sizes, 95% intervals and limitations. Anyone can redraw the same rows from the seed.

05 · Keeping it accurate

Guards that run every day, not once

Hide, never delete

When a check or a review finds a bad record, it is hidden from customers with a written reason and a backup, so every fix can be undone. 104K rows are held back today.

Hand-checked before any bulk fix

A fix that touches many rows first passes a seeded hand check of 30 at 90% or better, then runs in batches of 200 or fewer.

Every link carries its method

Each event-to-company link stores how it was made and how sure we are. 99.9% of 301,128 links made since the method record went live carry one.

Fresh output on every lane

A daily signal-health check asserts that every advertised lane produced new records inside its expected window, and a site walker uses the product like a customer twice a day.

Identifiers held on the 22,040 companies with an event in the last 12 months

Web domain73%
LinkedIn page56%
SEC CIK43%
IRS EIN19%
LEI4%

Automated duplicate checks and their gate

Auto-apply only at gold precision ≥ 0.97 over at least 30 rows the check fired on. Below the gate, a check can only suggest.

Duplicate checkPrecision on the gold setWhat it may do
Same parser ID, 2+ canonical orgs100.0% (95% CI 96.2%–100.0%; 98 of 98)Fixes on its own
Same normalised name (needs a 2nd signal)75.9% (95% CI 63.0%–85.4%; 41 of 54)Suggests, a reviewer decides
Shared domain (never merged on this alone)46.6% (95% CI 34.3%–59.2%; 27 of 58)Suggests, a reviewer decides
Domain accretion (3+ unrelated roots) and junk hosts100.0% (95% CI 95.2%–100.0%; 76 of 76)Fixes on its own
LinkedIn slug consistent with name/domain (Apollo accounts[0] fallback)26.5% (95% CI 14.6%–43.1%; 9 of 34)Suggests, a reviewer decides
Ingest churn: a new org whose key matches an older canonical69.4% (95% CI 53.1%–82.0%; 25 of 36)Suggests, a reviewer decides
Funding round's own EDGAR link names another org's CIK (review only)100.0% (95% CI 93.5%–100.0%; 55 of 55)Suggests, a reviewer decides
9 more checks still building a gold sample (they can only suggest)
Duplicate checkPrecision on the gold setWhat it may do
Same CIK after trimming zerosnot yet measured: sample n=26 is below the publication minimum n=30Held: under 97% on gold
Shared LinkedIn company slugnot yet measured: sample n=8 is below the publication minimum n=30Suggests, a reviewer decides
Event text names the linked org (role check for acquisitions)not yet measured: sample n=7 is below the publication minimum n=30Suggests, a reviewer decides
8-K filing CIK vs parser company CIK vs org CIKnot yet measured: sample n=1 is below the publication minimum n=30Suggests, a reviewer decides
EDGAR company name for the CIK vs org namenot yet measured: sample n=1 is below the publication minimum n=30Suggests, a reviewer decides
Homepage title / og:site_name vs org name; redirects; parked/scanner hostsnot yet measured: sample n=4 is below the publication minimum n=30Suggests, a reviewer decides
WARN notice name vs org, with state agreementnot yet measured: sample n=3 is below the publication minimum n=30Suggests, a reviewer decides
Investor alias clusters / person-as-investor / new row duplicates an old onenot yet measured: sample n=3 is below the publication minimum n=30Suggests, a reviewer decides
Investor named in the source article (NOT MEASURABLE: only a ~400-char summary is stored)not yet measured: no labelled gold rows fire this checkSuggests, a reviewer decides

06 · AI and machine learning

Where AI does the work, and how we check it

JobWhat the model doesHow it is checked
Reading textLanguage models turn press releases, articles and filing text into structured records.Spot-checked against the source in every audit; a model never sets a government identifier.
ClassifyingModels tag each deal and story (whole company or assets, new deal or a later mention, launch or feature update).Intake guards built on tags were validated on hand-checked samples before going live; a bulk hide built on them must pass a 30-row hand check at 90% or better.
ScoringMachine-learned models estimate the chance a company raises, hires an executive, buys or is bought within a set window.Predictions are frozen daily, hash-committed and graded against what really happened.
AuditingAI raters label the monthly sample; a second model labels the same rows blind.Agreement is published as Cohen’s kappa, disagreements are adjudicated in writing, and the raters are named.

Where we deliberately do not use a model

  • Matching events to companies is rule-based: government identifiers first, then web domains, then a name backed by a second signal.
  • SEC filings are tied to companies by their CIK, the government identifier, not by a model.
  • Amounts and dates in SEC filings are read from the filing itself.
  • A check that changes data on its own must first pass a frozen gold set; a model alone never deletes anything.

Predictions you can hold us to

Each day’s predictions are frozen and committed as a salted Merkle tree whose root is timestamped in the Bitcoin blockchain through OpenTimestamps. Any single prediction can be revealed later with its proof, without exposing any other, so a track record cannot be rewritten after the fact.

Calibration is graded once each horizon has closed. For the current model:

  • Company raises a funding round within 180 days: first gradable 24 Jun 2027
  • Company is acquired within 90 days: first gradable 26 Mar 2027
  • Company acquires a company within 180 days: first gradable 24 Jun 2027
  • Company hires an executive within 90 days: first gradable 26 Mar 2027

07 · Check it yourself

Don’t take our word for it

  1. Every number on this page is version-controlled JSON: api.fundz.net/v1/data_quality.
  2. 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.
  3. 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.
How we measure, in full
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.
Whole-record audits
No material error (lenient, the published gate): nothing that would mislead a buyer on the company, what happened, the amount or when; strict: every row a reviewer marked arguable also counts as wrong.
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.

08 · Limitations

What this page does not prove

  • 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.
  • Whole-record audits are one fresh draw of about 30 rows per lane, judged by AI reviewers against the original source, so each interval is roughly 15 points wide. Each lane is measured on its own scope (shown on its card), and a lane below its bar shows no figure until a fresh draw clears it.
  • 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.

See it on your own market

Your scored leads, from these signals, matched to the companies you sell to.

See plansLicense the data