CohortWatch

The startup momentum tracker that shows its work.

momentum-v2 · data as of 2026-08-09Explore the data Methodology Source

Current published read

No company in the current-startup read clears the evidence bar for a gaining read.

CohortWatch tracks 5,746 current accelerator startups through public hiring and code signals. 921 more have graduated, meaning acquired or public, and have left this read; they stay searchable, keep their own pages, and remain in the published file of 6,667 rows. Capital is reserved in the scoring design and contributes no observations in this export. The current-startup read is 5,561 uncertain, 177 holding, 8 cooling, and zero gaining. Every public history stands at length 0, so the instrument publishes level readings, never trends, and says so in its largest type.

That is not a gap in the product. That is the product.

Deterministic score · signal record visible · free & open source

Published 2026-08-09 · 5,746 current startups

Uncertainnot enough public signal yet5,561

Holdinglevel reading: holding steady177

Gainingevidence bar not cleared in the current-startup read0

Coolinglevel reading: cooling8

Of the 5,328 scored, 5,143 are low-confidence; 418 more are watched, not yet scored.

Who this helps

One public record, several practical uses.

The same dated record can support different questions. CohortWatch keeps those uses separate from prediction, valuation, and employment advice.

Founders and operators

See how the public evidence reads, inspect available sources, and file a correction when the record is wrong.

Early employees

Compare observable activity. CohortWatch does not measure culture, stability, compensation, mission, or job quality.

Accelerator teams

Review a cohort consistently and find where the public record warrants a closer look.

Researchers and investors

Reproduce peer-relative context. The result is neither a prediction nor a recommendation.

Run the check yourself

Watch a score earn its way

Every momentum read passes the same four steps: the claim, the signal record, the cross-checks, and the published read. Pick a company and watch its check run. These are the only 177 the instrument will even call steady, and even their momentum claims stay unsupported.

Motion off. Showing Mindee’s completed check. Choose any company to run another.

Holding in this export · 177 companies · not a leaderboard

List order as exported, never sorted by score. “Holding steady” is the instrument’s most modest read, not a win. Scores are level readings.

Browse all tracked companies →

1 · The claim

“Mindee gains momentum”

Under check: not published

History length on file: 0. The record cannot carry a trend claim for anyone yet. The instrument sets the claim aside and reads the available signals.

2 · Signal readings

Hiringweight 40%

Codeweight 30%

Capitalweight 30%

Capital lane: Designed: collectors not yet emitting data. This export contains no capital observations.

includes code signals as of Jun 17, 2026 · code lane: Frozen: collection paused; last collected Jun 17, 2026. Historical data through that date remains in the record.

The bars show exported signal readings. Open the signal record and available source links.

3 · Cross-checks

Cross-checks agree: all 2 checks run agree

The instrument cross-checks every signal family it can. When a check cannot run, it says so. A missing check is never scored as a no.

4 · The published read

Momentum claim: not supported, insufficient history

What we publish instead

Holding steady against its YC AI & ML Infrastructure peers.

Score 58, a level readingMedium confidencevs 294 peers · YC · AI & ML Infrastructure · older vintages pooled

YC · W21 · data as of 2026-08-09 · full analysis →

Where the evidence comes from

Three signal families. One standard.

A momentum score is only as honest as its signal record. The design recognizes three public signal families. This export carries hiring and code observations, never a founder’s name, and no capital observations.

Hiring40%

Hiring signals

Signal present for 5,318 of 5,746 tracked companies. Direct source links appear where the public export carries them.

Lane: Checked Aug 9, 2026

Code30%

Code signals

Signal present for 500 of 5,746 tracked companies. Direct source links appear where the public export carries them. Frozen: collection paused; last collected Jun 17, 2026. Historical data through that date remains in the record.

Lane: frozen since Jun 17, 2026

Capital30%

Capital signals

Reserved as a one-sided family in the scoring design. This export contains 0 capital observations, so capital does not affect any displayed score.

Lane: designed: collectors not yet emitting data

The scorer starts with accelerator, sector, and vintage. When that pool is too thin, a documented fallback ladder broadens it. Cross-accelerator comparison statistics are withheld from the public display. The method is deterministic and the instrument covers companies, not people.

The corridor

Every reading, one narrow band

Drag the reading head across the scale. The instrument prints the exact company count and confidence mix at every score, straight from the export.

The corridor · drag to read

level readings · 0–100 · 5,328 scored · not a leaderboard

1,027 companies read 50

1,016 low · 11 medium confidence

Every scored company sits between 43 and 66; the median reads 50. That narrowness is the evidence being honest, not the instrument being timid. The ends stay unlabelled because 0 and 100 have never been read. Bar heights sit on a square-root scale so single-company readings stay visible; the exact count always prints above.

Confidence grades

How a number earns its confidence

Thin evidence pulls every estimate toward 50. Low confidence never deletes a number and never turns it red. It washes the number out and sets the label beside it, on the page, where you can see it.

5,143Low50: this is how a low-confidence number displays: washed out, labelled, published anyway.

185MediumThe firmest grade in this export. No fanfare, just full-strength type and a label.

0HighThe instrument has issued no high-confidence grade. It will not borrow one.

Reserved for momentum
Teal appears only on a confirmed GAINING read. In the current-startup read: 0 gaining, so the colour goes unspent here. An instrument that keeps its brightest colour unspent is telling the truth. A graduated company can still hold a gaining read on its own page.

Cross-checks

Every check, on the record

Before anything publishes, the instrument cross-checks its own signals. A total of 4,211 individual checks on the record. The tallies below count companies by their outcome (a company can carry several checks; 1,779 carry none).

3,908SupportedCross-checks agree: the available records corroborate. 3,790 of these are unanimous.

59CappedSignals are not fully consistent. Consistency is not claimed, and the score is capped rather than argued.

1,779AbstentionsNothing to cross-validate yet. An abstention is never a score. The record says “no publishable checks” and moves on.

How a check is phrased, verbatim

Supported: “Cross-checks agree: {N} checks run”
Capped: “Signals not fully consistent: consistency not claimed”
Abstention: “Nothing to cross-validate yet: no publishable checks for this company”

The standing commitments

What we never do

Four rules keep the corrections rare. They sit here in full view, not in a footnote.

Never claim a momentum leader.

No "fastest-growing," no leaderboard. When the current-startup read has nothing that clears the evidence bar, we say so plainly.

Never invent a trend.

When there isn't enough history, we keep the number honest: an estimate pulled toward 50, labelled low-confidence.

Never hide a weak number.

Low confidence washes the number out with a label. It is never deleted, and it never turns red.

Never track people.

No founder names, no personal data, no opaque prediction model, and no client-side analytics.

The dataExplore the data/search →The methodRead the methodology/methodology →The sourceInspect the sourcegitlab.com/krahul02004/cohortwatch →

CohortWatch: startup momentum, tracked in the open by an instrument that shows its work and declines to claim what the evidence cannot carry.

Data as of 2026-08-09. This page reads a dated export, not a live feed. Model momentum-v2.