Verifiable company intelligence

Company briefs with the receipts attached.

Type a company name. Kompete researches it across public sources, writes the competitive brief and the investment layer, and binds every number to the page it came from. Where sources disagree, it says so. Where research came up empty, it tells you what to ask.

04 Research agents
×2 Passes per agent
Claim states
00 Documents uploaded
Brief · Northwind Robotics
ILLUSTRATIVE · FICTIONAL COMPANY · v3 · 4 AGENTS · 27 SOURCES
accuracy 94%
Sources disagree
  • Company press release$600M post-money
  • Secondary market report$540M, citing investor filings

Detected arithmetically before synthesis. The brief shows the range and who claims what.

Hover a figure to see what stands behind it

§ 01 · Product

The report is the easy part. Trusting it is the product.

Anyone can prompt a model into writing a confident memo. Kompete is built for the moment after that, when a partner asks where did that number come from? and the answer has to be a link, not a shrug.

A Kompete brief is a competitive-intelligence report and an investor layer, generated from sources the pipeline actually retrieved. Every fact is bound to one of those sources. Disagreements between sources are shown as ranges with attribution instead of being quietly resolved. Fields research failed to establish are listed explicitly, each with the question that would close it.

A change that makes a claim look better-supported than it is would defeat the product, so those guarantees are enforced in code, not written into a prompt and hoped for.

Sourced

Bound to a retrieved URL through a chain the model cannot fake.

Contested

Two sources, two values, both shown with who said what.

Not established

Stated as a gap, carrying the question that closes it.

Inference

Model reasoning, labelled so it is never mistaken for a fact.

§ 02 · How it works

One name in. Five phases. Nothing invented on the way.

The pipeline is arranged so that the model is never the source of truth for a number, a URL, or whether two sources agree.

  1. 00

    Retrieve

    The company is classified as public, startup or obscure, which gates which scrapers run. News, community discussion, encyclopedic pages and the company site are collected for everyone. Public companies also get fundamentals, a multi-quarter series and consensus beat-or-miss, attached deterministically so no number passes through a model.

    Playwright · yfinance
  2. 01

    Research, twice

    Four agents cover news, financials, reviews and social signal. Each runs two passes: a grounded prose pass with search on, which yields grounding metadata, then a search-free extraction pass that turns the prose into structured facts citing sentence indices. Grounding only comes back for prose, so this split is what makes provenance possible.

    4 agents × 2 passes
  3. 02

    Check before writing

    Contradictions are found arithmetically across the fact ledger, guarded against false positives: same canonical metric only, never two explicitly different periods, rounding-insensitive, with a per-unit noise floor. Gaps are the set difference between what the brief template promises and what the ledger resolved.

    contradictions · gaps
  4. 03

    Synthesise

    The SWOT brief is written with the conflicts already in hand and an instruction to disclose them rather than pick a winner. The investor layer follows: thesis, unit economics, market sizing, competitive map, team and ESG. A second-order napkin thesis runs last.

    brief · investor layer · napkin
  5. 04

    Resolve every citation

    The model only ever cites an integer or a tag. Each citation is resolved in code from fact to sentence to grounded span to URL. Anything that does not resolve is dropped, never guessed. The result streams to the browser, exports to PDF and is versioned so the next run can be diffed against it.

    fact → sentence → span → URL
§ 03 · Trust layer

Four guarantees. Each enforced in code.

T·01
Sourced

Claim-level provenance

Each extracted fact is bound to a source the pipeline retrieved. The model cites a sentence index; code maps it to the grounded span and then to the URL. A citation that does not resolve is dropped rather than guessed.

fact → sentence → span → URL
T·02
Contested

Contested figures

When two sources put different values on the same metric, the brief shows the range and attributes each value. Detection is arithmetic and runs before synthesis, so the writing step is told to disclose the conflict, not resolve it.

same metric · same period · noise floor
T·03
Not established

Could not establish

Every brief section declares what it promises. What the ledger failed to resolve is listed as a gap, split into missing versus unsourced, with critical fields flagged and the question that would close each one.

expectations − ledger = gaps
T·04
Diffed

Changes since last run

Re-running a brief diffs the fact ledger, not the prose. Synthesis rewrites every sentence, so a text diff is noise; a ledger diff shows exactly what moved, what was added, and which figures lost their sourcing.

content-hashed versions
Published accuracy

One metric. Two graders. And the agreement between them.

Accuracy is the share of extracted facts whose cited source actually states that value. It is scored by a human golden set and by an automated judge, and the agreement rate between the two is published alongside. The judge number is never reported on its own.

§ 04 · The brief

Competitive intelligence on top. Investor layer beneath. Trust throughout.

Core brief6 sections
  • 01
    Executive summaryThe brief in one screen, with each claim labelled by how it is supported.
  • 02
    SWOT gridStrengths, weaknesses, opportunities and threats, disclosed conflicts included.
  • 03
    Financial snapshotRevenue, valuation, growth and funding history, each figure sourced or flagged.
  • 04
    Quarterly performanceBeats, misses, EPS surprise and YoY/QoQ moves, attached straight from market data.
  • 05
    Market sentimentA sourced score with the insights, strategic moves and threats behind it.
  • 06
    SourcesEvery URL the pipeline retrieved and cited, ready to open and check.
Investor layer7 sections
  • 01
    Investment thesisBull, base and bear cases, VC-fit score, stage read and likely acquirers.
  • 02
    Unit economicsLTV, CAC, gross margin, payback, burn, runway, ARR and NRR, confidence-scored.
  • 03
    Market sizingTAM, SAM and SOM with the reasoning, plus CAGR and lifecycle stage.
  • 04
    Competitive mapDirect and indirect rivals with threat levels and a moat breakdown.
  • 05
    Team & ESG scorecardsFounder quality, key-person risk, hiring velocity, and E/S/G with controversies.
  • 06
    Napkin thesisThe second-order read: what has to be true for this to be a great outcome.
  • 07
    Peer benchmarksCohort median, quartile and percentile against sector and stage peers.
Trust layer4 sections
  • 01
    Provenance summaryCounts of sourced, contested, unestablished and inferred claims across the brief.
  • 02
    Contested figuresEach disagreement as a range, with who claims what.
  • 03
    Could not establishThe gaps, critical ones flagged, each with its closing question.
  • 04
    Changes since last runWhat moved, what was added, and what lost its sourcing.
§ 05 · Sample briefs

Two briefs, as generated. Read the receipts yourself.

One public company with audited filings, one private company with none. Both briefs are served exactly as the pipeline produced them: every figure tagged, every disagreement arbitrated, every gap listed with the question that would close it.

01

Alphabet

Public company

Record GAAP earnings that are mostly a mark-to-market gain, a capital model that inverted, and a six-year court judgment on the search index.

Revenue up 24.2% year over year with Google Cloud growing 81.8% against a market growing 43%. Q2 net income of $112.2bn includes a $99.0bn securities gain, buybacks went to zero, and $105.8bn was raised in six months to fund capex that cut free cash flow to $4.3bn.

  • Five contested figures arbitrated, including GAAP versus ex-gain EPS
  • Eight gaps listed with the question that closes each one
  • All five critical fields filled from audited filings
103Facts
16Sources
7Blocked
77.7%Sourced
Generated 8 September 2026
02

Notion

Private company

Category leader at $865M ARR and an $11B secondary mark, profitable since 2018, with growth the one load-bearing number nobody can settle.

Over half of ARR now comes from AI-enabled customers. The same research house publishes both 82% and 53% year-over-year growth for the identical $865M figure, and no audited statement exists to resolve it. Unit economics are left blank rather than invented.

  • Seven contested figures, three critical gaps flagged
  • Eight of 39 sources refused retrieval and are recorded, not circumvented
  • LTV, CAC and gross margin marked unavailable instead of estimated
86Facts
39Sources
8Blocked
83.7%Sourced
Generated 5 September 2026
§ 06 · For investors

From a name in a feed to a diligence-ready file.

I·01
Investor

Discovery feed

Typed, de-duplicated events ingested from public feeds and pages: launches, raises, hires, filings. Each event keeps the source it came from.

RSS · pages · robots.txt honoured
I·02
Investor

Evidence-first profiles

Companies are profiled into a taxonomy from collected snapshots, with each profile fact tied to the snapshot that supports it. Profiling stays useful without a model and never invents a fact.

content-hashed snapshots
I·03
Investor

Explainable comparables

Comparable matching explains why two companies match: which taxonomy facts overlap, which differ, and what the evidence for each was.

overlap · difference · evidence
I·04
Investor

Peer benchmarks

Metric strings are normalised to numbers and placed against a cohort of sector-and-stage peers as median, quartile and percentile.

sector + stage cohorts
I·05
Investor

Coverage and watchlists

Keep a set of companies under coverage, re-run briefs on a cadence, and read what changed on the ledger instead of re-reading the prose.

monthly coverage plan
I·06
Investor

Startup side

Founders can point Kompete at their own public materials so the brief is built from what they actually published, collected with the same rules as everything else.

same rules for everyone
§ 07 · Data posture

A reviewer will test the claim. So it is enforced, not asserted.

Kompete reads what is public and permitted, and nothing else. Two flags govern the posture and both default to off, so the shipped behaviour is the strict one. All ingestion of user-submitted URLs passes the same checks; there is no code path that fetches around them.

ALLOW_RESTRICTED_SCRAPERSfalse
ALLOW_DOCUMENT_UPLOADfalse
posture · public sources only · uploads off
  • Public, permitted sources onlySources whose terms forbid automated collection are off by default and stay off in the shipped posture.
    ON
  • Nothing uploadedDocument upload is disabled. Kompete works from what a company and the public web already publish.
    OFF
  • robots.txt honouredEvery fetch of a submitted URL checks robots.txt and identifies itself with a named crawler user agent.
    ON
  • No private networksSubmitted URLs pass a public-address check before any request is made.
    ON
  • Live, not assertedThe running posture is reported by the system itself, so a reviewer can check it rather than take the copy at its word.
    LIVE
§ 08 · Book a demo

See a brief built live, on a company you choose.

Thirty minutes. We run Kompete on a name you pick, walk the provenance chain on a few figures, open the contested ones, and read the gaps together. You keep the brief.

  • Priced per brief, with a monthly coverage plan for portfolios.
  • Works with your own model key if you prefer to hold the spend.
  • No documents to upload, nothing to install.
Demo requestmritunjaypandey0789@gmail.com

No list, no sequence, one reply.