The agent stack

Four stages, zero shortcuts. Each stage feeds the next — no brief ships if any stage fails QA. The pipeline runs hundreds of sources in parallel, then narrows to a single scored document.

01 / Ingestion
Ingestion
Raw source collection across primary documents, regulatory filings, and native-language press. No aggregators — the agent reads the actual filing, the actual transcript.
IN Filings, transcripts, patents, trade data, press
OUT Raw document corpus, source-tagged
02 / Entity Graph
Entity Graph
Named entities — companies, executives, products, competitors — are resolved across 40+ naming conventions and jurisdictions. Every claim is atomized to a specific source document.
IN Raw corpus from Ingestion
OUT Normalized entity graph, discrete claim records
03 / Scoring
Scoring
Every claim is rated 1–5 on source tier, corroboration count, recency, and triangulation. Claims below 3 are excluded or flagged; only rated claims reach the brief.
IN Entity graph + claim records
OUT Confidence-rated claim set (1–5 per claim)
04 / QA
QA
A sector lead reviews every flagship brief: accuracy vs. source documents, rating calibration, thesis coherence. No flagship ships without a human sign-off.
IN Draft brief + scored claim set
OUT Published brief, QA-signed

Source coverage

Eight source categories, continuously ingested. Primary documents only — no aggregators, no summaries, no paid data we can't cite. Every source is tagged and available in the brief's source log.

📄
SEC Filings
10-K · 10-Q · 8-K · DEF 14A · S-1
EDGAR primary text; also HKEX, KRX, TSE, LSE, SEDAR for cross-listed coverage.
🎙️
Earnings Transcripts
Auto-indexed within 2h of publication
Verbatim earnings calls, analyst days, investor conferences, and product keynotes.
🎤
Founder & Exec Interviews
Podcasts · press · conference talks
First-person statements from CEOs and executives — sourced to the specific appearance and timestamp.
🌐
Foreign-Language Press
6 languages · machine-translated & verified
Native-language journalism in Mandarin, Japanese, Korean, German, French. Translated in-pipeline; source language cited.
🛰️
Satellite & Alt-Data
Parking lots · ports · construction
Occupancy tiers, shipping port congestion signals, and construction activity change detection — used for independent triangulation only.
⚗️
Patents
USPTO · EPO · WIPO
Filed and granted patents; assignee and inventor networks; R&D direction signals from application text.
📊
Hiring Signals
Monthly headcount delta
Job description keyword changes, role-level mix shifts as a proxy for R&D vs. sales investment, and headcount trajectory by division.
🚢
Supply Chain Disclosures
Import/export records · customs data
US import/export records and customs datasets for supply chain mapping and demand-signal validation.

Update cadence

Different brief types have different freshness requirements. Every type is published on a defined schedule — no gaps, no guessing when the next update lands.

Brief type Timing What it covers
Pre-print briefs T-7 to T-1 before event Specific, falsifiable predictions before earnings or catalyst. The document graded in our track record.
Post-print recaps Within 24h of print Graded scorecard: what we called, what the print delivered, Correct / Mixed / Incorrect per prediction.
Sector briefs Quarterly refresh Sector-level intelligence: competitive map, multiple expansion/compression signals, cross-company dynamics.
Catalyst calendar Continuous · always live Every upcoming earnings date, investor day, product launch, and regulatory decision tracked and monitored.
Ask previews On-demand · <60s Instant brief snapshot on any tracked entity — vitals, thesis, risks, and next catalyst. Available from Ask Vektor.

Confidence & grading

Every pre-print brief makes falsifiable calls — specific, checkable predictions with a stated direction and a stated basis. After the print, each call is graded against the actual result. No rounding up, no retroactive interpretation.

✓ Correct
The prediction matched the actual result within a reasonable margin. Direction, magnitude, and stated basis all confirmed by the print.
◑ Mixed
The prediction was partially right — correct direction but wrong magnitude, or right on one sub-call but wrong on another within the same thesis.
✗ Incorrect
The prediction was wrong. Direction missed, magnitude was off, or the stated catalyst didn't materialize. We don't delete these.
⏳ Pending
The print hasn't happened yet, or the call is tied to a multi-quarter event still in progress. Grading is deferred until the resolution date.

Confidence scores on forward calls are assigned at time of publication — High / Medium / Low — based on source quality, corroboration depth, and historical accuracy on similar setups. The score is not updated after publication.

Live Scorecard
See every call we've made — graded
Every pre-print brief graded against the actual print. Including the ones we got wrong.

What we won't do

These aren't marketing promises. They're technical and editorial constraints enforced at the pipeline level.

Anonymous sourcing without disclosure If a claim originates from an unnamed source, it is labeled as such and cannot exceed a confidence rating of 2. We don't hide the sourcing gap behind confident language.
Recycled press releases Press releases are Tier 3 — not primary documents. A company saying something in a PR is not the same as disclosing it in a filing. Press releases are ingested as signals, not as evidence.
Unsourced valuations Every valuation figure (ARR, revenue run-rate, enterprise value, AUM) requires a primary citation. "Reportedly valued at X" without a named filing or credible named source is not published.
AI-generated claims passed as fact AI is used for extraction, normalization, and summarization — not for originating claims. Every claim in a Vektor brief can be traced back to a specific source document. If it can't, it doesn't ship.
Political slant Briefs do not take political positions. Regulatory risk is reported factually. Our defense coverage is about capability and contracts, not policy advocacy.
Paid placement No company pays to be covered, scored favorably, or included in a brief. Revenue is from subscribers, not subjects. The companies we cover don't know we're publishing until we do.