01
Ingest filings
SEC EDGAR · 10-K + 2× 10-Q + 4× 8-K
812 ms
The standard behind the research hub
How we keep equity research repeatable.
The same workflow runs whether you press the button or Taufolio runs it for you on a schedule: source collection, structured analysis, model cross-checks, synthesis, and source citations. The machinery is proof of quality, not the product promise itself - the promise is an organized research base around the companies you track.
pipeline · NVDA · deep-researchLIVE
- ✓ 01812 msIngest filingsSEC EDGAR · 10-K + 2× 10-Q + 4× 8-K
- ✓ 021.4 sExtract transcriptsLast 4 earnings calls · Azure Doc Intel.
- ✓ 03410 msFetch market dataYahoo → Finnhub failover · 90d OHLC
- 04running…Draft ×3 in parallel3 frontier agents draft independent viewsClaude1,842 tokGemini ⚑2,108 tokGPT1,604 tok
- 05—Reconcile3 drafts → agreements, flagged disagreements
- 06—Score9 dimensions · −5 to +5 scale
- 07—CitePer-claim anchors to source pages
stage 4 of 7 · 3.2s elapsedeta 36s
The pipeline · 7 deterministic stages
Seven stages, one report.
Reports follow the same route: ingest filings, extract transcripts, fetch market data, draft independent research views in parallel, reconcile them, score the result, and attach citations. That consistency is what makes reports comparable across companies and useful over time.
02
Extract transcripts
Last 4 earnings calls · Azure Doc Intel.
1.4 s
03
Fetch market data
Yahoo → Finnhub failover · 90d OHLC
410 ms
04
Draft ×3 in parallel
3 frontier agents draft independent views
4.1 s
05
Reconcile
3 drafts → agreements, flagged disagreements
2.1 s
06
Score
9 dimensions · −5 to +5 scale
180 ms
07
Cite
Per-claim anchors to source pages
240 ms
Specialised drafting roles
Seven research agents. Seven angles on the same evidence.
Each agent looks at the same source set through a different lens - business model, moat, management, tone, analyst concerns, risks, and a final evidence check. The output of every agent is a structured fragment that the reconcile step stitches into one report.
AGENT 01
Business model
How the company makes money, who its customers are, and what drives growth.
→ revenue.segments[]
AGENT 02
Moat & competition
What protects the business, where the advantage may be strong or weakening.
→ moat.signals[]
AGENT 03
Management
How leadership communicates, allocates capital, and explains strategy.
→ management.tone
AGENT 04
Earnings-call tone
Tone, key themes, analyst questions, and possible red flags from the call.
→ call.signals[]
AGENT 05
Analyst concerns
Recurring questions in the Q&A - what the buy-side keeps pressing on.
→ concerns[]
AGENT 06
Key risks
Risks invisible in the numbers: regulation, cyclicality, customer concentration.
→ risks[]
AGENT 07
Evidence check
Final pass that confirms each material claim resolves to a primary source.
→ citations[]
Model Council
Three frontier models. One reconciled report.
Three providers run in parallel on the same primary sources. A reconcile pass merges what they agree on and flags where they diverge. No chat. No clarifying questions. Disagreement is surfaced, not hidden - and we tell you which models disagreed about what.
Why three, not one
One model can be confidently wrong. Three rarely are at the same time.
Single-model outputs lack the cross-check that an analyst team produces naturally - someone reads the 10-K, someone listens to the call, someone challenges the view. The Model Council approximates that team: each frontier model drafts independently, then we reconcile.
Where two or three agree on a material claim, the reconciled report inherits the claim with a confidence pin. Where they disagree, we surface the disagreement instead of picking a winner.
The reconcile pass is deterministic. It's not a fourth model voting - it's a rule-based merge that aligns claims by source anchor and flags structural disagreements.
Agent α · Claude
draft #412
"Services acceleration is structural; expect 18–22% YoY on the line item."
Agent β · Gemini
draft #187
"Services growth may moderate as tough FX comps return in Q1."
Agent γ · GPT
draft #603
"Services growth trajectory remains a key driver; analyst consensus at 17%."
reconcile
Reconciled✓ 18 agreed · ⚑ 3 flagged
"Services growth is the key driver, running ~18%12, though ⚑ models disagree on the Q1 impact of FX comps - one drafter expects moderation; two expect structural continuation13."
The four meanings of τ
What Tau means, in plain English.
Tau isn't an acronym. It's a symbol borrowed from maths and engineering that maps cleanly to the four things a company research folio should be: marked by evidence, checked from multiple angles, fast enough to matter, and complete enough to be useful.
TAU I — Evidence
An insight should leave a mark back to its source.
Material findings tie back to filings, annual reports, transcripts, and management commentary. The citation chip is the mark.
Symbol: marks & signs
TAU II — Agreement
No single model is the final word.
Multiple frontier models analyse the same evidence independently. A reconcile step compares findings and flags disagreement instead of hiding it.
Symbol: Kendall's τ
TAU III — Speed
Research at the tempo of new information.
Monthly recaps keep you current. Long reports help you understand the business. Tau is also the engineer's symbol for system response time.
Symbol: time constant
TAU IV — Full view
One full turn around the company.
Business model, moat, management, earnings tone, analyst concerns, risks - and the evidence behind each. τ = 2π means the complete circle.
Symbol: 2π / one turn
How citations are constructed
Material claims carry numbered pins.
Material factual claims resolve to a specific source anchor: a section of a 10-K or 10-Q, a passage from an earnings-call transcript, an annual-report page, or a named external article. Background framing - language about narrative or context - is deliberately uncited so you can tell fact from interpretation.
Read the chip
Click any pin to jump to the source page.
Each N chip resolves to a deep link inside the original document. Open the 10-K page or transcript line that backs the claim - without leaving the report.
Lines without a chip are framing language - connective tissue between cited facts. We don't pretend that "this looks like a strong quarter" is a citable claim; we mark interpretation as interpretation.
Cache fan-out is invisible. If your ticker was researched recently, you may get a cached report - but the citations always resolve to the source as it existed at the time of generation. Citation freshness is timestamped in the report header.
NVDA · long report (excerpt)§ 2 · MOAT
CUDA's deep integration with the developer ecosystem creates a switching-cost moat that competitors haven't matched in over a decade04. Management framed this directly on the most recent call, with CFO Colette Kress noting that "the installed base of CUDA developers is approaching four million"05.
The implication for investors is that displacing NVIDIA on accelerated computing means displacing CUDA - and that is a much higher bar than displacing GPUs alone06.
04NVIDIA 10-K FY2025 — Item 1, "Software"p. 14 · §1.3
05Q1 FY26 Earnings Call TranscriptL. 412 · CFO
06Bernstein Research Note — May 2026external
What this is - and isn't
The honest caveats.
Every reasoning step you'd want to inspect, we expose. Every limitation we know about, we name. Both belong on a methodology page.
It's a research base, not a recommendation engine.
Reports are starting points for your own analysis - not buy/sell signals. Every report carries a research stance, not a price target. The investment decision stays with you.
Models can still misread sources.
Reconcile catches most disagreements but not all subtle mistakes. If a claim looks important to your view, click the pin and read the underlying source yourself.
Late-breaking filings may not be picked up.
Each report has a cut-off timestamp in the header. Filings or transcripts published after that time aren't in the source set - we won't pretend otherwise.
Coverage varies by company.
Public companies with deep filing histories produce better reports than recent IPOs. We tell you the source count and freshness on the cover sheet so you can calibrate confidence.
Cache is invisible by design.
If your ticker was just researched, you may get a cached copy. Generation time differs; the report content doesn't. Force a fresh run from the report header if you need it.
Read what comes out the other side
Now read a real report. It's the only honest way to evaluate the workflow.
Open a Taufolio long report on NVIDIA - every citation pin live, every reconcile flag visible, every agent's contribution traced. Then run your own.