project grandview focus markets × agents status under development page v0.1 updated 2026-11-18
Grandview

An open working page — not finished, and saying so

The models do
the reading. A
person does the arguing.

Grandview is where markets commentary gets built with AI models and agents on purpose — and where the machine part stays visible instead of hidden. Which agents ran, what they got wrong, where a human overrode them, and what the resulting view is worth. The point is not that a model wrote it. The point is which judgment survives contact with a tape that doesn't care how it was produced.

It is under development: the method below is real and in use, the notes it produces have not been published yet, and the page says which is which.

collects: nothing  ·  forms: none  ·  analytics: none  ·  third-party requests: none

01 — What this is

A working page about markets,
not a place to be sold something.

is

  • Commentary on markets and on the AI tooling now used to study them
  • An honest account of the pipeline: models, agents, prompts, what they fail at
  • An archive in the making, held to one rule: notes that age badly stay visible
  • A reference for how a claim gets checked before it gets published

is not

  • A service, consultancy or firm taking clients
  • Software, a tool or a subscription for sale
  • A mailing list — there is nothing to sign up to
  • Investment advice, research for compensation, or a performance record

In one line: a markets notebook that treats its own tooling as part of the subject. Written and published by Grandview Financial Markets LLC, an independent company based in Puerto Rico — the About page has the rest.

02 — Thesis

Two curves, different shapes.
The interesting thing is the gap.

Writing about AI and markets usually collapses them into one story. They are separate: model capability moves on one schedule, the market's willingness to pay for judgment on another. Where they cross is where the questions are.

What the models already beat you at

  • Reading all of it — filings, transcripts, local-language press, the stack, before the open
  • Cross-document contradiction: the guidance that moved quietly between two quarters
  • First pass from research prose to something testable
  • Boilerplate diffing — the clause that changed
  • Breadth on demand: forty names in a sector, the same hour

What they still lose to a human

  • Knowing what is already priced, therefore worthless as news
  • Sizing something you are only 55 % sure about
  • Telling a regime change from a noisy week
  • Reading positioning, crowding, who has to sell
  • Owning a view when a model's confidence is indistinguishable from its invention

So the working question is not can a model do it. It is how fast does the capability stop being an edge once everyone holds the same model, and which judgment stays scarce.

03 — Method

How a note gets written

Four steps. Two of them are the machine's, and neither of them is the last word.

  1. step 01 · machine

    Sweep

    Agents run the day's material at a scale no person reads: filings, transcripts, press across markets, the back file. Output is a stack of candidates with citations, not a conclusion.

    out: candidate list, sourced

  2. step 02 · machine

    Quarantine

    Model output stays in a holding layer. Nothing it asserted enters a note as a number, a date or a quote until the source behind it has been opened by hand.

    out: claims with a human-opened source

  3. step 03 · human

    Argue

    Where the note is actually made: which of those candidates matters, what is already in the price, what would be evidence against the reading. A person writes this, in their own words, and takes the position.

    out: a view with its falsifier attached

  4. step 04 · both

    Adversarial pass, then publish

    A model is pointed at the finished note to hunt unsourced numbers and sentences the evidence doesn't carry; a person resolves every hit. The note ships with the tooling that made it disclosed.

    out: a published note, tools named

hard rules

  • no_agent_publishes — an agent never ships a note or replies to a reader
  • every_number_has_a_source — opened by a person, dated in the note
  • failures_stay_up — a wrong reading is amended in place, never deleted
  • tooling_is_disclosed — which model, which date, which prompt mattered
  • no_investment_advice — views are views; sizing is yours

Of the four steps, exactly one is where the outcome is decided, and it is not the one with the most compute behind it. Reading, arithmetic and source-chasing are delegated freely. Deciding what matters, and what is already priced, is not.

04 — Tools & agents

Five roles, each with a job
and something it may not do.

Which model fills a role turns over faster than any list of them stays true, so the specific name and version travel with the note they were used for — every published note says what ran and when. What sits here is the permanent part: the division of labour, and where each machine's authority stops.

05 — Notes

Nothing published yet

The index is empty because nothing has cleared the adversarial pass at the end of the method yet. Rather than fill the space with summaries of things that don't exist, the first entry arrives when there is one.

published in progress outlined

Those are the three states a note can be in, and each future entry carries one. The first entry will arrive with its tooling named and its evidence attached — the same standard the rest of this page is written under.

06 — Log

Development log for this page

Kept because a page about verified claims should be able to show its own. Newest first; entries are never edited backwards.

07 — Disclaimer

Commentary, kept to its own standard.

Grandview is commentary on markets and technology, written for its own sake. It is not investment advice, research for compensation, a recommendation, or an offer to buy or sell anything. Views are the writer's own and may be wrong; nothing on this page is a record of performance.