BUILT NOT BOT
← back to the overview

How it runs

The plumbing behind the copilot.

The schedule, the shared files, how a call is built, the books that grade every call, and the checks that keep it from guessing. The parts that make it work when nobody's watching.

01 · THE SYSTEM

One front door, a schedule behind it, and a scorekeeper.

I talk to one agent. Everything else runs on a schedule, reads the same files and the same market data, and every call it makes gets graded.

Eric on Telegram asks, approves real money: by hand only asks answers Chief of staff always on the one agent I talk to reads first Shared files: the team's memory instructions decisions in force daily notes lessons → rules change log status board Data feeds 04:30, before the open insiders, SEC filings, macro, 13F, Congress 42 scheduled jobs about 170 runs a week Research: briefings, hunts Marketing: news scan, trends Chief of staff: brief, theme check Ops: watchdog, health, backups read and write Market data layer rebuilt every weekday every listed stock and about 52 ETFs splits applied, bad prints blanked reads logs prices grade them Every call logged with its style and horizon Shadow books graded by prices, from the next open Stage books each hunt stage vs what it rejected Scoreboard what's proven, and what isn't its line is quoted next to every call Every change a before-and-after review my yes, then a backup install, proof, change log one command to undo ships

Swipe to see the whole diagram →

None of it is a chatbot waiting for a prompt. Each job wakes up, loads its own instruction file and only the files that job needs, does one job, writes its work down and exits; long logs are searched, never loaded whole. The marketing agent reads the news and never grades a company, the chief of staff's 7:40 theme check asks who gets paid, and prices grade every call. The scoreboard decides what the system is allowed to claim.

02 · A WEEKDAY

Every dot is a scheduled job. The green lines are agents reading each other's work.

The agents never message each other. They hand work over through shared files, on a clock.

Data feeds Research Marketing Chief of staff Scorekeeping Ops 4:00 6:00 8:00 10:00 12:00 14:00 five feeds: insiders, SEC filings, the macro calendar, 13F, Congress briefing · open check · momentum close review news scan content plan trend check always on 7:30 brief to me · 7:40 theme check midday pulse after the close: prices, the data layer and lenses, reflections, shadow books, stage books, scoreboard login keep-alive health check watchdog, every hour reads the briefing reads the scan, asks who gets paid quoted next to every call

Swipe to see the whole day →

Weekday times, Pacific, from the server's schedule. Evenings add a journal and two backups. Sundays add the screens and the hunt.

03 · HOW A CALL IS BUILT

A name that survives the hunt gets seven steps.

Rewritten on Oct 10 because of a measurement: the evaluations were too cautious, and the names they turned down kept beating the ones they kept. Now every call is built from separate claims that can each be checked, and the confidence comes from how many are proven.

  1. 0The facts first: the live price, then the settled history, with every warning shown
  2. 1Why it moved: the market, its sector, or the company itself
  3. 2The thesis in legs, each with dated evidence and what would break it
  4. 3What's already in the price
  5. 4The base rate: how stocks in the same spot have done
  6. 5What else moves it: rates, oil, the dollar. They set the size, not the conviction
  7. 6The call, in odds rather than adjectives, with the scoreboard's line for its method
  8. 7Logged, so prices can grade it

04 · KEEPING SCORE

Every call gets graded. A part that loses, loses its say.

An agent's confidence isn't evidence. So every call, mine included, is logged and graded by prices, and the system can only claim what the grades support.

  1. Loggedevery call, with its style and how long it should take to work
  2. Shadow booksentry at the next open, graded against its sector and the list it came from. Prices only, no orders
  3. Stage bookseach stage of the hunt, graded against what it rejected
  4. Scoreboardwhat's proven and what isn't, rebuilt every weekday

Plain rules are the baseline.

Five lenses pick their lists by a written rule, with no AI judgment, and each one is graded in its own book. The AI's version of a lens has to beat the plain rule to keep its say.

  • deep drawdown
  • leaders
  • themes
  • after the news
  • defense

Every pick carries a card: the rule that fired, the facts, how stocks in the same spot have done, and a stop and a target sized from the stock's own swings.

Oct 11 · it demoted part of itself

8 of 8

hunts where the names the quick evaluation turned down beat the names it kept, over the next 20 sessions.

So it lost its veto. Its verdict is advice now, and every name it sees goes on to the critic.

The scoreboard's head line, Oct 10Measuring is proven; picking is not yet.

WORKINGthe data layer: nightly checks on every listed stock
NOT WORKINGthe quick evaluation as a gate
EARLY EVIDENCEthe deep-drawdown tilt, in one kind of market
NOT YET MEASUREDthe five plain lenses: too new to grade

With every call, the chief of staff quotes the scoreboard's line for the method behind it, next to its conviction, so it never sounds equally sure about everything.

05 · THE CHECKS

When the evidence isn't there, it says so.

A research agent is only useful if you can trust where its answers come from. These are the checks that keep it from guessing.

RECEIPTS

Every number has a receipt.

Each figure is stamped with its source and the time it was fetched, so any claim traces back to where it came from.

GAPS

An empty lookup isn't a fact.

When a filing lookup comes back empty, it has to say whether the company didn't disclose it or the lookup failed. That rule came from a real miss.

CRITIC

A critic tries to kill every finalist.

Before the full workup, a separate pass argues against the idea. Its only job is to find the flaw.

REASONS

Every no gets graded too.

A rejected idea is logged with the reason it failed, and the stage books check later whether turning it down was right.

BROKEN RUNS

A broken run stops itself.

Checks between stages halt the hunt if a stage loses or garbles its inputs, before anything reaches a verdict. It has done this on a real Sunday.

EVEN THE NO

It doesn't trust its own no either.

A run that rejects everything raises a flag, because a broken pipeline says no too.

ORDERSOrders go through one guarded script that only knows the paper account. Real money moves only by my hand.

Research and education. Not investment advice. The grades measure methods, not returns, and nothing here claims returns.

WANT ONE

Want an agent like this? Tell me what it should do.

The overview has the short version and the form. The hunt has its own page.

Tell me what it should do →How it hunts →