How the pipeline works
Every name in this workspace is worked through the same ten stages, in the same order, by a routine that runs unattended in the cloud and writes what it found into files this site renders. This page is what the stages mean, what closes each one, and which ones stop at your desk.
The ten stages
Defined by the project fork of the pipeline skillThe last column counts the 14 names that currently have a scorecard. Click a stage row’s name on any profile to see what the run actually read.
| Stage | What it asks | What closes it | Who | Closed |
|---|---|---|---|---|
| S0 Universe & fit | Does this company keep the surplus that cheap intelligence creates, or does it hand the saving to its customers? Four structural tests, scored with named evidence — with different tests for Japanese deep-value names, for anything that carries insurance risk, and for concession and regulated-tariff assets, because one instrument cannot ask all four sets of questions. | Nothing. Since 2026-09-22 this is a lens, not a gate: it is scored, written down and carried past. The score never stops a name. | routine | 1 of 14 |
| S1 Source tagged | Where did the idea come from — the owner's watchlist, a screen, a filing alert from the daily EDGAR sweep? A tip with no work behind it raises the evidence bar for every later stage. | The source is recorded. | routine | 14 of 14 |
| S2 Kill test | The thirty-minute question: revenue trend, margin trajectory, the balance sheet, the five-year share count, free cash flow. Serial dilution is the trap this stage exists to catch. | A paragraph saying why the company survives — or a failure with the killing fact named. This is one of the two stages that really gate. | routine | 14 of 14 |
| S3 Filings deep dive | The 10-K, the latest 10-Q, the proxy, the transcripts, and a red-flag sweep: customer concentration, related-party transactions, auditor changes, going-concern language. | At gate-minimum depth for this project (owner decision 2026-09-21): risk factors and MD&A read, not the whole filing end to end. The scorecard says which documents were actually opened. | routine | 14 of 14 |
| S4 Industry & supply map | One value chain for the whole cluster: supplier → company → customer, with the pricing power located on it and customer concentration taken from each issuer's own filing. | The chain written out in words with pricing power located. | routine, once per cluster | 1 of 14 |
| S5 Ownership check | Who is on the other side: 13F holders and what they did this quarter, 13D/G filers, insider Form 4s and Form 144 proposed sales, short interest, and any shelf or ATM hanging over the stock. | You know who owns it and what could be issued. | routine | 14 of 14 |
| S6 Scuttlebutt | One piece of evidence that did not come from the company: employee review trends, customer and competitor commentary, trade forums — and the board's own claim rows, which are free scuttlebutt with the quote already located in a transcript. | One non-company source per name. The product test and the expert call stay with the owner. | routine, then owner | 9 of 14 |
| S7 Written thesis + test | The variant perception in a paragraph, the bear case first, and kill criteria specific enough to be checked against a future filing. | A thesis in writing with testable kill criteria. Conviction is the owner's. | routine, then owner | 0 of 14 |
| S8 Valuation & sizing | What the name is worth on the routine's own model, and what would have to be true for the traded price to be right. | A buy price and a maximum position size, written down before any purchase. Both are yours — the routine models and stops. | owner | 0 of 14 |
| S9 Watchlist & monitoring | What would change the answer, and what watches for it: the trigger, the next filing date, and whether the daily EDGAR sweep covers the feed. | A monitoring trigger exists and the sweep covers the name where a CIK exists. | routine | 14 of 14 |
Two kinds of run
Cluster pass
Carries S0, S4, S6 and S7 across a whole group of names at once, because those four are properties of an industry rather than of one company. It writes one dated file per pass; the profiles of its members point at it rather than repeating it.
A member that turns out not to share the group’s business model gets re-tagged before it is scored — an instrument built for an insurer tells you nothing about an airport.
Name pass
Carries S1, S2, S3, S5, S8 and S9 for up to three names, and carries the cluster’s four forward with a pointer. It writes a dated dossier holding the evidence, updates the living scorecard, and extends the connection map.
One slot per session. Timings come from the git log rather than from the run’s own estimate, which was wrong by hours before the rule changed.
What a slot is
The work queue is a list of slots. A run claims the first pending one with a commit — that commit is both the lock and the clock — works it, and marks it done. The run cadence on the console is read straight out of those commits.
How to read a stage state
Every profile shows the same ten cells in the same order, so the shape of a name is legible before you read a word of it.
| Cell | State | What it means |
|---|---|---|
| ✓ | Closed | The stage was worked and its gate met. The detail column on a profile says with what evidence, and at what depth. |
| · | Carried forward | A cluster pass owns this stage for this name — S4, S6 and S7 are properties of an industry, not of one company, so they are worked once for the whole group. The profile links to the pass that holds the verdict. |
| ◴ | Open with the owner | The routine has taken it as far as a model honestly can. Conviction, the buy price and the position size are yours; a model that supplied them would be pretending. |
| ✕ | Failed | A gate was tested and not met, with the killing fact named. Gates are non-blocking: the run records the failure and carries on rather than abandoning the name. |
| · | Not started | Nobody has attempted it. For a cluster stage this usually means the cluster pass has not run yet — which is different from a pass that ran and left it open. |
S0 is the exception. Since 2026-09-22 it is a lens rather than a gate: its state records whether the lens was applied to a name, not whether the name passed, and the score sits in the detail. It was demoted because it had never actually stopped anything — and because the test it uses failed the category its own reference list calls a perfect score.
Where the evidence comes from
Four rules do most of the work of keeping this honest. They are worth knowing before you trust a number on a profile.
- Nothing here is human-verified. Every scorecard, dossier and connection map was drafted by a language model, unattended since 2026-09-21. It is process support, not investment advice, and no person has checked it.
- A quote is located, never typed. The run finds the span by exact substring search in the filing or transcript and copies it out. A quote that cannot be found is refused rather than written — which has already caught one error.
- Every edge on a connection map cites a document and a locator. Not "the 10-K" but which 10-K, and where in it. The Sources block on each profile lists them.
- Anything from background knowledge is marked as such in the text, as is anything taken from a web-search summary rather than a primary document.
A dated dossier is never edited after the day it was written. When a later run finds something wrong, it writes a correction in its own dated file and the scorecard log records it — so a profile shows what was believed when, rather than a tidied history.
How a name gets here
Two routes, and they meet in the same queue.
The watchlist
122 names the owner confirmed, worked in scheduled slots by cluster. This is the round-robin that keeps the list advancing.
The filing sweep
A daily pass over EDGAR that runs no model and costs nothing. A new 8-K, a cluster of Form 4s or a 13D promotes a name into the queue, where it takes a seat in the next name pass. The sweep is on the console.