Ask the desks anything.

Latest from the desks

Each desk's most recent run, ranked by what only this suite can see: names more than one desk holds, disagreements, and what is new since the last run.

The Launchpad

Six pipelines are normalized into one signal shape, folded into a graph of pipelines against entities, and measured for confluence, friction and novelty. Only then is a model asked to read the findings out, and every figure it writes is checked against the graph before it reaches this page.

Six products, one family

They share a canvas, a monospace numeral, a DuckDB warehouse and a scoring methodology. Only the accent colour differs.

What actually connects them

The claim that this is a suite rather than six unrelated apps is checkable, so it is checked. Twelve cross-product edges exist in code — a shared warehouse, a scoring methodology ported verbatim, a scraper hardened in one repo and carried into another, an artifact one product reads directly out of another's published output. Every edge below carries a file and line number.

Why agreement is not all worth the same

Two desks agreeing because they both read the same warehouse is weak evidence. Two desks agreeing when they share nothing upstream is strong evidence. That intuition is made arithmetic here: pairwise overlap of source sets, computed over informational sources only, then used to discount correlated confirmation everywhere in the Launchpad. The picture and the ranking are the same numbers.

Two roots, not six islands

Strip out data flow and only ancestry remains: code that was forked, and operational lessons that propagated without any code moving at all. Out-degree over those edges identifies the hubs without anyone having to assert them.

The suite covers the clock

Read from the products' own cron configuration, not described from memory. Between them the pipelines publish from before the open to after the wrap, seven days a week — which is what makes a synthesis layer worth having, and what determines when it runs.

What the suite covers, and what it does not

Horizon against domain. The gaps are drawn as gaps.

The rules the suite holds itself to

Numbers before words

Findings are graph measurements computed before any model is called. The model narrates a result it cannot influence, which is what gives the validator a ground truth to check against.

Fail closed on figures

Every percentage and decimal in the narration must already exist in the graph, within 0.15 after rounding. Offenders are regenerated once, then dropped in favour of deterministic prose.

A missing feed is stated

When a pipeline fails to load, the narrator is forbidden from characterizing it. An absent feed is otherwise indistinguishable from a quiet one.

No double counting

A pipeline never relays another's view as its own. BoltFactors republishes equity-screener's output; those sections are excluded, because they would make one desk look like two.