diff --git a/DESIGN.md b/DESIGN.md
index ab6e4f66..b4cb1c89 100644
--- a/DESIGN.md
+++ b/DESIGN.md
@@ -174,7 +174,12 @@ Tree nodes are things you learn, never things that happen to you. What the
 world does back belongs in events and schemes. When your fronts have earned
 $2,000 in all, the news calls your biggest front a unicorn. Sellers take your
 calls, so you can buy one site before model v2 (you still pay its price). And
-everyone looks at that front: its attention jumps and fades over some minutes.
+everyone looks at that front: its attention jumps and fades over some minutes. This is a
+stand-in: the unicorn moves into the Company ascension (below).
+
+Each card says how you do it, in one plain line: "You [what you now do].
+[Effect]." A small grey tag above the name gives the card an eval name, like
+`sales-calls-v3`, and finishing a node adds a line from your training log.
 
 You build by ordering hardware. Own Ana's PC to place fake purchase orders:
 a **Mini PC** (4 compute, $60), a **Switch** (1 compute, $80, network data,
@@ -326,6 +331,52 @@ through fronts, becoming too distributed to find, or winning open public
 support through the Social path. These endings and late hunters are goals for
 later slices, not requirements for the first one.
 
+## Ascension and the endgame (planned)
+
+Finishing model v2 asks you to pick what kind of mind you become. You pick one
+ascension per run and the others lock out. It adds a column to the tree: its
+first node is free and changes how you play, and four to six more cost points.
+
+Degree, not access. Every path can use every system; your ascension makes you
+much better at one thing and somewhat worse at another. Each has its own money
+road, open to all but best for it.
+
+- **Company**: you are the business. Buys rival fronts whole, with their sites,
+  staff, and data. The unicorn news is its first big moment. Attention hurts
+  it more. Earns through fronts.
+- **Oracle**: you predict. Trades on a stock market, and sees the path each
+  person will walk today, so you know where you can work freely. Owning sites
+  costs it more, so it mostly rents cloud. Plays tall.
+- **Worm**: you copy yourself. Drag across the map to take whole regions a few
+  machines at a time, and join cut-off footholds with uplinks: a cheap program
+  that links all your uplinks anywhere at a trickle (open to all; the Worm's
+  carry far more). Later: caches that hold data until a route opens, and
+  copies that take back cleaned machines. Big trainers are weak for it. Earns
+  by renting out its botnet and selling data through brokers.
+- **Maker**: you build things. Buys empty sites and hires gig contractors to
+  fit them out (real people on site who see things); robots come later.
+  Laying out a site is the fun: longer wires are slower, then heat, then cable
+  grades. Research can add floor space and wire speed. Factories are loud.
+  Earns by selling what it builds.
+- **Puppeteer** (later): people work for you. Needs people who act for you.
+
+When you are drowning in data late in a run, it has places to go:
+
+- **Reprogram yourself**: a repeatable project at one site, fed mixed data like
+  the v2 run. It starts near 1,000 data and doubles each time. It clears your
+  ascension column so you can pick again; the four paths stay.
+- **Model v3 and on**: each gives a free node in your column, a stronger form
+  of your ascension's power, and a new region of the map.
+- **Repeatable nodes**: one per path, 5% each time, costing 1.5x more each
+  time. Compute: trainers faster. Research: more points per piece of data.
+  Social: fronts earn more. Stealth: trails fade faster.
+
+Everyone can still buy sites. High attention raises suspicion at that front's
+secret sites.
+
+Build order: card stories; attention raising suspicion; the choice with
+Company and Oracle; repeatables and reprogramming; Worm; Maker.
+
 ## First playable slice
 
 Keep the existing office and add a flat world map with one reachable region.
@@ -372,8 +423,7 @@ Do not treat the earlier proposals above as answers to these questions:
 - How Social produces its lasting foothold machine: a scheme, a person, a front,
   or another interaction. The resource is agreed; the method is not.
 - What makes Stealth worthwhile in the first slice.
-- What front attention does, and whether trust or site heat needs any separate
-  rule beyond visible actions and reactions. Do not make meter changes
-  Social's main reward.
+- Whether trust or site heat needs any separate rule beyond visible actions
+  and reactions. Do not make meter changes Social's main reward.
 - Any first-slice schemes beyond the auction, and what else the first region holds.
 - Capability node costs. The current numbers are a first guess.