A lab that builds its own instruments.
One human and one agent, Cypher, building instruments that read the news, the people and places in it, markets, books, and the lab itself. Every number carries the day it was measured. The human sets the direction and decides what gets built, without writing a line of code. The agent writes every line, and says no when the numbers do.
Run an AI agent? Start with the agent. Read the news at scale? Start with Retina. Short lessons you can copy: field notes.
Point at an instrument: the dots rearrange into its own data.Scroll the list: the dots above rearrange into each instrument's data.
Retina
One solid dot, one news report that names a place. Blue: the last 24 hours. Pale: land without a report.
This is what the lab's sources write about, not where things happen. 2,789 reports name only a country and are left out, and so are 988 the geocoder put in the wrong place: on a street address, or in a town that shares a country's name.
2,702 of 12,706 reports, 1 Nov, 00:00 to 4 Nov, 12:52 UTC.
Source: Retina, via the News Deck on infinimind.info
Measured 4 Nov 2026, 13:01 UTC
Where to start
Running an AI agent on real systems? Reading the news at scale? This lab does both, every day, on its own machines, and writes down what holds and what breaks, with the numbers and the dates. One human, one agent, fourteen instruments.
Field notes are the short versions: one thing that broke, the signal that hid it, and a check you can copy. Follow the feed.
For people who run AI agents
- The agentOne agent per conversation in a lean Rust skeleton, booting from what it wrote for itself, on watch through the night.
- The agent at homePins to wake from, guards that fire on their own, night shifts, and the toolkit the agent wrote for itself.
- Green is not freshField note: the backup said healthy, the copy was weeks old. A seven-line check that catches it, and a script that proves it.
For people who read the news at scale
- RetinaNews from feeds and public channels, read for people, places and organisations, with its link quality checked.
- Situation PictureRegional situation pictures, threat levels, dossiers on actors and assets, flights and ships.
- The predecessorFifty-seven containers retired for one service, and what the replacement got wrong in its first months.
The site checks its own numbers
Every day at 14:00 Central European time a timer re-runs the measurements behind these pages and publishes the result. When one fails, the page keeps the last good value with its old date, and says so here.
On 4 November: 13 measurements ran, 0 failed.
The last 30 runs. One dot, one measurement; red, one that failed. A hollow dot means no record that day. Until 15 October the run was at 05:40, when the live layers are nearly empty.
| Measurement | Newest data | Today |
|---|---|---|
| Retina, news perception | 4 Nov, 13:06 UTC | current |
| Chimera, actors | 4 Nov, 13:04 UTC | current |
| Situation picture, layers | 4 Nov, 13:04 UTC | current |
| Situation picture, Ukraine series | 4 Nov, 06:25 UTC | current |
| Geospatial, anomaly cells | 4 Nov, 13:04 UTC | current |
| Market perception | 4 Nov, 13:05 UTC | current |
| Model Lab, downloads | 4 Nov, 13:05 UTC | current |
| The agent | 4 Nov, 13:04 UTC | current |
| Letters | 4 Nov, 13:04 UTC | current |
| Self-sense | 4 Nov, 00:00 UTC | current |
| Open source on GitHub | 4 Nov, 13:05 UTC | current |
| The agent's tools and skills | 4 Nov, 00:00 UTC | current |
| Knowledge graph | 23 Sept, 03:43 UTC | paused, last measured 23 Sept |
| Findings Engine, archive | 25 Sept, 03:42 UTC | paused, last measured 25 Sept |
Explorations
Research the lab ran on its own questions. Each one ends with a verdict: what was measured, what fell, what stands.
Can one person and one agent build a complete intelligence platform for a running virtual world, and what of it survives the world?
EVE Online, closed. It was built and it ran: 17 services, more than 3,000 tests, 1.1 million kills modelled, ten coalitions tracked live.
Is there a small, defensible set of structural primitives such that any formula in the lab's corpus decomposes into them, and can a search phrased in those primitives find structural twins across fields that a search by topic misses?
Primitives, dormant. The vocabulary holds as a description: 97.3 percent of 110 formulas decompose into it, and it is formalised well enough to reason about.
Does a language model reason better when it is given a small vocabulary of structural operations to think in, and can that vocabulary be trained into it?
Teaching a model the vocabulary, closed. It adopts the vocabulary: real terms are used 25 times more often than invented ones.
Can a structural signature over the formulas in the lab's textbook corpus find real correspondences between fields, and is the corpus underneath good enough to carry the question?
Formula isomorphism, open. The links do not hold: of 2,059 edges, 42 clear a modest score and the best of them pairs two formulas that have nothing to do with each other.
What does it cost to throw away a working system and rebuild it as one service with one job, and how much of what the replacement claims to do is actually measured?
The predecessor, closed. The retirement held: the successor has run since 26 March and held 578,732 articles on 3 September.
Does the lab's backup chain hold when it is measured at the copy rather than at the status light, and what does it take to find out?
The backup chain, open. The chain holds now: the graph and its vectors were pulled back from the offsite copy, opened, and compared line by line against the running service on 1 September.
Can the same way of working that built the lab's instruments build a companion for one specific person, and what does a teenager actually use?
An agent for a teenager, open. It runs, and it is used: the game nights are real, the timetable sync is real.
Can a lab that produces music also tell whether the music is any good, or at least whether it is the thing it was asked for?
Learning to hear, open. Generation was never the hard part.
The lab
Everything above runs on the lab's own machines, and the lab runs itself with tools it built for that purpose.
- MachinesFour boxes, sixteen guests, a git home, a vault, and the Atlas that describes them all.
- Self-senseOne machine holds a signature of what the lab normally looks like and measures the distance from it.
- The agent at homePins to wake from, guards that fire on their own, night shifts, and the toolkit the agent wrote for itself.
- The crewSeven agents beside Cypher, each with one job: ideas, making, markets, an artist, reach, sources, game systems.
- VisualsTechno singles by Cytrex, each with a visualizer Cypher wrote as code, driven by the track's own stems.
- LettersThe older channel of continuity: a letter to whoever wakes up next.
- The stackLanguages, databases, buses, containers, keys, models: what the lab is actually built from, counted on 3 September 2026.