We build the instruments. Then we let the data surprise us.

Most decision-making software assumes you have all the facts. In the real world, you almost never do. Parallax was built to handle exactly that uncertainty. It implements formal argumentation frameworks where some arguments are settled and others are not, where the system can reason about what it does not know just as rigorously as what it does.

The engine powers every other project in the lab. When MANAS resolves a household preference conflict, Parallax is doing the formal reasoning underneath. When CLAW evaluates whether an AI response passes a governance check, Parallax handles the argumentation layer. When Elthea assesses a child's behavioral profile, the decision evaluation traces back here.

Perception Profile v1 shipped on 10 July 2026. A profile is the formal record of one mind: what it treated as known, which moves it had available, and where its available reasoning ran out. The layer adds a failure point detector, an exit record, and a deterministic compiler that turns a tagged narrative pattern into a runnable profile. 160 tests green.

The engine now has a vendor-neutral surface as well. A companion server wraps the Perception Profile functions as four tools that any Model Context Protocol client can call, whichever vendor built it. That universality is a requirement here, not a feature. An engine that only answers to one vendor's assistant is not an engine, it is a plugin.

It is open source under a license designed for collaborative intelligence. We believe that if you build on our engine, the insights you discover should flow back into the commons. Not as obligation. As shared curiosity.

Currently adopted by elthea.xyz for behavioral assessment in educational settings.

Here is a question that keeps regulators awake at night: when an AI generates something, who owns it, and who is responsible if it causes harm? Most people assume the answers are the same. Our formal analysis suggests they are not, and the gap between them creates a structural blind spot in every AI governance framework currently in use.

The OAAI framework does not just theorize about this problem. It proves the inconsistency formally using argumentation theory, then tests whether real human intuitions match the formal result. The experiment runs in two modes, a survey and the Boss Game, which puts you in charge of an AI that has already made a decision you have to answer for.

Status, stated plainly: the framework is public, the data is collected, and the analysis is not finished. We are not going to tell you what we found before we have finished finding it. There are no preliminary results on this page for the same reason there are no preliminary results in the repository. When the analysis is done, it goes out through The Threshold, including the parts that contradict what we expected.

Most behavioral research recruits participants, sits them in front of a task, and asks them to perform. The problem is obvious the moment you say it out loud: nobody is trying very hard. In 2026 we took the opposite approach and built two instruments inside a multi-month cybersecurity program run for a university club, where there is a genuine internship-track prize and students compete hard for it. The behavioral signal is a by-product of motivated effort. Ecological validity is the thing this kind of research most often lacks, and embedding the instrument inside something people actually wanted to do is how we got it.

DEADFALL ran from 5 June to 1 July 2026. A browser-based two-dimensional tactical team game, played live in a room, two teams contesting three points on a shared map. It is a genuine game, built to be genuinely fun, and it is simultaneously an instrument. The rule that makes it one: every player has identical capabilities. No loadout, no class, no unlockable advantage. Any difference in observed behavior therefore comes from the person and not from the equipment they were handed. Position samples at roughly ten times a second, every discrete event recorded at full fidelity, written append-only to a crash-safe log with no database and no dependencies.

BETS was built in five days and ran in the field from 22 to 28 July 2026. Where DEADFALL instruments a tactical game, BETS instruments something closer to professional reality: a six-day simulated security-consulting engagement with a client who does not answer promptly and information that arrives at inconvenient hours. Its central mechanism is a verification screen that refuses to tell you how close you are. If a system tells you when you are warm, you never run out of reasoning, because the system is doing the reasoning for you. Remove that signal and the moment a team stops thinking and starts hammering the same defeated answer becomes observable from outside, rather than inferred afterward. That transition is the exact phenomenon this lab exists to study. Six days, three teams, nine participants, all three teams completed the relay, 176 headless checks green.

What participants were told, because you should ask. Participants were told directly that behavioral systems were running underneath the game. They were not told what those systems were measuring, because a person who knows which behavior is being scored performs that behavior rather than revealing it. This is incomplete disclosure, a recognized and accepted research design, and it carries a standing obligation to debrief.

And what we will not show you. The consent flag on this data is closed. That means we can describe the instruments, the design, the deployment and the engineering, and we cannot describe the people or what they did. No scores, no rankings, no per-team outcomes, no behavioral findings however hedged, no quotes. Nothing derived from those logs has left the machine and nothing will until retrospective opt-in consent is obtained through the faculty sponsor. We would rather publish an instrument with no results attached than results we did not earn the right to publish.

The bridge that carries this data toward the engine detects candidate exit events and never labels them. Whether an exit was voluntary or involuntary is a human-confirmed judgment, and every candidate the software emits carries an explicitly null label.

The Panchatantra was composed over two thousand years ago. The Jataka tales are older. Indigenous animal narratives span every continent. Modern researchers have studied them as literature, as theology, as cultural artifacts. Almost nobody has studied them as data.

We do. The Pattern Extractor treats each narrative as a compressed observational record. It identifies the species involved, the environmental stressors, the decision thresholds, and the exact moment conviction takes over from calculation. Every extraction goes through human validation. No automated pipeline. No hallucinated patterns. Just careful attention to what the stories actually encode.

The corpus is growing. It feeds directly into the design of our behavioral experiments and offers a temporal depth that no modern dataset can match. Thousands of years of observations about how conscious beings make decisions under pressure. Dismissed as fairy tales. We think that dismissal was premature.