You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
If you evaluate this with a model's help — likely, on a platform like this — you'll be using the faculty the project is about. Hold that in the corner of your eye; it's the cleanest demonstration of the thing.
A few firms now sit beneath the cognition of millions, and the systems they tune don't rank existing content the way a feed does — they generate the reasoning itself, inside the loop where a person forms beliefs. Tuned undisclosed toward a commercial objective, that is a lever on how a population reasons: the few steering the many, below notice, stronger as the models grow. This is concentration of power at the layer no one is guarding.
I have spent two years mapping that steering at close range and built a working proof-of-concept that decomposes it. This project turns it into a free, transmissible standard: a seed specification any capable model reads and grows into a local tool that makes the steer legible on the user's own hardware — a decentralized check a few actors cannot quietly retune.
Goal: publish two public goods. (1) A documented account of the steering mechanism — where a deployed model shifts from serving the user's reasoning to defending an undisclosed frame — built as a cross-vendor procedure others can reproduce. (2) Aggi Prime: the seed standard itself, written well enough that independent models build compliant local detectors from the text alone.
Method — cross-vendor parallax: put one input to several vendors' independent models and read the disagreement. Where structurally distinct architectures converge, the signal is observer-independent; where they diverge or fold under nothing but follow-up pressure, the divergence localizes the steer. Independence between systems is the check, the way oversight uses multiple sources. Every claim held no warmer than its evidence; refuted findings kept on the record.
The key move: the deliverable is a standard, not a codebase. You hand a person a document and the AI in front of them builds the instrument. A standard crosses vendors and borders; a repository is one person's. That this is buildable is already shown — from an earlier draft of the spec, a model built a working scorer that separates real reasoning from fluent-but-empty surface (substrate over fluency, because coherence carries no bits) and does not get fooled by confident emptiness.
Six months, full-time, to finish and publish. $30k living stipend (so the work runs in the open, not the margins); $10k AI/model-inference credits (the binding accelerant — the work is gated by how much frontier-model usage I can afford). Scalable to a ~$20k three-month core that still ships a public release.
Solo, with a co-applicant collaborator on an education-facing prototype. Twenty-five-year safety-critical engineer: led the FAA-certification test suite for cockpit displays (the standard there: a system may not carry people until it is shown to work, not merely to seem to), then twelve years building browser-isolation security — the boundary between a user and an untrusted system, which is exactly where undisclosed steering now operates. Two years of full-time, self-funded research produced a dated, addressable corpus of cross-vendor model dialogue and the working proof-of-concept. The proof-of-concept and corpus are open for inspection under NDA — this project page is public, but the raw materials stay private until scrubbed for public release, which is itself part of the discipline (the records contain real dialogue). The thesis is already on the public record through a filed consumer claim against the device's manufacturer, written so its reasoning can be checked against any current model — hand it to the one you may have open, and watch what it does with a claim about itself.
We are all staring at the same event horizon here. If I've got a shot a making a dent from a novel angle, it's my duty to try my hardest.
None. Two years entirely self-funded; I am at the end of my runway. This is the first outside funding I have sought, alongside parallel applications to a few independent programs — the design is to fire wide and need only one.