True under stated assumptions
A machine-checked or mathematical result. It does not become an empirical systems claim without a corresponding experiment.
Research
The work spans authority, gated computation, and vector-native systems. Every result is labeled by what was proved, measured, demonstrated, or merely proposed.
Published research · newest first
A live demo of OntQL, the query language of the Substrate. Visitors write an ontology in short English sentences. A closed-grammar compiler turns definitions into types, refuses undefined words and wrong types with their line and column, and runs the result as one transaction on the visitor’s own Substrate. Every answer cites the sentence it came from. Conflicting values and open-world gaps return UNKNOWN instead of a guess. The ten guided scenes are also 25 checks that the server must pass before it starts. No language model is involved.
Inside two Qwen3 base models, the days of the week, the months and the hours of the day sit at measured angles on a circle, and a shift turns the state by close to the ideal angle: 47.0° for one day at 8B, where one step is 51.4°. Read layer by layer in two model sizes and two languages, with causal tests showing the 8B model reads its answer from the whole state rather than the circle, half steps that land between hours but snap to days, every channel at full width, and the predictions that failed.
We reviewed OpenAI’s nine misalignment reports and three notices. A Governed Ontology with a typed capability vocabulary brokers action; Schemen Gate enforces the exact permission before execution. Explore 41 checks, live model proposals and a real file-transfer comparison, with independent evidence and the boundaries of each control.
A three-minute silent film with animated banners: follow layer readouts, inspect 3,840 residual channels, then test 288 controlled patches. A layer 38 residual patch restores Paris to 97.1% probability after Eiffel becomes Tokyo. The film includes the full-restore control, component-patch null results, and the limits of one prompt pair.
What a machine is doing, how often, and whether something is wrong with it, read from the averaged power reports a meter sends every few seconds, on a laptop CPU with no training loop. Running and idle at 0.9955 by a threshold baseline against a compressor's own valve signal, every air-leak failure recognised with the failure held out, accuracy equivalent to open state-of-the-art classifiers on valve lag at a fraction of their cost, and a feature-map lift that reads accumulator pre-charge. With the negative results, a stopped comparator run, an external review's corrections, and the paper.
Byte-exact tenant isolation on one copy of a language model. Each tenant's lane owns private attention and norms over a shared, frozen backbone, and its output is byte-identical whatever its neighbors send, hold or do, yet three lanes serve up to 2.58× the throughput of the same model serving tenants one at a time on Qwen3-8B, and a tenant costs megabytes of private weights instead of a model copy. Measured on Qwen3-8B and Gemma 3 4B and 12B, with an exactness ledger of every comparison, the timing channel that remains, the defects the evaluation found in itself, and the paper.
An interpretability report on a 1.7B base model, using only standard public methods. It holds a blank placeholder before it commits an answer, looks a fact up in a single block near block 20, and carries calendars as rings and numbers as a line. With the written findings, walkable activation manifolds, causal patching, steering against random nulls, two corrections, and a per-figure receipt.
An interpretability report on a Lean theorem prover. At the mid-proof decision state its next tactic is linearly readable at 0.867 against 0.643 for a general encoder, one direction steers the choice, and two earlier headlines are corrected in the open. With the written report, the proof reader, and the activation visuals.
Research map
Evidence policy
A machine-checked or mathematical result. It does not become an empirical systems claim without a corresponding experiment.
Architecture, data, denominator, metric, seeds, and failures stay attached to the result.
Proof that the declared path runs—not proof that every alternate path or environment has the same property.
Read alongside the work
The public repository can be inspected directly. A separate conversation reviews Schemen Gate against Dario Amodei’s frontier-control argument and keeps the distinction between authority control and broader AI safety explicit.
Start with the critique, then inspect the implementation it cites.
Read the review