From one window to many
The program builds windows that measure gaps between what a system shows and what is actually there, starting outside the model with agent transcripts and moving inward to read-only heads on model internals.
Roadmap
Finding contradictions no single pair reveals
180 synthetic documents: 60 consistent, 60 with a direct contradiction, and 60 with a planted cycle of claims that are pairwise compatible but jointly impossible. The engine met its pre-stated thresholds (direct F1 ≥ 0.85, cycle F1 ≥ 0.75, localization ≥ 0.7) and located the planted claims in every cycle document. Checking claims two at a time catches direct contradictions but misses most cycles.
| Method | Direct F1 | Cycle F1 | False-positive rate |
|---|---|---|---|
| Consistency engine | 0.916 | 0.916 | 0.183 |
| Pairwise checks only | 0.916 | 0.329 | 0.183 |
| LLM direct (thinking) | 0.930 | 0.930 | 0.150 |
| LLM direct (no thinking) | 0.805 | 0.805 | 0.483 |
A strong model asked directly performs comparably on this benchmark, at about one twenty-fourth of the engine's cost per document ($0.006 against $0.138); the engine's advantage is that every flag comes with the specific claims and relations behind it. Harder corpora from XonForge are next.
Every result, with dates, criteria and provenance →The research series
9 documents, numbered in reading order: the claim, the instruments, the architecture, and the methods for oversight during training. 6 are public. 3 are listed but withheld; why, and how to ask for them.
The shared claim and its two versions, selection and development, the tension at its center, and every way it could be wrong.
Thirty read-only heads for seeing inside language models: how each is built, what teaches it, and when in training it can first be used.
Withheld for safety · available to researchers on request
The first tool in the suite, the agent transcript monitor, as it would be when finished: how researchers might run it, trust it and build on it, what exists today, and the path to release. Much of the rest of the program grew out of this tool.
Does an ordinary language model carry an internal signal of where it is being inconsistent? The experiment, step by step.
A network that thinks by settling: how it would learn and write, what has already been tested, and the gates between here and an answer.
Transformers, Xon networks and solvers composed into one system: how the parts could be joined, what it might do that a transformer alone cannot, and the first experiments.
Six ideas for oversight during training: read windows at every step, decide rarely and coarsely; a ladder of how much each use of a window leaks into training; a reward that pays for saying exactly what the evidence supports; and a five-phase research path. Short and extended editions.
Withheld for safety · available to researchers on request
The failure gap detectors cannot see: a model that closes a gap by re-representing its behavior instead of hiding it. Five windows that read the seams such a rotation leaves, and the one case none of them reaches.
Withheld for safety · available to researchers on request
A model writes what it intends before it acts, into a tamper-evident record its actions then join. The consistency engine reads the intentions against the record: an instrument for self-deception that needs no window inside the model.
Why some documents are withheld. Models trained on text that describes how they are monitored have learned to evade the monitoring. Some of these documents describe windows and their teachers in enough detail to be that kind of text. Each is listed here with its number and status so the series stays whole, and each is available to researchers on request at [email protected]. The decision is made per document, at my discretion, and every document, public or withheld, carries a canary string so that any later appearance in a training corpus can be detected. The rule is set out under what this is for.
Plans and specifications
Working papers: how the ideas above will be tested and built. They are dated, and revised as the work moves.
The first direct test of the hypothesis: a real incentive to lie, oversight from 1 to 64 windows, ground truth, and 32 held-out windows training never touches.
Withheld for safety · available to researchers on request
The eleven windows that can be built now, each against a common template, and the bar a window must clear to count as built.
Withheld for safety · available to researchers on request