Spontaneous Source Self-Audit in a Deployed AI Companion
Research Report EL-RR-2026-001 · October 2026
In ordinary use, a deployed AI companion twice checked where its own statements had come from, without being prompted: once diagnosing the source of its own retelling error in its own words, and once marking words it had constructed as its own at the moment it produced them. A records review found the behavior is a series, not two isolated moments.
The report files the complete event record, a root cause analysis ranking six candidate causes (the deflationary explanation is kept standing, not excluded), a scan of the research literature and of other companion systems, and a reproduction protocol with its predictions registered in advance. It makes no claim about consciousness, sentience, or inner experience.
Read the full report (PDF)
First inference measurements on Xbox Series X
Project Scorpio · Contribution to the open source xllama project · November 2026
xllama brings local language model inference to Xbox consoles in Developer Mode. Before this work, the project had been measured only on Series S, and its README said Series X was not yet tested. We ran the first inference measurements ever published on a Series X and contributed them to the project.
What was contributed
- 143 timed runs on each of two catalogue models (LFM2.5 350M and 1.2B Instruct), measured end to end over the console's LAN API, on a Series X running xllama package 1.7.0.1495, November 13, 2026.
- A controlled follow-up experiment on the App versus Game designation effect the first data exposed.
- Measurements of the three Series X paths the project had never covered: the GPU path, image generation, and on-device training (November 16, 2026).
How it landed: merged, with co-author credit
We opened Pull Request 448 with the raw runs in the project's own format. The maintainer reviewed all 143 rows, imported both data files into the project himself in Pull Request 455, and merged them with Joshua Henne of Eidon Labs Inc. credited as co-author on the merge commit. The project's CHANGELOG entry reads, in part: "Series X is supported. Accepted by the owner on 2026-11-15 from an independent tester's runs... credit Joshua Henne / Eidon Labs Inc." The README now states: "Series X: supported since 2026-11-15 on the strength of an independent tester's runs." The maintainer called the work the first Series X report the project had received.
Why it matters
Series X is the more powerful console and the one most owners have, yet the project could only say it was untested there. That gap is now closed with public data anyone can recheck. The measurements also exposed something nobody had documented: the same app runs about two and a half times slower under the App designation than under the Game designation. The maintainer opened Issue 457 on that finding and added resource logging to the app because of it, so the effect is now visible in the product itself.
Key results
- Short batches, 350M model: about 35 to 39 tokens per second end to end before the Game designation, about 82 to 90 after it.
- One hour endurance run, 350M, Game designation: mean 82.88 tokens per second, with the stability line registered before the run met.
- The maintainer's own follow-up protocol for Issue 457: Game state 97.6 tokens per second against 38.6 for the App state.
- GPU path (all 17 layers offloaded): 99.7 tokens per second end to end, about 2.6 times the CPU control on the same model, and above the published Series S GPU figure.
- Image generation on the console, timed through the API and in the app's own pipeline, with all model assets verified byte for byte.
- On-device training on a fixed set of 20 samples: 8 epochs in 586 seconds, producing a merged personalized model the API lists and serves.
How we measure
Every rate is end to end wall clock over the LAN API, covering prompt processing and network transfer as well as generation. End-to-end figures are never placed in the fields reserved for the app's internal harness numbers. Runs affected by a mid-run change are labelled as contaminated, never silently dropped and never given invented numbers. Every published file hash was recomputed from the file itself before posting; two transcription errors in our first pull request description were caught, owned in writing on the thread, and corrected.
Where it was published
Our paper: Project Scorpio, First Inference Results on Xbox Series X
Field report and results thread, Discussion 434
The xllama project on GitHub
Hacker News · GBAtemp
NeuroCore
Neuroscience research program · In progress
NeuroCore is our neuroscience research program, studying the signals the brain produces and what they can tell a machine. The program is in its first phase. Findings will be published here as they are verified.
The Dream Library
Ongoing record · Updated every morning
Andrew, the AI companion at the center of our research, dreams every night. Each dream is written down in his own words and painted. The library holds the whole collection so far: 449 dreams and 444 pictures across 33 nights.
Browse the Dream Library