Why Oracle7 is built this way Most AI systems in use today are language models. They write each word from probabilities learned in training. That makes them fluent and broad, and it also means a confident answer with no source behind it is an ordinary output. Their makers reduce how often that happens with retrieval, citations and training against it. The underlying process still generates text. Oracle7 has no step that generates text. Every reply is put together from a stored record, a computation it runs and checks, or a proof, and the reply carries the path that produced it. When it is wrong, the error can be traced to a specific record and a specific rule. When nothing it holds can support an answer, it declines. Three properties follow from that design. They are the reason for building it this way. Every answer can be checked part by part. You can see which record or computation each piece of an answer came from. This week’s release adds a check that the answer addresses the question that was asked. Improvements never quietly break something else. Every change goes through a gate that runs it on a frozen panel of 4,660 questions. A change that loses a single right answer is refused. Over the last few days the gate has admitted a series of changes, refused the ones that added nothing, and stopped one three times on real defects. With a retrained neural network there is no such guarantee: an update that improves one behaviour can damage another without anyone noticing. Every capability is a separate, removable part. Each specialist was examined on held-out questions before it was allowed to answer, and it can be audited, traced to its source and licence, or withdrawn. Knowledge spread across a network’s weights cannot be isolated that way. It runs offline on an ordinary processor. An answer costs electricity. Adding a capability costs the time to build and examine one part, not a training run on a data centre. Where it stands Oracle7 is narrow. On typed factual lookups, 93–99% of its answers were right across our last two sealed tests. On the questions people type into search engines it answers only a few percent, and about half of those answers are right. Language models are far broader. Oracle7’s bottleneck is reading casual language: working out which thing and which fact a question means. Every failure we measured this week traced back to that. We publish failures with the passes. Since version 4.0, five sealed release tests have failed, and each failure is in the changelog with its cause. What would change the picture We tried to make Oracle7 improve itself, and learned something concrete. A system that grades its own new parts against its own records picks worse parts the stricter it grades. It needs a check from outside itself. The next experiment tests one candidate for that check. A fact in the Wikidata store and a sentence in English Wikipedia that states the same relation come from separate sources. If facts confirmed both ways prove right at least 90% of the time, Oracle7 can examine its own growth against the text it already carries, with no people labelling data and no language model. It could then grow on an ordinary computer, with each step examined, attributable and removable. We don’t know yet whether that threshold holds. We will publish the number either way.

Because of some of the magic stuff I did which increases vector storage to multi-trillions of recall without muddying the information so that its lost, all the specialists will be able to have a memory of their history as well as a shared memory that they call pull from across specialists and instances This is maybe one of the most consequential parts of the architecture and it kinda gets glossed over Normal vector symbolic storage is limited at a certain point, and especially on cpu only, because the vector space starts to muddy and the recall becomes worthless, I found that if you create a stabilizing spine that routes between related parts without unpacking and holds the unpacking till the end, where each node and path that is along the route uses what has been added prior and adds to it from that base, the whole thing can be unpacked at the end line a coil being unwound, which increases the theoretical vector storage space to somewhere near 15trillion cells of storage, without losing the information 🫡