Behind the builder / 3 October 2026
A WorkMachine needs to know when its memory is wrong.
R3.9.6 brought contradiction and forgetting into UBIKITOS. The day connected a sharper memory system, a mathematical audit and the practical work of learning and publishing.
Remembering everything is only half the problem
On 3 October, the memory question became more interesting. UBIKITOS had reached R3.9.6, called Contradiction and Forgetting. A system that carries experience into its next task also needs a way to stop carrying the wrong experience forward.
A fact can become stale. A later observation can contradict an earlier conclusion. A better candidate can replace an older one. Keeping all of that in storage is useful for reconstruction, but it does not mean all of it should remain available for current reuse.
The local implementation gave those cases explicit dispositions, including active, stale, superseded, contradicted, quarantined, retracted and archived. The historical content and evidence remain. Forgetting here means excluding a candidate from current reuse; it does not mean deleting the record or changing the model’s weights.
A conflict is a reason to investigate
Imagine a campaign brief based on an earlier catalogue detail. A newer source disagrees. The useful response is to preserve the disagreement, identify which candidate may no longer be reused, and establish what fresh evidence is needed. Quietly choosing the more convenient answer would make the next draft look cleaner while making its basis harder to trust.
The R3.9.6 result implements reviewed relationships and a current reuse gate. Returning a restricted candidate to active use requires fresh review and source validation under the declared rules. Superseding an old candidate does not automatically approve its replacement.
The report records 2,056 selected passing tests and three inherited subtests, alongside a separate recovery campaign. Those are meaningful local engineering results. Native integration, independent review and formal acceptance remained open.
The mathematics needed an audit too
The day also brought a contract and mathematical-substrate audit of R3.8. The retained record reports 3,014 distinct collected test cases with final passing status and a separate 324-case mathematical-oracle suite. Architecture checks passed.
The audit retained the failures encountered along the way and the fixture-recovery work that resolved them. That history matters: a final green result should not erase the conditions that made the earlier run fail.
It also identified remaining work around real-provider context and exposure, native restoration, retrieval quality and object coverage. A test campaign can establish the behaviour it exercised. It cannot quietly stand in for connections or acceptance checks it never performed.
The products gave the architecture somewhere to go
The publisher integration advanced as a locally tested candidate, with native activation still disabled at the reporting checkpoint. Learn acquired a separately packaged local demo. Neither result alone established an ordinary live workflow for a customer or learner.
Those products make the same architectural question concrete in different ways. A tutor needs the learner’s relevant attempts and source material. A publishing assistant needs the right book information, editorial context and approval. In both cases, memory must remain tied to what it supports and what may happen next.
A complete jazz-crossword manuscript draft also reached its automated structural checkpoint. Human clue checking, fairness and blind solving remained distinct tasks. Across the portfolio, a finished artifact was increasingly becoming the beginning of a different kind of test.
The next receipt has to come from use
The Chairman close and its reporting cascade agreed on the practical constraint: technical capability had advanced, while paid customer proof and measured work transfer remained open. The cascade offered several views of the same checkpoint, not several independent demonstrations.
I want the next useful win to connect the machinery to a bounded job someone can judge. One approved workflow, an inspectable outcome, and an honest account of the attention it required would tell me more than another unqualified completion badge.
The memorable part of the day was that memory gained a brake. Experience can help the next decision only if the system can also recognise when that experience should stop guiding it.