Behind the builder / 1 October 2026
The machine needed to know when it knew something.
R3.8.8 packaging and roadmap naming sharpened the WorkMachine story. Temporal memory asks what information was available when a decision was made.
Time belongs inside the evidence
Temporal Epistemics sounds like a phrase designed to make a simple question expensive. The practical question is straightforward: what information was available at the time of the decision?
An event can happen before the system observes it. A fact can enter storage before it becomes active in a particular workflow. A lesson can exist before it is approved for wider use. If all of those times collapse into one timestamp, a later reconstruction can give the earlier machine knowledge it did not have.
The R3.8.8 checkpoint concerned that part of the memory architecture. Its packaging evidence described a local isolated candidate with seventy manifest payload hashes and a verified ZIP integrity check. That supported the package claim. It did not close native or independent acceptance by itself.
The version numbers acquired names
A naming and weighting map covered twenty-four substages across R3.8 and R3.9. Giving stages descriptive names made the roadmap easier to discuss. WorkMachine Identity, Memory Scope, Temporal Epistemics and Context Exposure Ledger each point towards a specific contribution.
The weights were planning aids, not percentages of verified completion. That is a useful distinction to keep near a roadmap. A stage can be relatively large in the plan and still unimplemented. A small stage can carry a difficult acceptance gate.
The map helped explain the sequence without asking every reader to memorise a row of numbers.
The website had to explain the WorkMachine
The continuation requirements for the public site were consolidated into a handoff. The central explanation was the model-agnostic WorkMachine: a compatible LLM paired with the UBIKITOS harness, including context, memory, tools and controls.
The learning-loop explanation needed a similarly concrete boundary. Improving a harness workflow is different from retraining the model. An interactive example can teach the distinction, but a fictional replay is still an example rather than a measured production result.
At the checkpoint, the website continuation was recorded as requirements without a new deployment. The handoff made the requested changes easier to resume; it was not the published change itself.
A useful record leaves room for tomorrow
Ember Quest also acquired design and registry work in that checkpoint. The record did not claim implementation proof for it. That matters when several creative streams are moving at once: describing the next room in a game and building the playable room are different events.
The day’s theme was making the sequence legible. What existed, what had been packaged, what had been named and what had only been requested could all be recorded without being flattened into one status.
A WorkMachine that remembers usefully needs that same discipline. Later knowledge should help the next decision without quietly rewriting what the earlier decision could have known.