A new level of learning
As of 26 October 2026, ALEPH has demonstrated Level 3 learning within the bounded scope covered by the latest LL04 implementation and audited test evidence. This is a meaningful capability milestone: useful, retained, revisable procedural discovery. She can learn a method from solved work, carry it into fresh tasks, retain it after restarting and revise it when corrected.
The qualified implementation is 36893a3, published in the audited repository state 17ac5f8. The latest audit accepted that work within its declared scope. This account assesses the supplied evidence rather than claiming a fresh engine run. Source: ALEPH_LL04_SOL_AUDIT_20261026.md.
A connected learning cycle
First, she investigates a requested outcome. Given a target and recorded execution data, she selects inspections and lets intermediate observations influence what she examines next. Ambiguity can lead to a question; missing information can lead to deferral.
Second, she constructs a reusable procedure from solved tasks. She learns selectors and identifier relationships that connect relevant records. Separate captures must successfully exercise the proposed procedure before it gains support.
Third, she applies that learning to fresh tasks. The learned procedure changes her later queries and lets her skip discovery work. Answers still come from freshly inspected records.
Fourth, she retains and reuses it after restarting. The demonstrated learning lineage spans six process sessions.
Fifth, she responds to correction. Correction suspends dependent procedures and pending work. Replacement versions preserve their history and require fresh validation.
Sixth, she recovers from interrupted work. She preserves uncertain outcomes while continuing independent work, without silently repeating the unresolved action.
Retained learning made later work better
The strongest evidence is the measured benefit of retaining what she learned. In two counterbalanced comparisons, retention saved six inspections and approximately 7.2–7.3 seconds per eight-task block, with 8/8 correct reports in both arms. The comparison arm could relearn, so the benefit was demonstrated against an operating alternative. Those results establish usefulness for this workload.
The next frontier
The boundary is the richness and reach of what she learns. Her demonstrated procedures concern querying execution captures across two formats, using supplied operators and validation rules. Learning a task dependency, precondition or failure relationship—and using it to choose a better later plan—remains the next substantial extension. Broader organism integration also remains unfinished.
This is our biggest capability milestone yet: evidence that retained procedural learning can make subsequent work measurably better. Completion of the full organism-wide Level 3 vision remains ahead.