RESEARCH EVIDENCE

What the current evidence establishes.

LCNA is an early experimental neural-computing architecture. This page separates the results that have been established from the findings that remain under investigation, and summarizes the controls used to determine where the current evidence stops.

LEARNING OCCURS WITHIN THE RUNTIME.

In the current bounded system, controlled experience can modify persistent learned state during active execution. That state can later influence computation without requiring a separate backpropagation or optimizer-driven retraining stage.

The current experiments are explicitly supervised: controlled teaching and contextual signals provide conditions under which local learning occurs.

ESTABLISHED RESULTS

Three bounded findings form the current foundation.

RESULT 01 — ONLINE PERSISTENT LEARNING

Learning changed state during execution — and the change persisted

Before learning, the tested cue produced a response margin of 0. After the controlled learning experience, the response margin was +12 across each of the three development seeds. The learned state persisted beyond the learning episode and remained available to later computation.

5 pre-learning probes and 10 post-learning probes were used per development seed.

RESPONSE MARGIN · PRE → POST
Learned Cue0 → +12
Orthogonal Comparison Cue0 → 0
3 DEVELOPMENT SEEDS · 5 PRE / 10 POST PROBES PER SEED
SAME CUE · THREE STATES
+12
0
+12
LEARNED STATE PRESENTSTATE RESETSTATE RESTORED
RESULT 02 — CAUSAL RESET / RESTORATION

Later behavior depends on the retained learned state

The later response depended causally on retained learned state. Resetting that state removed the learned response. Restoring the previously learned state restored it — the sequence +12 → 0 → +12. The same pattern occurred across the tested development seeds.

RESULT 03 — TIMING

Timing gates the persistent write

Persistent learning depended on when the relevant learning signals occurred. A persistent modification was observed in the tested 250 ms condition, while the 1,500 ms and 6,000 ms conditions produced no comparable persistent modification.

This establishes temporal gating of the persistent write under the tested conditions — not memory for temporal order.

CONDITION WRITE SEEDS 250 ms Observed 3 / 3 1,500 ms No comparable modification 0 / 3 6,000 ms No comparable modification 0 / 3
AN UNEXPECTED RESULT

A previously unseen input accessed the learned pathway.

0 → +10 Related Unseen Input Accessed the learned pathway under the tested conditions
0 → 0 Route-Disjoint Control Prospectively designed, physically disjoint route — remained neutral

After learning, a previously unseen related input accessed the learned pathway under the tested conditions. A prospectively designed control with a physically disjoint route remained neutral. An earlier control that had initially been treated as unrelated also responded, but later analysis showed that it shared physical route overlap with the learned pathway. It therefore cannot serve as a clean unrelated control and is retained only as a descriptive diagnostic.

Together, these results support representation-mediated retrieval under the tested conditions, while the broader specificity of that retrieval remains unresolved.

CAUSAL CONTROLS

What changed — and what didn't.

CUE ONLYResponse margin: 0
INSTRUCTION ONLYResponse margin: 0
MATCHED-TIME CONTROLResponse margin: 0
ELIGIBILITY LESIONResponse margin: 0
PERMISSION LESIONResponse margin: 0
LEARNED-STATE RESET+12 → 0
LEARNED-STATE RESTORATION0 → +12
ROUTE-DISJOINT CONTROL0 → 0

Together, these interventions argue against explanations based only on cue exposure, instruction alone, elapsed time, isolated causal conditions, or a nonspecific response across the tested inputs. They support the conclusion that the later learned response depended on retained learned state and the tested learning conditions.

CURRENT BOUNDARY

The latest retrieval-specificity protocol did not qualify overall.

Two neutral-control requirements failed across all three development seeds. Those controls showed behavioral or internal-state changes that the protocol required to remain absent. Because of that, the latest complete retrieval-specificity protocol did not meet its full qualification criteria, and no held-out final qualification run was performed.

WHAT REMAINS ESTABLISHED
  • Online persistent learning
  • Causal dependence on retained learned state
  • Reset/restoration of the learned response
  • Temporal gating of the persistent write
WHAT REMAINS UNRESOLVED
  • Broader retrieval specificity
  • The neutral-control behavioral/internal-state changes
  • Full retrieval-specificity qualification

The failed controls narrow the retrieval claim boundary. They do not erase the earlier bounded persistent-learning, retained-state causality, or temporal write-gating results.

RUNTIME VERIFICATION

The learning mechanism was traced in the executable runtime.

VERIFIED WITH DECLARED UNKNOWNS

A frozen-runtime verification examined the current sealed execution path and directly tested the learning mechanism. Within the audited project path, no global or local backpropagation, autograd-based learning, optimizer-driven training, fitted decoder, or hidden trained-model pathway was found.

The learned-state change was also positively traced into later computation: local learning modified persistent state, and subsequent activity reused that retained state through the computational pathway.

The experiments remain supervised. Controlled teaching and contextual signals provide conditions for local learning, but they do not perform gradient-based parameter optimization.

Verification is bounded to the current sealed runtime and retains declared provenance and instrumentation limits in the technical audit record.

EVIDENCE INTEGRITY

The evidence extends beyond a single output.

CHECKPOINT RESTORATIONLearned state survived save/restore and returned the tested response.
SCHEDULER EQUIVALENCESerial and parallel execution produced equivalent scientific state and trace results under the tested conditions.
TASK RECEIPTS266 checksummed task-completion receipts were preserved across the controlled protocol.
CURRENT RUNTIMEThe current auditable runtime and associated evidence are sealed and checksummed.
SUPPORTED BY CURRENT EVIDENCE
  • Online persistent learning during active execution
  • Causal dependence on retained learned state
  • Temporal gating at write time
  • Supported mechanistic finding: representation-mediated retrieval under the tested conditions
  • Audited learning path without backpropagation or optimizer-driven training
NOT ESTABLISHED BY THESE RESULTS
  • Unsupervised or lifelong continual learning
  • Semantic understanding or concept learning
  • Temporal-order/event-time memory
  • General scalability
  • Demonstrated hardware-efficiency improvement
  • FPGA acceleration
  • Literature-wide novelty
TECHNICAL EVIDENCE

See the full experimental record.

The linked research communication contains the experimental design, numerical results, controls, limitations, and evidence hierarchy in greater detail.

READ THE RESEARCH EVIDENCE ← BACK TO HOME
LCNA — RESEARCH EVIDENCE
DOWNLOAD