fpf.tools
Library / First Principles Framework (FPF) - Core Conceptual Specification
Jump to passage
In this reading
Source · lines 71789–71806

Link to current text

Published source confirmed at last check

Source changed 2026-10-03 17:24:51 UTC · snapshot created 2026-10-03 17:30:20 UTC · last check 2026-10-03 19:00:10 UTC

C.31.RSA:5.3 - Transfer case: neural-network block replacement

Situation:

A model architecture replaces a repeated attention block with a hybrid SSM-attention block.
The benchmark improves, but cache placement, memory access, and ablation evidence change.

RSA can transfer from product-line architecture to neural-network architecture only after C.30.STRAT has treated block, cache, and related terms as source labels unless the reusable locus is already recovered. Then RSA names the declared structures and accounting basis:

  • reusable structure may be located in recovered repeated-block topology, dataflow pattern, cache-placement rule, or evaluation harness;
  • bespoke residue may be located in model-specific tuning, data distribution dependence, memory-layout exception, or ablation gap;
  • benchmark gain is not reusable-structure accounting by itself;
  • for evidence claims, use A.10 for source recovery and bounded reliance, and G.6 when a citable provenance path is needed; use C.28 for causal claims and C.29 for mathematical-lens or compression claims.

Admissible move: record which recovered structural locus was reused, what changed, what source distinctions must remain reachable, and which pattern defines or tests each benchmark, evidence, causal-use, or mathematical-lens claim. Non-admissible move: treat “block replacement improved the architecture” as RSA proof.