# Research basis

The experiment is an original educational recurrent-network study on a fully specified exact synthetic rule. No external dataset, pretrained model, publisher illustration or reference implementation is copied. Original code/data/weights/results are Apache2.0, Renchao Wu. It does not reuse quarantined HAR data or licensing.

Primary methodological reading: [Dive into Deep Learning, Section9.7, Backpropagation Through Time](https://d2l.ai/chapter_recurrent-neural-networks/bptt.html), official author textbook site, actually read using official CUA on2026-11-04. Its unrolled shared-parameter chain-rule explanation and gradient-risk discussion informed the verification plan. This implementation is independently written with tanh and biases; the detailed textbook section's simplified identity/no-bias example is not claimed as an exact reproduction. No textbook prose/code/figures redistributed. Own full recurrent-gradient finite-difference tests establish the actual implementation rather than citation alone.

Least-significant-bit-first encoding exposes carry as a finite-memory task; that is a designed input representation. A106parameter tanhstate does not establish a general arithmetic language model or novel exact algorithm. The exact XOR/ANDcarry reference is deliberately separate and remains the authoritative answer. The learned model's wrongbits/integers are never replaced by that reference.

Unordered operandpair grouping excludes orientation/width and prevents familiarreverse/leadingzero representations crossing smallfit/val/test. Longer/largernumber probes change both magnitude and sequence length, with a same-valuepaddingcontrol. Carry/zero/max stress can reuseknownsmallpairs; labels explicitlyretain theirsource roles. Overlapping evaluation suites dedupuniqueorderedpair-width cases, with memberships preserved. These are dependent mathematical probes, not human observations or independent random experiments.

Reset-each-bit is an intervention on the same trained model, not a secondfit or a trained memoryless comparator. It changes hidden-state distribution as well as removing recurrence. Hiddenactivationplots do not identify a causalcarryneuron. Finite successfulprobes do not prove the algorithm for arbitraryintegers or infinite lengths. Probabilityoutputs are uncalibrated; wholeintegererrors remain distinct from bitaccuracy. This is not a substitute for exact financial/engineering computation.
