# Interpretation limits

This is a known mathematical task in a designed least-significant-bit-first representation. Addition, carry and BPTT are established methods. “Original synthetic” describes independently generated examples and authored artifacts, not invention or ownership of arithmetic facts. No external dataset, pretrained model or textbook implementation is copied.

One 106-parameter RNN was trained on widths 1–6. The selected model passes 64,856 declared finite cases, including 64,499 held-out/new-range/padding cases. That does not exhaust all 64-bit operand pairs or prove arbitrary lengths. The wider random suites contain only 500 cases each. Larger values and longer sequences change together; padding is a separate control, without claiming full causal isolation.

The stress inventory includes 17 exact fit representations. Reverse pairs and leading-zero views share numeric groups. Overlapping suites retain their memberships while unique ordered operand/width cases are predicted once per mode. Repeated cases and modes are dependent. Neither 64,882 suite memberships nor 129,712 mode-case rows are independent new problems or extra training runs.

Resetting state changes the SAME weights' hidden-state distribution. It is not a separately trained or optimal memoryless baseline. The difference shows sensitivity to this intervention on these inputs, without identifying a causal carry neuron or proving a universal mechanism. Hidden-activation colors are numeric guides, not semantic labels.

Bit and integer accuracy have different denominators. Reset predicts `14` for `1 + 15` at widths 6 and 64: 3/7 versus 61/65 correct bits, while the integer remains wrong. All wrong positions, full integer strings, log probabilities and traces remain. Probabilities are uncalibrated; loss uses stable log-softmax rather than clipped or rounded display values.

Canonical string operands are bounded to 64 bits; sums may need 65. JavaScript Number is never used for operands or decoded integers. Exact XOR/AND carry is checked against BigInt addition and remains authoritative. It does not repair neural answers: a synthetic zero-output-weight test leaves the learned answer wrong while the reference stays correct. Use exact arithmetic for real calculations.

Current inputs and reports stay in page memory. Clear and refresh remove them; deliberate downloads remain where saved and include full inputs and hidden activations. There is no external API, font, analytics or browser storage. Hashes check consistency, without establishing mathematical proof or measurement authenticity. Local validation, independent acceptance, public availability and adoption are separate states.
