Open questions and limits

What is not known yet, and the claims Requant deliberately does not make.

Open research questions

1. Lookup-table precomputation, measured. The advantage of bit-plane lookup tables over tensor cores is bounded analytically at 4–12× slower; no optimized kernel has been written to confirm it.
2. Dedicated hardware. A chip that does only int8 GEMM is, in effect, an AI inference chip. Its existence is not in doubt; its advantage over existing AI accelerators is unknown and has not been modelled.
3. Newer architectures. TNet has been measured on Turing (CMP 50HX) and Ampere (RTX 3090) with the research benchmark, and mined on Blackwell (RTX 5070). Ada, Hopper and dedicated AI accelerators have not been measured.
4. Energy and fairness across hardware. No matched energy-and-time comparison between CPUs, GPUs and accelerators exists. The only energy figure is a coarse 52 µJ per ticket on an RTX 3090.
5. A fused requantization step. Requantization is a separate pass today (7.2% of an attempt on the RTX 3090). Fusing it into the GEMM epilogue is a miner optimisation that must stay bit-exact.
6. Light clients. Checking a header needs the 512 MiB epoch weights and about 11 ms, or trust in a full node. No light-client design exists.
7. External review. Neither the work function, the chain rules nor the Magnet negative result has had an independent cryptographic review. Independent replication of the measurements is also missing.
8. Useful work. Whether external, useful data could be bound into such a proof of work without reopening reuse attacks is an open problem; TNet's inputs and weights are random.

Claims Requant does not make

• No ASIC resistance. The design argument is limited: the work is the int8 inference kernel and the weights change every epoch, so they cannot be fixed in silicon. That is not a proof of resistance.
• No useful work. Mining does not perform anyone's AI inference or training.
• No post-quantum label. The proof of work is hash-based; ed25519 signatures are not post-quantum.
• No novelty claim for matrix-product proof of work. Pearl deploys an int8 GEMM proof of work; TNet borrows its ticket structure.
• No security or value claims. Requant is an experimental, pre-alpha test network. Its coins have no value, and the network may be reset when the rules change.
• No hardware rankings. Performance figures describe the listed devices and settings only.

Scope of the record

• The research record was produced quickly and over a short period; most GPU results come from a single CMP 50HX, with TNet repeated on one RTX 3090.
• Raw data lives outside the repositories; the research notes record source hashes and environments.
• Main-network parameters (genesis, development-fund key, address prefix in use) are not final.

Sources: spec/07 §4, ADR 0016, STATUS.md, THREAT-MODEL.md, RATIONALE.md.