LIMITATIONS.md

# Scope and limitations

The central contribution proposed in v3 is a gauge-aware reserve compiler for a restricted passive network class. Independent novelty is unverified. Ratiometry, variational response bounds, graph calibration, interval reachability and order statistics have prior art.

1. **Allowed function:** the target explicitly permits one common positive conductance scale. Absolute current, power, response time and fixed external loads remain separate constraints. Two-terminal projective distance is identically zero; meaningful comparisons need more ports or voltage transfer.
2. **Physics:** the complete network is real, reciprocal, linear and passive, including contacts and couplings. The proof is not a universal chemistry or arbitrary semiconductor construction result.
3. **Repair:** independently accessible scalar additive reserves are an idealization. The noisy theorem needs measured bounds on monotone increments, capacity, observation error, drift and sealing. Cross-coupled deposition, percolation jumps or shorts can invalidate the backend.
4. **Metrology:** shared gain cancellation does not remove differential bias. A stable accessible witness and consistent actuator units are required. Cycle checks cannot identify a pure gradient bias.
5. **Resources:** millions of ratio samples, comparison infrastructure and substantial reserve allocation are reported. Ensemble time excludes inference latency. Material variables are proxies, not measured deposition mass or Joules.
6. **Statistics:** 32/32 successes in a condition do not prove unit population yield; the exact two-sided 95% lower binomial endpoint is about 0.891. Conditions sharing seeds are paired. Phase sweeps reuse arrays across reserves.
7. **Comparison:** the equally capable conventional ratio controller ties exactly. The fixed-scale controller solves a stricter objective. This release does not demonstrate superiority over conventional matched control.
8. **Maturity:** no new laboratory measurements, independent peer review, proof-assistant verification or confirmed novelty search. Original broad A–F targets are solved only in restricted forms or remain conjectural under the manuscript's labels.

The most important unresolved physical obstacle is bounded differential-bias paired metrology together with reproducible bounded post-conversion actuation. The strongest negative controls, failed runs and earlier baselines are included.