Preflight

Exactly budgeted local acquisitions before a maintained hardware run.

A hosted server cannot call a measure() function that lives beside your instrument. Client 0.6.0 therefore orchestrates the handshake: the endpoint returns a safe configuration, your process applies it and acquires one complete response, and the client posts that response before requesting the next configuration.

The default budget is seven acquisitions: six response checks plus one timed return-to-x0 read. That final read separates readout noise from drift and supplies the session recommendations. Preflight setup is reported separately from controller cold start and does not debit the hosted runtime key. Pass measure_drift=False only when reproducing the earlier six-acquisition behavior.

The measured diagnostic compares response motion with caller-declared minimum resolution and measured repeatability. A structural registry DECLINE or an out-of-contract local check can decline preflight. A local FIT means every bounded measurement check passed on this plant; evidence_backed_fit separately reports whether a matching positive registry cell supports it.

Preflight does not infer target feasibility from configuration bounds alone, estimate an unknown plant from nothing, or apply a mesh-specific drift number to every substrate.

Preflight also asks what will consume the maintained observable, because the plant can be perfectly maintainable and the application still see little gain. Measured across three campaigns: numeric consumers (an energy, a force, a spectrum, an expectation value) gained +34.0 points on hardware; argmax consumers, which keep only a ranking or best index, gained +0.7; a first-order variational loop re-tunes its parameters each iteration and absorbs coherent drift as a gauge; an optimiser that averages many gradient estimates (SPSA, Adam) converges regardless. Declare consumer as numeric, argmax, variational, or averaged. The last three return an advisory at session start: the substrate will be held in spec, and the application-level gain should be expected to be modest. Where a variational workload also produces a number, that number is where the layer pays.

layer = StabilityContract(key, n=len(target), target=target, tolerance=0.05)
report = layer.preflight(measure, x0)
x = layer.start(x0)
for _ in range(rounds):
    x = layer.step(measure(x))
proof = layer.end()