A precise map of the plants, problem structures, and domains TrueLoop Compute drives — separated cleanly into what we have demonstrated in controlled simulation and what we project from those results. No claim here rests on a regime we have not run.
TrueLoop Compute is a model-free, retained-state feedback runtime. It sits between your application and a physical or parametric substrate and closes the adaptive loop: each round it reads one measured statistic, compares it to where you want to be, and returns the next configuration — driving on the measured residual itself, not on an estimated gradient or a model of your device. Because the state lives in the configuration, it carries forward across related problems instead of restarting. It discovers each channel’s drive direction from your measurements, so it adapts to the actual sign and shape of your plant rather than assuming one.
That design defines exactly where it wins — and where it does not. The requirement is a locally monotonic, writable loop: a configuration you can set, a measurement you can read, and a response that moves consistently in one direction as you move each control. Below is everything that satisfies that, and the honest boundaries that do not.
Each of these was run through the live runtime in regulation mode across multiple seeds. The runtime learns each channel’s sensitivity sign from a short measured probe, so polarity, gain, and mild coupling are handled automatically.
| Plant type | Status | Notes |
|---|---|---|
| Phase–power (cos²) p = ½(1 + cos φ) | Validated | The canonical interferometric transfer. Converges across the full output range including extreme targets. |
| Negative-sine m = −½ sin φ | Validated | Includes targets that require crossing φ = 0 into the negative branch. The sign-discovery probe handles this. |
| Positive-sine m = +½ sin φ | Validated | Opposite polarity to the above; learned from measurement, no configuration needed. |
| Linear, positive gain m = kφ, k > 0 | Validated | Any positive gain, including small (0.3×) and large (5×). |
| Linear, negative gain m = kφ, k < 0 | Validated | Negative gain is driven correctly via the measured sign — a case fixed-direction loops get wrong. |
| Saturating (tanh) m = tanh φ | Validated | Smooth saturating response; converges within the monotonic region. |
| Coupled / crosstalk outputs share inputs | Validated | Mild channel coupling (neighbour crosstalk) is absorbed by the per-channel residual drive. |
| Mixed sign per channel polarity varies by channel | Validated | Each channel discovers its own direction independently in the same session. |
| Drifting / time-varying target or plant drifts | Validated | The core strength: retained momentum tracks a moving target under one read per round. |
| Fold-back / periodic many φ → same output | Out of scope | A probe measures a local slope; a globally ambiguous response is genuinely out of envelope and trips the saturation flag. |
The runtime drives objectives that live in the measured marginals. When the value of a problem is carried in those marginals it fits; when it lives in correlations the marginals cannot see, it does not.
QUBO, resource allocation, feature selection, threshold setting — cost reads directly from the measured bits.
Routing on local graphs, scheduling, MPC plans, beamforming — local structure the marginals resolve.
Calibration, sensor fusion, drift correction, setpoint holding — the most-validated regime.
Attention allocation, clustering, graph partitioning, inference paths — value lives in pairwise structure marginal feedback cannot reach.
Tested thoroughly; the objective sits in two-qubit correlations the loop cannot observe. We do not sell it here.
Problems whose optimum depends on a global low-rank structure rather than local readouts.
| Domain | Status | What the runtime does there |
|---|---|---|
| Photonic mesh calibration MZI phase locks, drift | Validated | Holds per-tap power at setpoint under thermal drift, one detector pass per round. Self-configuration is mature in the field; our edge is the retained-state tracking. |
| Drifting sensors / NDT ultrasonic, GPR, arrays | Validated | Continuous residual correction holds calibration as the instrument drifts — the measurement-limited regime it is built for. |
| Variational quantum (calibration) gate-angle, readout point | Projected | Holding a single-rotation observable at setpoint under drift. Control, not variational solving. |
| RF / microwave tuning phase shifters, match nets | Projected | Local, monotonic, writable — fits the envelope; projected from the validated control results. |
| AI inner-loop decisions routing, scheduling, MPC | Projected | The fast decision layer beneath perception, warm-started across related decisions; projected from the deadline and warm-start results. |
| Variational energy minimization VQE / QAOA solving | Out of scope | Structurally excluded — see problem structure above. |
Two results are demonstrated in benchmarks: the sub-quadratic convergence cost (measured ~n1.3, 95% CI [0.9, 1.5], over n = 12–48) and the deadline crossover, where classical search collapses to near-random while the runtime still returns a usable solution. From these we project — we do not yet claim — that the feasibility advantage extends to larger problems under tight deadlines, within the structural envelope above.
That projection rests on two hypotheses we state openly: that the exponent holds beyond the sizes we can brute-force-verify, and that hardware delivers a favourable optical-to-digital cost ratio. The deadline timing model is the one load-bearing assumption still awaiting on-hardware confirmation; a hardware pilot is the experiment that converts it from projected to demonstrated. We would rather tell you that than overstate it.
The fit finder on the homepage classifies your problem in a minute, and the client returns DECLINE when a conventional method already wins. Honesty about the boundary is the product.