Every number TrueLoop publishes is tiered by evidence and travels with the code that produced it. This page is the ladder: what has been demonstrated on live hardware, what is established in controlled simulation, what is a projection, and where the runtime loses. The losing rows are published with the same care as the winning ones, because a regime map with no red cells is marketing.
107-channel closed-loop regulation, Rigetti Cepheus-1-108Q. The runtime held the excitation probabilities of 107 physical qubits to a hidden measured target under disclosed injected drift, converging in twelve parallel 150-shot acquisitions, in both of two runs. Finite differences at the identical measurement budget probed 11 of the 108 evaluations its first gradient requires: zero completed updates, no improvement. From billing metadata: one runtime update cost one task, about 20 credits and 100 ms of machine execution; one finite-difference update costs 108 tasks, about 2,180 credits and 10.8 s; the cloud loop carried a median 48.9 s of wall time per cycle, a measured 489× infrastructure-over-physics overhead.
Retained-versus-reset ablation, IQM Emerald, pre-registered. Two controllers identical in every component except that one carries the configuration between cycles and the other re-draws it each cycle. Five hypotheses were committed before any hardware submission; five passed. Pooled reset-to-retained error ratio: 7.41 against a committed threshold of 2.0. Retention is the mechanism, isolated.
One million channels. The unmodified production runtime holds 106 simulated coupled, drifting channels flat, locks within seven cycles, and recovers from a full configuration scramble in one cycle. The deployed dither practice, granted oracle-exact corrections, wins at small scale (published) and fails between 105 and 106 channels by staleness arithmetic. The direction law. Per-channel readout converges at every tested size to 65,536 channels whether the detector is signed or unsigned, so plain power detection suffices; a broadcast scalar is pinned at the do-nothing level. The runtime’s drift-rate frontier is flat from 64 to 32,768 channels; SPSA’s frontier is zero by n=512. Structured QUBOs. Fed per-variable structure signals, quality reaches 0.95 (constraint) and 0.84 (parity decoding) at n=4096, flat with scale. Variational shot budgets. At a budget of twelve acquisitions the runtime converges to 1.5× the sampling floor while finite differences completes zero updates; at sixteen times the budget the best tuned incumbent is still five times the floor.
Scalar-score steering at scale starves by a provable information ceiling, and we publish the decay curve. Unstructured spin glasses trap at 0.70 of the free-digital ceiling. Dense arbitrary QUBOs pay a measured quadratic feedback cost. Landscapes a digital computer evaluates for free belong to digital solvers. In-solve stabilization of annealing solvers is measurably harmful. Multiplexed dither beats the runtime at small channel counts, and the small-scale cells where incumbents win are pre-registered requirements of our own experiment designs: if the deployed practice did not win its home ground, the model would be rigged.
Each repository ships the plant models, gold-standard baselines tuned on calibration seeds disjoint from analysis seeds, pre-stated verification bars, per-cell checkpointed result JSONs, and cells runnable to n=4096 on a free evaluation key. The runtime is driven strictly as a hosted black box; no update law or internals appear in any repository.
TrueLoop Compute: Reproducible Headline Results →
Variational parameters under shot budgets · photonic mesh hold to one million simulated channels · phased-array calibration in dB
TrueLoop QUBO Reproduction →
The solvability law regime map: the feedback channel swept in size · the structured solve regime · the boundary rows
All repositories →
A full manuscript formalizes the framework: the stateful-wave-computing class, a two-sided direction law (per-channel feedback converges in a cycle count independent of channel count under stated assumptions; scalar ranking feedback carries a provable alignment ceiling), the measured estimation tax and elimination order, both hardware campaigns, and a claims ledger with every pre-registered bar and every miss. It is prepared for journal submission; a preprint will be linked here on release. For the results themselves, see the headline results doc or run the repositories above.