Between December 2025 and February 2026 I filed my inventions as Swiss patent applications. Technology is moving faster than patents can keep up, and my own work has moved well beyond them, so I am publishing them in full as research papers instead. Each one below links to its permanent copy.
Compound Learning with Cross-Model Transfer in Machine Learning Systems
A way for what a system learns from its own predictions to carry over when the model, the deployment or the instructions change.
Abstract
A machine learning system in production builds up a record of where its predictions were right and where they were wrong. Most systems lose that learning whenever the model is replaced, moved to another deployment or run with a different system prompt. Learning tied to one model, one deployment or one configuration has to be gathered again from nothing each time.
This paper describes compound learning. Prediction-outcome pairs accumulated from production inference are turned into learned patterns (calibration parameters, task classifications and model capability profiles), stored in a provider-independent form, and applied across AI model providers, across deployment types (cloud, local and edge), across reasoning perspectives (different system prompts and settings), across task types and across languages. The paper also covers fine-tuning on accumulated pairs; a privacy-preserving mode in which pairs never leave local hardware and only abstracted patterns are shared, through aggregation, differential privacy, embedding projection or federated aggregation; safeguards for continual learning; parameter-efficient methods such as LoRA; and multi-tenant operation.
Three results were measured. Fine-tuning a cloud-hosted model on 1,033 pairs selected from 9,254 accumulated pairs reduced its mean absolute error by 78.3% against the base model. Fine-tuning a model on a local Apple M1 machine on 612 pairs, kept apart from the cloud data, reduced its mean absolute error by 92.5%. Patterns learned under one reasoning perspective held under four maximally different perspectives (analytical, creative, adversarial and collaborative) in all 160 directed tests, 120 of them across perspectives (exact 95% lower bound 97.0% for those 120). Transfer of the patterns across model providers, task types and languages is expected and not yet measured; related measured results are in the author's first two papers (doi:10.13140/RG.2.2.11961.12645 and doi:10.13140/RG.2.2.12380.55688).
This is the full technical text of Swiss patent application CH000058/2026 (filed 7 January 2026; first filed as CH001465/2025 on 7 December 2025), published as research rather than pursued as a patent. Passages written only for the patent examiner have been removed, and every figure is marked as measured, expected, a design bound or a design target.
- DOI
- 10.13140/RG.2.2.31674.35520
- License
- CC BY 4.0
- How to cite
- Tom Killi, “Compound Learning with Cross-Model Transfer in Machine Learning Systems,” preprint, ResearchGate, October 31, 2026, https://doi.org/10.13140/RG.2.2.31674.35520.
Doubt-Based Model Routing with Self-Calibrating Factor Weights and Cross-Deployment Transfer in Artificial Intelligence Systems
A system that chooses which AI model gets each task before any model runs, and learns from how tasks actually turn out which of its warning signs to trust.
Abstract
Systems that route work between AI models of different cost and capability usually decide too late or learn too little. Cascades such as FrugalGPT send each task to a small model first and escalate only after judging its output, so the first inference is paid for even when it is thrown away. Learned routers such as RouteLLM decide in advance, but are trained offline and do not adapt to outcomes in production. Calibration methods such as temperature scaling adjust one global parameter and cannot tell which source of error to correct.
This paper describes a router that decides before any model runs. For each task it computes a doubt score, a predicted probability of failure between 0 and 1, as a weighted combination of four independent factors: the task's complexity, its novelty relative to past tasks, past performance on similar tasks, and the current state of the system. The score selects a model tier from any number of tiers, across cloud-hosted, locally deployed and edge models. When each outcome is known, the accuracy of every factor is tracked separately and the weights are recalibrated by inverse mean absolute error, so that factors that predict well gain influence and factors that predict badly lose it, continuously and without retraining or labelled data. The paper also covers factor isolation testing with synthetic tasks, shadow execution, transfer of calibrated weights to new models and deployments, routing in multi-step agent tasks, and extensions for surfacing doubt to human reviewers, per-category calibration, historical counterfactual insights and feedback-loop detection.
Two results were measured. Weights calibrated on cloud models and transferred to eight locally deployed model configurations, tested on a held-out set of 500 tasks, reached between 89.3% and 238.8% of the source routing accuracy; the mean of the six middle values was 97.7%, and seven of the eight reached at least 95%. Over 50,000 production tasks, the routing decision took a median of 0.47 ms (99th percentile 1.08 ms). The reduction in prediction error from factor-specific calibration, and transfer of the weights across model providers, are expected and not yet measured; related measured results are in the author's first paper (doi:10.13140/RG.2.2.11961.12645).
This is the full technical text of Swiss patent application CH000112/2026 (filed January 2026; first filed as CH001466/2025 on 7 December 2025), published as research rather than pursued as a patent. Passages written only for the patent examiner have been removed, statements about related systems have been checked against their published papers, and every figure is marked as measured, expected, a design bound or a validation target.
- DOI
- 10.13140/RG.2.2.12380.55688
- License
- CC BY 4.0
- How to cite
- Tom Killi, “Doubt-Based Model Routing with Self-Calibrating Factor Weights and Cross-Deployment Transfer in Artificial Intelligence Systems,” preprint, ResearchGate, October 31, 2026, https://doi.org/10.13140/RG.2.2.12380.55688.
Architecturally Enforced Prediction-Outcome Capture in Machine Learning Systems
A system that learns from its own mistakes has to record every prediction and how it turned out; this design builds that record into the system, so it is no longer optional.
Abstract
A machine learning system can only learn from the predictions it can link to their outcomes. In most production systems that link is made by logging, and logging is optional: it can be switched off in configuration, bypassed by calling the underlying function directly, forgotten in a hurry, or fail silently while the system carries on. Every prediction left without its outcome is a lesson the system never gets.
This paper describes an architecture in which capture is structural rather than optional. An enforcement layer replaces the target function in the runtime namespace at the moment the function is defined, so that calling the function by name necessarily passes through capture. The prediction is persisted before execution may begin; the outcome is captured in an exception-safe finalization block whether execution succeeds or fails; and the two are linked by a unique identifier carried in thread-local or task-local context. The paper sets out a taxonomy of enforcement mechanisms (function, request, event, compile-time, infrastructure, data and hybrid), and covers streaming and batch inference, delayed outcomes and an advisory mode.
In a production environment, 9,254 of 9,280 predictions were paired with their outcomes: 99.7% (95% confidence interval 99.58% to 99.82%). Measured by deployment type, pairing was 99.9% for cloud-hosted models, 99.6% for locally deployed models, 99.3% at the edge and 99.2% for hybrid deployments. Conventional approaches are estimated to pair 62% to 76%. Fine-tuning a model on 1,033 of the captured pairs reduced its mean absolute error by 78.3% against the base model.
This is the full technical text of Swiss patent application CH000033/2026 (filed 8 January 2026; first filed as CH001468/2025 on 7 December 2025), published as research rather than pursued as a patent. Passages written only for the patent examiner have been removed, and every figure is marked as measured, estimated or expected.
- DOI
- 10.13140/RG.2.2.11961.12645
- License
- CC BY 4.0
- How to cite
- Tom Killi, “Architecturally Enforced Prediction-Outcome Capture in Machine Learning Systems,” preprint, ResearchGate, October 31, 2026, https://doi.org/10.13140/RG.2.2.11961.12645.