$ cat research/decision-layer.md
Work on the class of models that answer questions instead of writing prose — and on how to verify whether their stated confidence is worth anything.
Models that answer bounded questions with probabilities instead of generating prose — and how to verify that their confidence means something.
open S.A.C.I. →Pre-registered bars, anti-contamination, honest repetition and measurements that block a run instead of describing it.
open the benchmark →Hybrid quantum-classical architectures, measured against a classical baseline on the same data and metrics.
open the paper →A generative model produces text, and the text is the interface — which means every downstream system has to parse prose and has no principled way to know when the model was guessing. A decision layer inverts that: it takes a state plus a set of typed questions with bounded answers and returns typed answers with probabilities.
The value is not accuracy alone, it is calibration: a stated 0.8 must actually be right about 80% of the time. When that holds, application code can route on confidence thresholds, escalate the uncertain cases, and abstain instead of guessing — which is precisely what an autonomous pipeline needs.
This is the research axis behind the distributed agent platform: a decision layer is what lets autonomous code act on model output without a human in the loop.
| system | task | accuracy | ECE | verdict |
|---|---|---|---|---|
| decision head v12 · tier | hardware-tier routing (4 classes) | 79.4% | 0.0415 | pass |
| decision head v12 · intent | user intent (6 classes) | 74.3% | 0.0738 | pass |
| 4B instruction model | jevbench sample · easy split — zero-shot baseline, no calibration | 67.0% | — | baseline |
| encoder classifier | jevbench sample · easy split — zero-shot baseline, no calibration | 44.3% | — | baseline |
| decision layer v1 | trading-action gate — calibration bar was ECE ≤ 0.08 — rejected, published as-is | 41.5% | 0.3427 | rejected |
Neural Network Analysis for CRISPR: A Quantum Computing Approach
25th Symposium on High Performance Computing Systems (SSCAD 2024) · Portuguese
Predicted CRISPR gene dependency from gene copy number with a classical and two hybrid quantum-classical neural networks, and assessed whether hybrid quantum neural networks are viable for regression.
This method runs in production as S.A.C.I. — typed questions, calibrated answers, confidence bands, and the measured results including the rejected ones.
Method write-ups or benchmark details on request — marcusnggg@gmail.com.