Audited (legacy) 2026

Persistent-Noise Multi-View Fusion

Implementation & benchmark of arXiv:2608.24819 — Reliability Limits and Decoding for Partial Nanopore Protein Rereads With Persistent State

Usefulness6/10

Difficulty4/10

Novelty7/10

Training Accuracy Stability Generalization Loss Training Regularization Probability Information theory Statistics

Source paper: Reliability Limits and Decoding for Partial Nanopore Protein Rereads With Persistent State arXiv:2608.24819 ⓘ · analyzed Aug 29, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Train a classifier or encoder to distinguish shared latent corruption from fresh per-view noise instead of treating repeated observations as conditionally independent given the target. A single persistent state corrupts all views, while each view may additionally receive independent observation noise; the fusion loss marginalizes the persistent state exactly. This should reduce overconfident predictions from repeated but systematically biased augmentations, sensor readings, or retrieved documents.

Formulas

$$T_{xu}(q,R)=(1-q)\mathbf{1}\{u=x\}+qR_{xu},\qquad q\in[0,1],\quad R_{xx}=0,\quad \sum_u R_{xu}=1.$$

$$p(y_{1:K}\mid x)=\sum_{u\in\mathcal U}T_{xu}(q,R)\prod_{k=1}^{K}S_{uy_k},$$

$$p(x\mid y_{1:K})=\frac{\pi_x\sum_{u\in\mathcal U}T_{xu}(q,R)\prod_{k=1}^{K}S_{uy_k}}{\sum_{x'\in\mathcal X}\pi_{x'}\sum_{u\in\mathcal U}T_{x'u}(q,R)\prod_{k=1}^{K}S_{uy_k}},$$

$$p_{\mathrm{ind}}(y_{1:K}\mid x)=\prod_{k=1}^{K}\left(\sum_{u\in\mathcal U}T_{xu}(q,R)S_{uy_k}\right)\neq p(y_{1:K}\mid x).$$

Mathematical statement

The paper defines a finite-alphabet persistent channel with clean symbol X in alphabet $\mathcal X$, persistent intermediate symbol U in alphabet $\mathcal U$, and transition matrix T. The key equation is $T_{xu}(q,R)=(1-q)\mathbf{1}\{u=x\}+qR_{xu}$, where $q\in[0,1]$ is the probability of a persistent deviation, R is a fixed row-stochastic matrix with $R_{xx}=0$, and $\mathbf{1}\{u=x\}$ is one when the persistent state is unchanged. For a neural-network adaptation, let x be the clean class or latent token, u be a shared corruption state sampled once per example, and $y_{1:K}$ be K observations generated conditionally independently from u through an observation matrix S, where $S_{uy_k}=p(y_k\mid u)$. The correct likelihood is $p(y_{1:K}\mid x)=\sum_u T_{xu}\prod_k S_{uy_k}$, not $\prod_k\sum_u T_{xu}S_{uy_k}$. With prior $\pi_x=p(x)$, the exact posterior is $p(x\mid y_{1:K})=\frac{\pi_x\sum_u T_{xu}\prod_k S_{uy_k}}{\sum_{x'}\pi_{x'}\sum_u T_{x'u}\prod_k S_{uy_k}}$. The independent-view expression redraws the persistent state for every observation and can therefore accumulate spurious evidence about a systematic error as K grows.

Implementation notes

Integrate this at the data-loader and fusion-loss level for a small multi-view classifier, or at the token-evidence aggregation level of a transformer. For each training example with clean target x, create one persistent state u using T: with probability 1-q set u=x; otherwise sample u from row x of R. Generate K observations y_k from u using an observation corruption operator S. For images, S can be a class-confusion channel applied after independent image augmentations; for discrete tokens or sensor categories, S can be a lookup-table sampler. Feed all K views through a shared encoder and either concatenate their representations or average their per-view logits. Compute the persistent target distribution with log-sum-exp. For every candidate class x, calculate a_x = logsumexp over u of [log T[x,u] plus the sum over k of log S[u,y_k]], then set target_x = softmax(log pi[x] + a_x). Minimize cross-entropy between the network output and this target, optionally adding the ordinary clean-label loss. As a control, train with the independence likelihood, which multiplies the per-view marginal after effectively redrawing u for every view. Initially set q and R from the synthetic corruption generator; in a second experiment, learn q through a sigmoid parameter and learn R through row-wise softmax, with S fixed. The first cheap experiment is MNIST or CIFAR-10 with a fixed class-confusion matrix R, q in {0.1, 0.3, 0.5}, and K in {1, 2, 4, 8}. Compare ordinary mean-logit fusion, independent-marginal training, and persistent-state marginal training using the same small CNN and augmentations. Measure accuracy, true-label NLL, expected calibration error, and confidence growth with K on systematically corrupted examples. Success means lower NLL and calibration error at equal accuracy, with confidence saturating rather than becoming unjustifiably extreme as K increases.

Verification

Audited (legacy)

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not tested

Stage 2 · Practical benchmark result: Not run

Methodology: Toy-system gate first; the benchmark stage runs only after a pass. How verification works

Stage 1 — Mechanism check agent confidence 9/10

Built exact persistent-state and independent-view fusion with stable log-sum-exp computation, vectorized synthetic sampling, and numerical identity checks. The core formulas matched direct computation with maximum error 3.6e-15, while the persistent model maintained low ECE (0.002-0.009) and NLL 0.868 at K=16 versus independent fusion's ECE 0.347 and NLL 4.774; accuracy was similar (0.652 vs 0.652). This is a clear calibration/NLL win under systematic shared corruption, not an accuracy improvement.

Agent confidence: 9/10

Baseline: Independent persistent-state redraw: K=16 NLL 4.774, accuracy 65.21%, ECE 0.3473, mean confidence 0.9993.
Idea: Exact persistent-state marginalization: K=16 NLL 0.868, accuracy 65.23%, ECE 0.00225, mean confidence 0.6500; at K=8, NLL 0.869 and ECE 0.00679 versus baseline NLL 2.420 and ECE 0.3270.

Limitations: Only a finite-alphabet synthetic channel was tested with known T and S; no learned CNN/encoder, trainable q or R, image augmentations, FLOP/speed comparison, or real dataset evaluation was included. The independent control is the mathematically misspecified redraw model rather than a neural mean-logit fusion baseline.

How to run: python3 experiment.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Super-audit (legacy) Aug 30, 2026 latest Worked ✓

Agent confidence: 9/10

The harder five-seed, variable-strength shared-corruption evaluation reproduced a clear calibration and NLL benefit for exact persistent-state marginalization. At K=16, persistent fusion achieved ECE 0.0147 +/- 0.0043 and NLL 0.9465 +/- 0.0081 versus independent redraw ECE 0.3866 +/- 0.0085 and NLL 3.7911 +/- 0.0653, with similar accuracy (0.5607 versus 0.5541). The no-shared-corruption ablation made persistent and independent fusion identical, confirming that the effect depends on shared corruption; this validates the fusion mechanism but not yet an end-to-end learned neural implementation.

Baseline independent K16: NLL 3.7911 +/- 0.0653, accuracy 0.5541 +/- 0.0086, ECE 0.3866 +/- 0.0085, mean confidence 0.9402 +/- 0.0013.
Baseline independent K4: NLL 1.4086 +/- 0.0310, accuracy 0.5276 +/- 0.0104, ECE 0.2296 +/- 0.0105, mean confidence 0.7556 +/- 0.0033.
Baseline independent K8: NLL 2.0854 +/- 0.0788, accuracy 0.5458 +/- 0.0095, ECE 0.3281 +/- 0.0093, mean confidence 0.8736 +/- 0.0026.
Data: 8-class harder synthetic channel; 3,500 examples per seed/K; q uniformly sampled from [0.15, 0.80], observation noise uniformly sampled from [0.10, 0.32], seeds 11, 22, 33, 44, 55; total runtime 14.49 seconds.
Idea persistent K16: NLL 0.9465 +/- 0.0081, accuracy 0.5607 +/- 0.0090, ECE 0.0147 +/- 0.0043, mean confidence 0.5613 +/- 0.0023.
Idea persistent K4: NLL 1.0211 +/- 0.0084, accuracy 0.5407 +/- 0.0072, ECE 0.0150 +/- 0.0038, mean confidence 0.5411 +/- 0.0032.
Idea persistent K8: NLL 0.9450 +/- 0.0110, accuracy 0.5578 +/- 0.0098, ECE 0.0195 +/- 0.0030, mean confidence 0.5615 +/- 0.0009.
Math checks: K=1 identity maximum error 0.0; posterior normalization maximum error 8.88e-16; strong-shared ECE reduction at K=16 is 96.2%.
No shared K16: Persistent and independent both NLL 0.000095 +/- 0.000180, accuracy 0.9999 +/- 0.0003, ECE 0.000072 +/- 0.000129.
Second baseline mean logit K16: NLL 1.2463 +/- 0.0058, accuracy 0.5541 +/- 0.0086, ECE 0.1782 +/- 0.0078.

Limitations:

The super-check uses a harder synthetic finite-alphabet channel rather than real images or a real sensor/document dataset. It evaluates exact known-channel fusion and a mean-logit control, but does not implement a learned CNN/encoder, trainable q or R, image augmentations, FLOP comparisons, or end-to-end training. The required supervisor protocol was fully covered with five seeds, K=4/8/16, variable shared corruption, independent observation noise, two baselines, and the no-shared ablation.

How to run: python3 supercheck.py

Artifacts

Implementation overview ⬇ Download all as ZIP 7 files · code, reports and structured results

🐍 experiment.py 3.9 KB View Raw
📄 report.md 1.4 KB View
📄 report_super_2026-08-30T102502.md 3.9 KB View
📊 results.json 2.8 KB View Raw
🐍 supercheck.py 4.5 KB View Raw
📄 supercheck_report.md 4.1 KB View
📊 supercheck_results.json 32.4 KB View Raw