Overview
Sparse waypoint analysis is privacy-sensitive: it must separate movement from irregular reporting, missingness, spatial coarsening, and corruption while preserving uncertainty about hidden location. Active FractalRabbit provides a controlled, artifact-bound benchmark whose headline lane uses a deterministic project-local synthetic FractalRabbit-format fixture; a separately retained lane exercises pinned open-source software from the National Security Agency as an independent simulator surface. The benchmark converts sporadic reports into categorical evidence, fits explicit hidden-state generative models, and compares transparent temporal, Markov, sequence, state-space, neural, latent-state, and active inference predictors under matched information sets. Under noisy partial-observability, Active Inference is the lowest-loss implemented predictor: it clearly leads point-estimate and raw-observation families and sits in a statistical tie with the strongest non-AIF belief-preserving comparator. The shared mechanism is soft Bayesian marginalization, which preserves probability across plausible cells instead of committing early to one state. Point estimates suffice for clean observations, an online base-rate predictor leads under regime switching, transparent temporal and disclosed kinematic controls anchor sparse reporting gaps, and withholding location sharply limits specific-cell recovery from metadata. The partially observable Markov decision process (POMDP) formulation also exposes variational and expected-free-energy diagnostics for belief, minimization, and integrity. These results establish a regime-specific synthetic model map and a reproducible evidence chain. The present contract covers synthetic software behavior; separate evidence protocols govern privacy and empirical evaluation. Code, fixtures, manuscript source, and the release manifest are public at github.com/ActiveInferenceInstitute/active_fractal_rabbit (https://github.com/ActiveInferenceInstitute/active_fractal_rabbit).
---
Associated artifacts
GitHub release: v0.2.0 (https://github.com/ActiveInferenceInstitute/active_fractal_rabbit/releases/tag/v0.2.0)
DOI: https://doi.org/10.5281/zenodo.21330636
Zenodo: https://zenodo.org/records/21330636
PDF SHA-256: d676159d149a12a1e990457329ad4cf52d2d8ab4ffcf09f55b42bad8e0c3c052
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- On the synthetic fixture, soft Bayesian marginalization beats a hard point estimate as emissions degrade: 0 nats at a clean channel to 0.739 nats at flip probability 0.400 on the primary draw, with a mean of 0.495 nats.
- Under noisy emission, active inference ranks 1 of 14 but leads the strongest non-AIF belief filter by only 0.005 nats, a statistical tie.
- No directional comparison in the clustered-bootstrap family survives multiplicity correction.
- Withholding location collapses the filter's mean mass on the true cell from 0.995 to 0.069.
- The pymdp lane matches the exact Bayes posterior (max abs. difference 3.99e-07) and contributes interpretability rather than predictive leadership on the default surface.
Methods
- Synthetic waypoint benchmark: FractalRabbit fixture plus NSA simulator lane — The headline lane uses a deterministic synthetic FractalRabbit-format fixture; a separate lane runs the pinned open-source NSA FractalRabbit software.
- Discrete POMDP generative model over grid cells with four categorical modalities — Waypoints are discretised into a 4x4 grid of hidden cells with location, time-of-day, speed and reporting-burst observation tensors.
- Matched-information comparison of 14 predictors incl. HMM, particle, pymdp — Temporal, Markov, sequence, state-space, neural, latent-state and active inference predictors are scored on held-out next-cell log-loss under matched information sets.
- Soft-vs-hard marginalization protocol under a K-ary symmetric noisy emission — Soft belief, hard MAP and observed-Markov predictors share transition, emission and priors, differing only in marginalize-versus-commit, across emission flip probabilities.
- Traveller-clustered bootstrap with Holm and Benjamini-Hochberg correction — Paired loss differences use traveller-clustered bootstrap intervals (B = 2000) with Holm-Bonferroni primary and BH secondary correction.
Summary sources: Paper metadata and evidence · Extracted source text.
PDF downloads
Archived files available directly from this site.
- Friedman_2026_Active_d676159d.pdf PDF · 14.25 MiB · extracted-text source
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related in Cognitive Security
Other catalogued works in the same domain.