Aionoscope
Summary
Aionoscope is the generator and benchmark ecosystem introduced by Aionoscope: Debugging Latent-State Accessibility in Time-Series Representations. It generates synthetic time-series streams from explicit latent process state, renders observations through views, and evaluates whether frozen model representations expose the categorical and dense process variables under a common probe protocol.
The current public MILETS 2026 snapshot uses Primitive Process Mixtures: single-channel synthetic streams with exact component labels and dense parameters for timing, phase, amplitude, frequency, trend, event, and regime-style variables.
Artifact Split
- Generator/library: https://github.com/langotime/aionoscope/ provides the PyTorch-native Process-to-View generation library.
- Benchmark harness: https://github.com/langotime/aionoscope-benchmarks/ runs frozen-feature offline probes and hosts benchmark artifacts.
- Interactive dashboard: https://aionoscope.langotime.ai/ visualizes current model, target, layer, and metric views.
- Preprint: https://arxiv.org/abs/2607.00956
- Raw paper Markdown: paper_aionoscope-2026.md
Benchmark Contract
Aionoscope separates three layers:
- Process: samples the true latent state, including active components, regimes, events, frequencies, amplitudes, phases, timing, slopes, offsets, and masks.
- View: renders the latent state into an observed time series through mixing, noise, nuisance variation, and sampling choices.
- Probe harness: extracts frozen model-plus-adapter features, pools by layer, fits categorical and masked dense linear probes, and scores against exact labels.
Role In The Wiki
Aionoscope is the local benchmark object for latent-state accessibility. It should be used when a source claims that a time-series representation preserves state, not only shape or label information.
Its strongest wiki lesson is diagnostic: coarse component identity can be linearly accessible while dense state variables remain weakly exposed. This directly supports the broader Latent-State Time-Series Modeling and Time-Series Benchmark Hygiene pages.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in aionoscope-2026 rather than duplicating verdict rows here.
At the entity level, Aionoscope is an evaluation instrument. It does not itself solve latent-state modeling, but it gives the wiki a concrete test for whether a candidate encoder exposes process state under controlled labels.
Caveats
- Current results are a public validation-seed pilot, not a hidden leaderboard.
- Primitive Process Mixtures is synthetic and single-channel.
- The common readout is mean-pooled and linear; failure under this probe is readout-limited.
- Native input length and adapter choice are part of each model fingerprint.
- No action-conditioned control, intervention, or deployment benchmark is included yet.