Zeno Divergent

A small model that predicts when other models are about to be wrong.

TL;DR

Forecasting models fail quietly. They keep returning confident numbers long after the world they were fitted to has moved on.

Zeno Divergent watches a stream of public observations and predicts the moment your expectation of that stream is about to break — a regime change, a cadence shift, a jump into a magnitude band the reference never expected.

It outputs calibrated probabilities, not anomaly scores.
It says "I don't know" out loud (explicit abstention gate) instead of guessing.
It is tiny (~31k params, CPU, milliseconds) and sits next to your existing model.
It is trained and evaluated on real public data, against rule baselines, with bootstrap confidence intervals.

pip install zeno-divergent

from transformers import AutoModel
import zeno_divergent                      # registers the architecture

model = AutoModel.from_pretrained("mbarbosa1/zeno-divergent-v1")
rows = model.score_events(observations, entity_id="grid_34.5_-118.5")

rows[-1].surprise_index    # 0..1 ranking of "expectation is breaking here"
rows[-1].probabilities     # calibrated probability per declared target
rows[-1].abstained         # True -> no call, not "all clear"

Why this matters

Every deployed forecaster carries an invisible assumption: tomorrow looks enough like the data I was fitted on. When that assumption dies, the model does not raise its hand. Accuracy degrades, uncertainty stays flat, and the failure is only discovered downstream.

The usual answers are partial:

Approach	What it tells you	What it misses
Predictive uncertainty (`σ`, ensembles)	how noisy this prediction is	stays low right up to a regime break
Drift / OOD detectors	the inputs moved	fires on harmless movement, silent on harmful stable-looking shifts
Change-point detection	something changed already	it is a post-mortem, not a lead time
Zeno Divergent	the probability that an observable departure happens in the next window, calibrated, with an option to abstain	it is not a hazard forecast — see limits

The falsifiable claim this model exists to test:

A learned representation of model-relative surprise contains information about future regime change and model failure that is not captured by predictive uncertainty or rule baselines alone.

The evaluation harness is built to reject that claim: every target is scored against a phase-0 rule baseline with entity-level bootstrap CIs, and a promotion gate refuses to move main unless the corpus is real and the lifts hold.

How it works

observations ──▶ schema contract ──▶ past-only features ──▶ shared encoder
                                                                  │
                                                                  ▼
                                                         surprise state z^S
                                                                  │
                             ┌────────────────────────────────────┼───────────────┐
                             ▼                                    ▼               ▼
                  calibrated p_k(t+Δ)                     uncertainty σ_k     abstain gate

Declare a schema. Categories, numeric channels, targets and a horizon. The schema is a contract, not a guess — it is hashed into the config.
Build past-only features. Row i is a function of observations 0..i and nothing after it. Leakage is re-derived from the raw stream, not asserted.
Encode the surprise state z^S. A 32-dimensional representation of how expectation is failing, shared across all heads.
Predict observable departures. One head per target: category-regime transitions, cadence shifts, magnitude-regime shifts, compound transitions.
Calibrate, then gate. Probabilities are fitted on a held-out validation split; low-evidence rows are marked abstained rather than scored.

What comes back

Field	How to read it
`probabilities`	calibrated probability of each observable transition inside the horizon
`uncertainty`	heteroscedastic spread — large means poorly determined, not safe
`abstained`	admission control. "Not enough evidence", never "nothing to see"
`surprise_index`	max over targets — a ranking, not a probability of harm
`latent`	`z^S` itself, for probing and transfer studies (`include_latent=True`)

Quick start

pip install zeno-divergent

from transformers import AutoModel
import zeno_divergent   # registers config/model/pipeline with the Auto classes

model = AutoModel.from_pretrained("mbarbosa1/zeno-divergent-v1")  # trust_remote_code=False

observations = [
    {"timestamp": 1717171717.0,
     "category": "light_shallow",
     "values": {"magnitude": 3.2, "depth_km": 8.1}},
    # ... chronological, one dict per observation
]

for row in model.score_events(observations, entity_id="region-A"):
    if row.abstained:
        continue
    print(row.timestamp, round(row.surprise_index, 3), row.probabilities)

Categories and channels must match the schema published in config.json. No trust_remote_code, no GPU, no network calls at inference time.

Architecture


architecture	`ZenoSurpriseModel` — shared MLP encoder → `z^S` → per-target calibrated heads
parameters	~30,859 (72 features → 128 hidden → 32 latent → 5 heads + log-variance + gate)
precision	float32, safetensors
domain	`seismic` (schema hash `8753db3bb6f6`)
runtime	CPU, sub-millisecond per row

Why is this smaller than v0.2.3?

It is a different model, not a shrunken one. v0.2.x was a multi-reference encoder over a synthetic ten-mechanism generator with a wide simulated input contract. v0.3 replaces that with a domain-neutral schema contract: the input is a declared public observation stream, features are derived past-only from that schema, and capacity is set by what the real corpus supports without overfitting. Nothing was pruned or quantised. The larger synthetic model stays frozen and loadable at revision v0.2.3.

Training data

Real public data — USGS FDSN event web service (public, credential-free).


events	47107
entities	224
window	2023-01-01 → 2025-07-01
threshold	minimum magnitude 2.5
splits	chronological, with a horizon buffer between splits

No personal data, no identity, no intent, no ownership fields — the input contract rejects them.

Results (held-out test split)

target	prevalence	AUROC (95% CI)	ECE	paired lift vs phase-0 rules
`event_regime_transition`	0.095	0.833 [0.784, 0.878]	0.010	0.202 (significant)
`cadence_shift`	0.504	0.651 [0.629, 0.674]	0.033	0.180 (significant)
`magnitude_regime_shift`	0.000	0.500 [0.500, 0.500]	0.000	-0.034 (not significant)
`significant_event_ahead`	0.407	0.917 [0.887, 0.940]	0.052	0.563 (significant)
`compound_seismic_stress`	—	—	—	not evaluable (insufficient_available_rows)

Abstention. Coverage 0.687; selective accuracy 0.859 vs base 0.763 (gain 0.095) — i.e. the rows the model chooses to answer are the rows it is right about.

Promotion gate: promoted. The full run report ships as report.json, digests in SHA256SUMS. If those digests do not match your rebuild, you are not looking at the same artifact.

Intended use

In scope

Research on model-validity forecasting and surprise representations
Advisory decision support with a named human owner
Prioritising which observation to collect next
Benchmarking change-point and surprise detectors against a locked protocol
Reproducing — or falsifying — the claim above on your own corpus

Out of scope

Automated or safety-critical actuation
Reading the output as the probability of a physical event (it predicts model failure, not hazard)
A single cross-domain risk number
Scoring people. Entities are models, sensors, regions, instruments, processes — never individuals.

Limits

Trained on one real public corpus (seismic). Transfer to other domains is unverified and needs a schema of its own.
Probabilities are calibrated on the validation split; recalibrate before using them as decision thresholds in a new deployment.
Abstention must be surfaced next to the score. Hiding it turns "not observed" into "not at risk" — that is a misuse.
The model predicts observable transitions in a public descriptive stream. It makes no claim about identity, intent, ownership or causation.

Revisions

revision	contents
`main` / `v0.3.1`	v0.3.0 real-corpus checkpoint, transformers-native (this)
`v0.3.0`	same weights, research release layout (`checkpoint.pt` + report)
`v0.2.3`	frozen v0.2 prototype, remote-code loader, synthetic corpus
`v0.3.0-rc1`	synthetic-validated architecture candidate

Revisions are immutable: nothing published is ever overwritten in place.

Citation

@software{zeno_divergent_0_3_1,
  title  = {Zeno Divergent: learned model-relative surprise},
  author = {Barbosa, M.},
  year   = {2026},
  url    = {https://huggingface.co/mbarbosa1/zeno-divergent-v1},
  version = {v0.3.1}
}

zenodivergent.dev — live workspace, interactive benchmarks, full documentation.

Downloads last month: 156

Safetensors

Model size

30.9k params

Tensor type

F32

mbarbosa1
/

zeno-divergent-v1