The model, not the dashboard
Zeno's centre of gravity is a learnable object: S_θ(X_t, R_t, C_t, M_t) → (z_t^S, U_t, T_t, I_t, A_t). Its question is not what happens next, but where the current understanding of what happens next is beginning to fail.
A. Native input representation
The model sees reality and its expectations about reality — X_t + R_t + C_t + M_t.
B. Multi-reference encoder
e_i = Encoder(prediction, uncertainty, residual, role, validity, freshness), pooled by attention over models — not over features.
| reference | pred | unce | resi | role | vali | fres | α |
|---|
C. Learned surprise latent z_t^S
The main representation: how reality relates to current expectations.
Latent topology
z^S is trained to be a topology of surprise kinds, not a single normal ↔ anomalous axis. The path is the scenario clock for this domain.
Predictive heads over z^S
z^S must be predictive, not descriptive.
| head | t+near | t+medium | t+long |
|---|---|---|---|
| P(model failure) | 0.00 | 0.00 | 0.00 |
| P(regime transition) | 0.00 | 0.00 | 0.00 |
| Credible tail mass | 0.00 | 0.00 | 0.00 |
| Expected consequence | 0.00 | 0.00 | 0.00 |
F. Learning policy π_update(z^S)
The model learns when it should learn from data.
D. Training objective
Not one anomaly loss: L = Σ λ_i L_i, with self-supervision over synthetic expectation breakage.
Predict each reference's residual and uncertainty from the shared latent, forcing z^S to carry model-relative — not observation-relative — structure.
Held-out sensor and coverage dropout with a proper scoring rule, so absence of signal raises U_t instead of lowering surprise.
Binary head over horizons h: P(regime transition at t+h) supervised by labelled structural breaks.
Generated trajectories must be low-likelihood under the main reference yet feasible under C_t; scored against realized tails.
Impact across horizons in domain units, so surprise is ranked by consequence rather than novelty.
Policy over {ignore, remember, new regime, rebuild, request evidence}, rewarded by downstream model performance after the decision.
Synthetic perturbations (reference swap, version rollback, sensor dropout, constraint violation) as free labels for surprise kinds.
Status: the architecture, tensor shapes and output contract are implemented and exercised end-to-end here on a frozen, hand-specified checkpoint over synthetic scenarios. No trained weights are claimed yet — the open work is corpus construction, supervision of labelled breaks, and the benchmark on the next page.
G. Hugging Face export
One versioned command builds two clearly separated repositories, tags the release, and verifies the upload against benchmark.json.
Semantic version, checkpoint digest and benchmark digest are resolved before upload, written into both repos as release.json, appended to model/artifacts/releases.json, and pushed as a git tag. Re-exporting identical artifacts reuses the version; any change to weights or benchmark bumps the patch.
The artifact of record. Weights, config, benchmark and the model card live here; the card is attached as the repo README with gated-use frontmatter stating the synthetic-only status.
- model.safetensors— weights for the exported checkpoint
- config.json— auto_map + custom_pipelines — loads with trust_remote_code
- modeling_zeno_divergent.py— ZenoDivergentModel — score_sequence + SurpriseReport
- zeno_architecture.py— the research architecture, flat for remote code
- pipeline.py— pipeline("surprise-detection", …) registration
- calibration_table.json— held-out percentiles of the surprise index
- kind_rules.json— typed-surprise regions and how to read each one
- README.md— model card + badges + rendered leaderboard
- benchmark.json— held-out metrics for SI and all baselines
- release.json— semver, tag, checkpoint and benchmark digests
- verification.json— post-export sanity inference against benchmark.json
- reference_implementation/— data, model, losses, baselines, evaluate, train
An interface, not a model. It publishes no weights, is stamped with the model release it renders, and links back to the model repo on its first screen. Published as a static Space (sdk: static) so it runs on a free account — the Gradio variant needs PRO and is skipped automatically.
- README.md— Space card, version stamp, links back to the model repo
- release.json— the model release this Space build renders
- app.py— Gradio demo — stream ingestion + SurpriseBench explorer
- requirements.txt— gradio, torch, transformers
rendered at the top of <owner>/zeno-divergent-v1The token is only used to fill the one-click command below; it is never sent anywhere by this page and never persisted beyond the tab session. The exporter itself reads HF_TOKEN, then --token-file, then a hidden prompt — so a token never has to appear as a command-line flag.
python -m model.sim.export_hf --owner <owner>
HF_TOKEN=hf_… bun run build && python -m model.sim.export_hf \ --owner <owner> --static dist/client --push
| detector | AUPRC | AUROC | lead@FAR | step-FPR | FA(clean) | ECE |
|---|---|---|---|---|---|---|
| surprise_index (Zeno Divergent) | 0.990 | 0.993 | 7.25 | 0.10 | 0.26 | 0.030 |
| logistic_signals | 0.894 | 0.921 | 4.10 | 0.07 | 0.34 | 0.042 |
| conformal_width | 0.715 | 0.706 | 4.83 | 0.07 | 0.00 | 0.139 |
| cusum_changepoint | 0.699 | 0.803 | 2.66 | 0.08 | 0.66 | 0.162 |
| ood_score | 0.691 | 0.719 | 5.13 | 0.08 | 0.35 | 0.186 |
| anomaly_score | 0.686 | 0.720 | 2.91 | 0.07 | 0.54 | 0.144 |
| disagreement_x_impact | 0.604 | 0.696 | 3.29 | 0.08 | 0.37 | 0.278 |
| knn_novelty | 0.589 | 0.720 | 7.40 | 0.14 | 0.62 | 0.258 |
| control_random | 0.399 | 0.500 | 2.99 | 0.10 | 0.32 | 0.260 |
| ensemble_variance | 0.396 | 0.371 | 1.95 | 0.09 | 0.35 | 0.327 |
| control_constant | 0.394 | 0.500 | — | 1.00 | 1.00 | 0.101 |
| predictive_entropy | 0.284 | 0.239 | 5.64 | 0.10 | 0.46 | 0.598 |
The card links this table directly to benchmark.json in the model repo — the same file this page reads — so the published leaderboard and the interface can never disagree. Verdict on the claim: supported.
- 1 — download model.safetensors back from the tagged revision.
- 2 — load it into the architecture; any missing or unexpected tensor fails the run.
- 3 — run inference over held-out episodes at the benchmark's seed and horizon.
- 4 — recomputed surprise-index AUC must match benchmark.json within 0.05, otherwise the export exits non-zero and writes verification.json.
The export refuses to invent evidence: if model/artifacts/sim.pt is absent it exports the initialised architecture and says so in the card. Nothing published claims real-world calibration.