# Zeno architecture preparation — 11 November 2026

This dated addendum describes an untrained candidate, not a release or a training result. Zeno v1.06 remains the sole default model: one jointly trained network across 13 fields.

## Implemented candidate

One shared set-transformer reads forecast panels, the caller reference and typed question through the same training/package encoding. The small configuration has 15,926,026 parameters (width 384, 8 layers, 6 heads); the large has 60,458,506 (width 768, 8 layers, 12 heads). Counts describe configurations; no trained full-size weights exist.

Stage A reconstructs hidden forecasters without outcome inputs. Stages B and C use mature supervised outcomes, an explicit caller reference and typed question inputs. Fixture checks cover optimizer steps, capped sampling, checkpoint integrity, raw-input reload parity and exact interrupted-run resume. Fixture completion is not a scientific evaluation.

Text tokens are deterministic hashes with randomly initialized embeddings, not pretrained semantic embeddings. Categorical distributions are not implemented. Causal dependence, useful warnings and decisive-fact supervision remain unfinished.

## Corpus boundary

The historical field registry lists 24 fields. It is an inventory, not a reconciled next-run training pack or an exposure receipt. Wind (air-wind-10m) and COVID-19 (health-covid) are removed from all three training stages for the unseen-field exam. Energy is held in the current preparation inventory because it has only one forecaster. Training eligibility requires row-specific time and input checks; inventory counts do not establish it.

Separately, Offdiagonal v6.49 at f9f7a7003975f65022fe168e4841439a8353c3b9 contains 235,207,134 ensemble member rows in 62 daily files. The audit reconciles 149,962,002 release-timed rows and 85,245,132 capture-time-only rows to v6.48. Capture-only rows must not borrow provider release proof. Even release-timed rows need usable_from <= decision_at < target_at. Member rows are not forecast panels or independent outcomes and are not added to the historical registry total.

Audit: /v649-ensemble-audit-20261111.json

## Before a run

Production pinned corpus readers, admission/consumption receipts, a fixed exam calendar, matched rivals and a quantitative warning protocol remain required. A real-data rehearsal and owner approval of that specific run are also required. No run has launched, no successor is served, and no new warning authority or demonstrated surprise-intelligence claim follows from preparation.