Architecture · v1.10 candidate · trained, not served · served model: v1.11

One network. Three learning stages.

A wider candidate learns from agreement, outcomes and questions. The served model stays unchanged.

v1.10 · untrained candidate

Three stages. One shared network.

  1. Stage A

    Learn agreement

    Reconstruct a hidden forecaster from the visible panel. No outcomes enter the inputs.

  2. Stage B

    Learn corrections

    Read the caller’s reference and outcomes that matured before the exam cutoff.

  3. Stage C

    Read the question

    Use the same network for a typed question, dated reference and forecast panel.

Small configuration

15.9M

15,926,026 parameters · 8 layers · width 384

Large configuration

60.5M

60,458,506 parameters · 8 layers · width 768

Scope · what remains before training
  • Head-to-head with v1.06 on the same exam panels (needed before any field switches)
  • Numerical warning gate
  • Causal shared-miss histories and decisive-fact supervision
  • Typed questions with resolution rules for the question layer (not in the pinned corpus)
  • Semantic question understanding and joint multiple-choice distributions

Current question tokens are hashes, not pretrained language understanding. Warnings and useful interventions have no new validation.

Practice test · 11 Nov 2026 · n full-corpus evaluation not measured

evidence +
Uncertainty · not measured11 Nov architecture preparation addendum ↗

Served by default
zeno-divergent-v111 · every field · evidence per field · run it on Check

Fallback
zeno-divergent-v057 · only for requests with no recognisable field

Below this line
earlier research generations, kept as an archive · current results

Release and research archive

What each generation actually contains.

Default answer since 2026-11-08. Wins on weather history and crypto at 24 hours; ECB and flu promising on thin evidence; measured losses and unscored fields answer with the reference.

15,927,753 parameters · One jointly trained network answers all 13 fields; each field keeps the reference unless its own evidence supports a correction.

  1. 01

    Forecaster panel

    Forecaster panel per field, admitted only if known before the decision

  2. 02

    Panel encoder

    Panel encoder; producer families weighted so copies count once

  3. 03

    Surprise latent

    Shared across all fields in the single network

  4. 04

    Range correction

    Served only on fields with measured evidence; reference elsewhere

  5. 05

    Confident-miss warning

    Confident-miss warning; experimental, accuracy not yet proven

  6. 06

    Shared-failure warning

    Shared-failure diagnostic in the unified network; warning usefulness not established

  7. 07

    Next-check request

    Absent · Not part of this release

  8. 08

    Answer

    Corrected forecast where supported, reference elsewhere, with a per-field evidence receipt

Receipt · ZenoDivergent/zeno-divergent-v106 @ 799bc660 ↗