Trifecta Reasoner

I’m testing what a local model can do when the application carries more of the workflow.

Format: Local model / retrieval / staged answers
Status: PRIVATE APP

Recorded Trifecta Reasoner answer showing four US states, capitals and admission years — Saved acceptance screenshot / Deep mode with search, October 2026.

01 / The project

What I’m working with

Reasoner has Quick and Deep routes. One takes a short path; the other can draft, inspect and repair an answer.

The designer wanted usable prose and tables. Structured records, deterministic comparisons and explicit retrieval do work that another confident paragraph cannot do by itself.

02 / Workflow

How it works

Choose the task

Select Quick or Deep and decide whether the answer needs current retrieved evidence.

Work through the answer

Run a bounded sequence of model stages, using separate paths for prose and tables.

Show the result

Display the answer with actual queue feedback and plain-text copy. Evidence retrieval failure is reported rather than hidden by a model-only replacement.

03 / Inside the project

Inside the project

Reasoner controller and clientJob handling, local-model calls and a reusable Python interface.
Retrieval and answer modulesTime context, evidence handling, typed tables and prose checks.
Recorded browser acceptanceThe saved states-and-capitals example shown here.

04 / Status

Where I leave it

The saved example shows the states-table workflow in use. The project continues around answer quality, retrieval and bounded repair.