DRACO-verified model fusion for Pi agents

Pi Fusion: DRACO-verified model fusion.

v0.2.0 scored 73.80 on a sealed fixed DRACO-10 validation run: +4.80 above the reported Fusion API headline result and +9.10 above the reported budget baseline.

View DRACO proof Install Pi Fusion Open GitHub
Sealed DRACO-1073.80
vs Fusion API+4.80
vs Budget baseline+9.10
Completed cases10/10

DRACO full10 scored · Prompt-only generation · Scorer-only rubric access

Latest validation v0.2.0
Pi Fusion score 73.80 +4.80 over reported Fusion API
Pi Fusion sealed DRACO-1073.80
Reported Fusion API69.00
Reported budget baseline64.70
01Prompt-only generation
02Scorer-only rubric access after seal
030 judge failures across 10 cases
Open benchmark evidence

The single-model problem

One model answer isn’t always enough.

When the stakes are high, a single response can miss tradeoffs, bury contradictions, or commit to one viewpoint before alternatives are even considered.

The fusion answer

Run a panel. Reconcile differences. Decide with evidence.

Pi Fusion runs multiple participant models in parallel, compares their answers, and uses a judge model to synthesize a final response — with optional evidence gathering, verification, and durable artifacts kept locally.

Capabilities

Built for decisions that deserve more than one guess.

Use it for research, planning, architecture decisions, debugging hypotheses, code review, vendor comparisons, writing, and document synthesis.

01

Parallel participant models

Independent answers from multiple models, side by side.

02

Judge synthesis

A judge model reconciles, notes contradictions, and reports the final answer.

03

Verification & revision

Optionally re-check and refine before the answer is final.

04

Optional evidence layer

Connect provider-agnostic search and fetch tools when current, sourced answers matter.

05

Local artifacts & costs

Evidence summaries, token usage, and cost reporting stay with you.

06

Fallback & retry policy

Configurable fallback and retry behavior keeps panel runs resilient.

Workflow

From prompt to scored artifact.

The panel fans out, evidence is attached when available, and the judge turns disagreement into a stronger answer that can be benchmarked after generation is sealed.

  1. 01

    Prompt in

    One prompt enters the fusion panel.

  2. 02

    Parallel answers

    Participant models respond independently.

  3. 03

    Evidence optional

    Search, fetch, and local read-only context can be used when configured.

  4. 04

    Judge synthesis

    The judge compares agreements, contradictions, and missing evidence.

  5. 05

    Verified artifact

    The final answer, evidence summary, token usage, and cost report are saved locally.

Pi-native

Three commands. That’s the surface area.

Configure the panel, diagnose availability, then run fusion from the place you already work.

Commands
/pi-fusion-config       Configure participants, judge, fallbacks, tools
/pi-fusion-doctor       Diagnose model and evidence backend availability
/pi-fusion <prompt>     Run a fusion panel
Examples
/pi-fusion Should we migrate this module to a plugin architecture?
/pi-fusion --fast Summarize the tradeoffs of these three plans.
/pi-fusion --quality Compare vendors and cite current sources.

DRACO benchmark

Scored DRACO validation above the reported Fusion API result.

Pi Fusion is not just a demo of multi-model prompting. On a sealed fixed DRACO-10 validation run, final answers were generated from prompt-only cases, then the scorer loaded rubric artifacts only after the answers were sealed. Pi Fusion scored 73.80: +4.80 above the reported Fusion API 69.00 headline result and +9.10 above the reported 64.70 budget baseline.

Latest DRACO-1073.80
Delta vs Fusion API+4.80
Judge failures0

Verification checklist

Cases10/10 completed
Generation inputSanitized prompt-only cases
Rubric accessScorer-only after seal
ComparisonFusion API 69.00 / budget 64.70
Public artifactsSanitized aggregates only

Public evidence scope

The homepage keeps the claim to aggregate scores and protocol checks. Raw scorer files, private rubric artifacts, and run internals are not published.

Published result73.80 DRACO-10
Protocol claimAnswers sealed before scoring
System / runDRACO scoreΔ vs reported Fusion APIΔ vs reported budget
Pi Fusion latest sealed DRACO-10 validation73.80+4.80+9.10
Reported Fusion API headline result69.00—+4.30
Reported Fusion API budget baseline64.70-4.30—

DRACO-verified here means completed scored runs under this fixed 10-case validation protocol. It is not official DRACO certification or a claim that Pi Fusion beats every Fusion API mode on every evaluation. Generation used sanitized prompt-only case files; answer/rubric/scoring artifacts were available only to the post-generation scorer.

Safety & control

Powerful tools, kept on a short leash.

No multi-agent writes to your real workspace. Evidence and artifacts are explicit, reviewable, and local.

Sandboxed bash

Deterministic calculations run in a sandboxed environment.

Read-only local context

Imported files are explicit evidence, not hidden state.

Optional backends

Fusion still works without search or fetch tools.

Pi-native control

You choose the models, fallbacks, tools, and reporting behavior.

Docs

English-first repo, Chinese docs included.

The GitHub repository is English-first for open-source discoverability, with Chinese documentation for local users and contributors.

README Chinese README Benchmark notes Chinese benchmark notes

Install

Put your next hard question to a panel.

Install Pi Fusion, configure your participant and judge models, then run your first fusion panel in minutes.

Open GitHub Configure models Run /pi-fusion-doctor
Install 
pi install git:https://github.com/aa2246740/pi-fusion@main