ZMLC+ Track 1: Local-First Token Router with Deterministic Gates and Fireworks Fallback

Artificial Intelligence (AI)
1

Hi everyone,

I’m sharing my Track 1 submission for the AMD Developer Hackathon ACT II: ZMLC+ Local-First Token Router.

The core idea is simple: not every benchmark task should be treated as open-ended LLM generation. Some tasks are deterministic or directly verifiable, such as arithmetic, exact JSON, sentiment, strict formatting, common reasoning patterns, and narrow code templates. ZMLC+ moves that safe slice out of the lossy remote-generation path and answers it with deterministic solvers, verifier checks, optional local Gemma/GGUF inference, and output normalization.

Only the remaining risky or genuinely lossy slice is routed to official allowed Fireworks models such as Kimi and Minimax, using compact task-specific prompts.

What it does

  • Handles deterministic/verifiable tasks with zero counted Fireworks tokens.
  • Uses local inference when available through an OpenAI-compatible local endpoint.
  • Applies verifier and parser guards before accepting local answers.
  • Falls back to allowed Fireworks models only when needed.
  • Implements the official Track 1 Docker batch contract: reads /input/tasks.json, writes /output/results.json.

Current validation

Fresh 12-task smoke results

I tested the current v2 router on a small fresh mixed set of 12 hand-written tasks, separate from the synthetic hard slice.

Task mix:

  • 3 math tasks
  • 4 QA tasks
  • 2 strict-format tasks
  • 2 code tasks
  • 1 reasoning task

Result:

System Accuracy Fireworks tokens/task Parser pass Verifier precision/recall Crash-free
Remote-only 0.8542 193.75 1.0 1.0 / 1.0 1.0
ZMLC+ v2 0.8875 19.33 1.0 1.0 / 1.0 1.0

That run gave:

  • 90.02% fewer counted Fireworks tokens
  • +0.0333 accuracy delta versus remote-only
  • 0 verifier false positives
  • 0 wrong local-only accepts

This is a small smoke test, not the headline benchmark, but it checks the same behavior on fresh tasks: exact or verifiable work is handled locally/deterministically, while risky tasks still go to Fireworks.

On a hard 60-task proxy slice:

  • Remote-only accuracy: 0.9117
  • ZMLC+ accuracy: 0.9225
  • Remote-only Fireworks tokens/task: 145.75
  • ZMLC+ Fireworks tokens/task: 47.68
  • Token reduction: 67.28%
  • Accuracy delta: +0.0108
  • Parser pass, verifier precision/recall, and crash-free rate: 1.0

On a separate fresh mixed smoke run, ZMLC+ reduced Fireworks tokens by 90.02% with +0.0333 proxy accuracy delta.

Docker image:
```text
christiandeluca/zmlc-track1:v4