Pre-Reasoning V4

Mia Labs' Pre-Reasoning Engine V4 is a super-optimized release: it reduces the parameter requirement by 93% compared with the previous V3 engine (13.7M to 1M).

check the file loss_function_optimization.md if interested on the how.

Usage

from pre_reasoning import analyze
result = analyze("Frontend depends on API. API depends on Auth.")
print(result["trace"])

The bundled release checkpoint is the weights-only file pre_reasoning/checkpoints/pre-reasoning-1m.safetensors. Training checkpoints and optimizer state are not distributed.

Contract

V4 retains the five-family structural output: dependencies, conflicts, requirements, conditionals, transitive closure, root blockers, and cycle detection. The deterministic harness helps with entities and preserves same V4 contract; the 1M model supplies the learned component behind that contract.

The promoted checkpoint passed the V4-compatible harness board at 3,000/3,000, including 1,000/1,000 F1–F5 cases.

Layout

• pre_reasoning/ — single public V4 package
• pre_reasoning/checkpoints/pre-reasoning-1m.safetensors — bundled 1M weights-only checkpoint
• pre_reasoning/engine.py — V4 API and harness adapter
• pre_reasoning/inference.py — minimal NanoChat model runtime
• pre_reasoning/engine_core.py — stable graph-core compatibility module
• pre_reasoning/harness.py — structural graph harness

No previous-engine checkpoint, training sidecar, or MoE is included (or needed) in this release tree.

License

MIT. See LICENSE.