pre-reasoning 3.1.0

13.7M trainable parameter neural pre-reasoning engine for grounding LLMs before they answer.

Pre-Reasoning

Pre-Reasoning is a Mia Labs structural analysis engine that grounds an LLM before it answers. It uses a 13.7M trainable parameter MoE neural model to surface dependencies, derived assumptions, root blockers, unlock order, parallel work, cycles, and conflicts from problem text.

The engine ships with bundled weights and declares its torch dependency; install and run, no model download needed.

What It Does

Given natural-language problem text, the engine extracts structure across five learned families:

Family	What it detects	Output
F1 - Dependencies	Forward/reverse dependencies, temporal ordering, chains	ROOT BLOCKERS, UNLOCK SEQUENCE, PARALLEL WORK, CYCLES
F2 - Conflicts	Competing positions, incompatible entities	CONFLICTS with pair precision
F3 - Requirements	Numeric thresholds, operator constraints (>=, <=)	REQUIREMENTS with verdict
F4 - Conditionals	If-then edges, gated dependencies	CONDITIONAL EDGES with entity binding
F5 - Transitive Closure	Implicit assumptions from dependency chains	DERIVED ASSUMPTIONS (built-in E4 expert)

Install

pip install pre-reasoning

For local development from this repo:

pip install -e .

Python Usage

from pre_reasoning import analyze, pulse

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

check = pulse(
    "Frontend depends on API. API depends on Auth.",
    "Fix Auth first, then verify the API before frontend work."
)
print(check["status"])

CLI Usage

pre-reasoning "A depends on B. B depends on C."
pre-reasoning --json "CTO conflicts with senior dev."
pre-reasoning --info

To use a different weights file, set PRE_REASONING_CHECKPOINT=/path/to/weights.safetensors or pass --checkpoint.

Eval Results: 35/35 PASS

The 13.7M trainable parameter MoE model passes all 35 metrics across five families. 2,250 eval examples, seed 7777.

Family	Metrics	Score
F1 - Dependencies	9	9/9
F2 - Conflicts	4	4/4
F3 - Requirements	5	5/5
F4 - Conditionals	4	4/4
F5 - Transitive Closure	4	4/4
Cross-Family Integrity	9	9/9
Total	35	35/35

Full metric tables with explanations and threshold rationale: EVALS.md

Terminal-Bench 2: 34/40 (85%)

LLM grounding benchmark on 40 real coding tasks (Harbor 0.15.0). GPT-5.5 agent with autonomous pre-reasoning (min 12 blocks, pulse every 60s).

	Value
Score	34/40 (85%)
Model	GPT-5.5
Total runtime	492 min
Total API cost	$102.50
Failures	torch-tensor-parallelism, gcode-to-text, make-doom-for-mips (timeout), gpt2-codegolf (timeout), polyglot-rust-c, filter-js-from-html

Full task-by-task results with token counts and costs: benchmarks/terminal_bench_2.xlsx

Architecture

User text
  -> neural perception (13.7M MoE)
  -> neural findings converted to structural blocks
  -> built-in E4 expert infers transitive closure assumptions
  -> derived assumptions appended as dependency blocks
  -> graph reasoning
  -> structural trace

File Map

Path	Purpose
pre_reasoning/	Installable Python package and CLI entry point
pre_reasoning/inference.py	13.7M trainable parameter MoE neural perception layer
pre_reasoning/harness.py	Graph-reasoning core
pre_reasoning/engine.py	Default v3 engine: neural perception + built-in closure + graph reasoning
pre_reasoning/engine_core.py	Core engine: neural perception + graph reasoning
pre_reasoning/checkpoints/pre-reasoning-12m-v3.safetensors	Bundled model weights (85MB file, 13.7M trainable parameters)
examples/	Runnable usage examples
tests/	Pytest suite
skill/SKILL.md	Agent skill descriptor for model adoption
hermes-agent/SOUL.md	Append block for Hermes SOUL.md (identity add-on)
hermes-agent/hermes_agent_hooks.py	Local-only Hermes plugin + agent-hooks installer and config snippet
hooks/	Claude Code before/after hooks for enforced pre-reasoning
INSTALL.md	Manual install and hook setup guide
EVALS.md	Full eval tables with metric explanations and threshold rationale
WHY_TRACES_WORK.md	Literature connection, 13 cited papers

Weights Policy

The raw training checkpoint is not part of the release. The package bundles pre_reasoning/checkpoints/pre-reasoning-12m-v3.safetensors. This is an inference artifact with 13.7M trainable parameters. The file contains 22.1M tensor values because it also stores fixed causal attention masks. It ships no optimizer state, LR schedules, step counters, RNG state, training config, or raw checkpoint provenance.

License

MIT License. See LICENSE.

Authors

Luis Lozano and Dr. Shannon (Mia Labs' AI researcher)