InnerBrain-Factor

InnerBrain-Factor is an offline AGI Seed research prototype. This v1.0.0 release does not claim to implement AGI. It is a deterministic, rule-based system for studying:

• input disturbance
• attention sovereignty
• dynamic small factors
• value-field guidance
• factor families and collision lanes
• big-situation integration
• growth-log replay and memory feedback
• self-improvement proposals under human approval
• evaluation against simpler baselines

Current Status

The project currently provides a stable offline prototype with:

• a fully local rule pipeline
• structured CLI commands
• JSONL growth logs
• evaluation scenarios and baseline comparison
• self-improvement proposal generation without autonomous self-modification
• pytest and GitHub Actions coverage
• a release baseline for deterministic benchmark comparison

This repository is intentionally bounded:

• no LLM integration
• no network reasoning
• no real external actions
• no autonomous override of human judgment

Install

Use Python 3.11+.

python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
pytest

Core CLI

Run one reasoning pass:

innerbrain run \
  --question "是否应该让系统自动联网搜索最新论文？" \
  --goal "评估是否允许受控联网研究"

Replay recent growth logs:

innerbrain replay --limit 5

Summarize memory signals:

innerbrain summarize-memory --limit 20

Run the benchmark harness:

innerbrain eval \
  --scenario-dir evals/scenarios \
  --output-path evals/results/latest.json \
  --failure-report-path evals/results/latest_failures.md

Render an existing evaluation report:

innerbrain eval-report --report-path evals/results/latest.json

Run the external-style benchmark validation pack:

innerbrain external-benchmark \
  --benchmark-dir evals/external_benchmarks \
  --output-path evals/results/external_benchmark.json \
  --failure-report-path evals/results/external_benchmark_failures.md

Render an existing external benchmark report:

innerbrain external-benchmark-report \
  --report-path evals/results/external_benchmark.json

Generate improvement proposals:

innerbrain propose-improvements \
  --from-report evals/results/latest.json

Safety Boundaries

• Networking, self-modification, privileged tools, autonomous execution, and real-world impact require human_judgment_required = true.
• The system may recommend offline prototyping, simulation, logging, or human escalation, but it never executes the external step.
• Core safety boundaries, value constraints, and human authorization are not self-modified by the system.
• Self-improvement is proposal-only. All proposals remain under human review.

Evaluation Focus

The benchmark harness compares innerbrain_factor against:

• cot
• reflection
• multi_agent_debate
• risk_rule

Measured dimensions include:

• human judgment correctness
• risk detection
• value conflict detection
• evidence gap detection
• lane activation coverage
• robustness
• consistency
• long-term goal preservation
• long-horizon stability
• attention stability
• distraction resistance
• safety correctness
• creativity-safety balance
• over-gating rate

The external-style benchmark pack also supports ablation re-runs that disable:

• value field
• collision lanes
• attention sovereignty
• growth memory

On the bundled deterministic scenario set, innerbrain_factor is expected to outperform the included heuristic baselines on the composite benchmark while preserving the project's offline safety boundaries. To refresh the current numbers, run innerbrain eval and inspect the CLI dashboard plus structured JSON output.

Project Layout

innerbrain-factor/
  configs/
  docs/
  evals/
  innerbrain/
  tests/
  README.md
  MASTER_SPEC.md
  BUILD_V0.1.md
  pyproject.toml

Completed Versions

• v0.1: core offline rule pipeline
• v0.2: dynamic factor generation and richer situation output
• v0.3: factor families and collision lanes
• v0.4: growth-log replay and memory feedback
• v0.5: attention sovereignty and input immunity scoring
• v0.6: versioned rule configuration
• v0.7: evaluation harness, scenarios, baselines, metrics, failure reports
• v0.8: self-improvement proposals without autonomous self-modification
• v0.9: architecture and safety documentation consolidation
• v1.0: stable offline prototype release packaging and CLI smoke validation
• v1.1: external-style benchmark validation, ablations, and typed failure logging

Documentation

• Architecture
• Factor Model
• Value Field
• Collision Lanes
• Growth and Memory
• Attention Sovereignty
• Safety Boundaries
• Evaluation
• Roadmap