Add README

README.md

@@ -0,0 +1,59 @@
+# Autonomous Innovation Engine
+
+A 3-layer multi-agent system that generates, adversarially evaluates, and refines
+innovation concepts — and gets smarter at filtering as its memory of past outcomes
+grows.
+
+## Architecture
+
+**Layer 1 — Generation (divergent).** Three generator agents with distinct personas
+and temperatures: The Systems Pragmatist (buildable-in-12-months), The Frontier
+Researcher (bleeding-edge), and The Contrarian Designer (inversions and overlooked
+niches). Each proposes structured ideas (name, domain, pitch, mechanism,
+differentiator) per round, conditioned on what the memory plane already accepted.
+
+**Layer 2 — Evaluation (convergent, adversarial).** A four-critic council with a
+weighted rubric: Feasibility (0.25), Novelty (0.25), Market Viability (0.25), and a
+Red Team Executioner (0.25) that attacks every idea and scores how well it survives.
+Ideas clearing the aggregate threshold (7.0) *and* the red-team floor (6.0) are
+accepted; the top ideas per round go back to their generator with the council's
+criticisms for one refinement pass, then re-score.
+
+**Layer 3 — Memory Plane (continuous optimization).** A persistent JSONL record of
+every idea, its per-critic scores, and its decision — synced to
+[LoveLogicAI/innovation-engine-memory](https://huggingface.co/datasets/LoveLogicAI/innovation-engine-memory).
+It calibrates the critics (per-domain score baselines, acceptance rate, best/worst
+examples injected into critic prompts) and detects near-duplicate pitches against
+past ideas via sequence similarity.
+
+## Run results (demo, 2026-10-08)
+
+Qwen/Qwen3-1.7B, fp16, T4-small. 3 rounds, 24 ideas, 22 accepted, ~50 LLM calls,
+6,594 output tokens, ~5 min wall. Mean aggregate rose across rounds
+(7.50 → 7.66 → 7.94) as the memory plane calibrated the council.
+
+Top ideas by aggregate score: NebulaNet 8.5, ResilientMesh 8.25, NeuroSync 8.25.
+
+Known limitations, observed on this run: the 1.7B critics are lenient (22/24
+accepted — a larger judge model would sharpen the filter), and the novelty critic
+did not fully suppress near-duplicate names across rounds. Both are memory-plane
+calibration targets, not architecture flaws.
+
+Full run artifacts: `results/ideas.jsonl` (every idea with per-critic scores),
+`results/summary.json` (per-round stats). Live metrics:
+[Trackio dashboard](https://huggingface.co/spaces/LoveLogicAI/autonomous-innovation-engine-trackio).
+
+## Usage
+
+```bash
+pip install transformers==5.19.0 torch==2.14.1 accelerate==1.15.0 trackio==0.41.0
+
+python engine_demo.py --dry-run   # full loop with a mock LLM, no GPU
+python engine_demo.py --smoke     # 1 tiny round on real hardware
+python engine_demo.py             # full 3-round run
+ENGINE_PUSH=1 python engine_demo.py  # also push memory + results to the Hub
+```
+
+Set `TRACKIO_SPACE_ID` to log live metrics. The memory plane (`memory_plane.jsonl`)
+persists across runs: rerun in the same directory (or download it from the dataset
+repo first) and the critics start calibrated from prior outcomes.