Risks, Limitations & Mitigations
| Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|
| Model hallucination in engineering specs | High | High | Simulation validation required before any design clears pipeline; physical ground-truth anchoring mandatory |
| RAM pressure under full multi-agent load | Medium | Medium | Celery async queuing; state serialization to SSD above 85% memory pressure; Nactive throttle enforced |
| Patent data staleness creating prior art blind spots | Medium | Medium | Live retrieval on every project — no cached patent data older than 24 hours used in novelty analysis |
| Ethics framework circumvention | Low | Very High | Hard-coded constraints in protected system layer; Ethics Officer runs as a separate model instance |
| Large lab models closing capability gap | High | Medium | HIIE's moat is the full pipeline + ethics governance + domain specialization + self-improving feedback loop |
| Single hardware node failure | Low | High | Daily HDD backup; Coolify restart policies; all project state checkpointed to SSD; Registered Agentics on-site maintenance |
| Legal uncertainty around AI-generated patent claims | Medium | High | HIIE generates drafts for human attorney review — not direct filing; human-in-the-loop documented at every step |
| Recursive loop / stack overflow | Medium | Medium | Hard depth limit Dmax = 8; cosine similarity divergence enforcement; Celery task timeout at 300s per agent |
| HCT economy miscalibration causing agent starvation | Low | Medium | Treasury Agent monitors spending velocity; reserve pool absorbs cascade failures; coefficients recalibrated per project |
| Alice intake gate creating friction rather than convergence | Low | Medium | Gate returns one targeted question per cycle; Exploratory Mode bypasses gate entirely; Arthur calibrates if pass rate degrades |
| ANE API instability across macOS updates | Medium | Medium | Adapter store and training pipeline version-pinned; rollback to last stable macOS version if ANE API changes |
| LoRA adapter degradation from low-quality training batches | Low | Medium | Adapter versioning with rollback; Treasury Agent monitors Fscore delta per adapter version; degraded adapters automatically flagged and rolled back |