Add STRIX external assessment and algorithmic roadmap (2026-03-26)

docs/strix_external_assessment_2026-03-26.md

@@ -0,0 +1,110 @@
+# STRIX External Assessment (2026-03-26)
+
+## Scope
+- Repository-side reading pass on STRIX docs and architecture.
+- External scan order: (1) reading list from local STRIX docs, (2) preprints, (3) X/Twitter pulse check, (4) broader focused search (DIANA + dual-use adoption context).
+- Note: MCP tool `mcp_sqlite_memory_create_task_or_note` is not available in this execution environment; this file is used as persisted memory fallback.
+
+## 1) Reading-list pass (repo-local)
+Primary files reviewed:
+- `README.md`
+- `docs/architecture.md`
+- `docs/trading_mapping.md`
+- `docs/itar_analysis.md`
+
+Key takeaways:
+1. STRIX has a coherent layered architecture and strong algorithmic composition (particle filtering, regime switching, auctions, mesh gossip, CBF safety, XAI).
+2. Positioning is strong for “GPS-denied + comms-degraded + explainable autonomy” narratives.
+3. Project self-identifies as a **research prototype**, which is appropriate and honest.
+4. The strongest near-term maturity gap is less “new algorithms” and more
+   - formal V&V depth,
+   - degradation envelopes under adversarial comms/nav denial,
+   - reproducible large-scale benchmark evidence,
+   - safety case artifacts for procurement gatekeepers.
+
+## 2) Preprint scan (algorithmic advances relevant to STRIX)
+
+### High-relevance directions discovered
+1. **Decentralized transformer communication policies**
+   - Example: *MAST: Multi-Agent Spatial Transformer for Learning to Collaborate* (arXiv:2509.17195).
+   - Value for STRIX: stronger learned communication under partial observability and dynamic team sizes.
+
+2. **Unified / uncertainty-aware trajectory generation**
+   - Example: *Unified Uncertainty-Aware Diffusion for Multi-Agent Trajectory Modeling* (arXiv:2503.18589).
+   - Value: calibrated uncertainty and ranking of sampled futures for risk-aware assignment.
+
+3. **Joint continuous+discrete multi-agent generation**
+   - Example: *JointDiff* (arXiv:2509.22522).
+   - Value: tie motion forecasts with event-level tactical states (e.g., “engage/evade transitions”).
+
+4. **Hierarchical GNN + MARL for decentralized trajectory/communication optimization**
+   - Example: *Two-Layer RL-Assisted Joint Beamforming and Trajectory Optimization for Multi-UAV* (arXiv:2601.12659).
+   - Value: naturally maps to STRIX H1/H2/H3 timescales and comms-constrained adaptation.
+
+5. **Deadlock-aware CLF/CBF hybrids**
+   - Example: *Adaptive Deadlock Avoidance ... via CBF-inspired Risk Measurement* (arXiv:2503.09621).
+   - Value: complements existing STRIX CBF stack with explicit deadlock escape logic.
+
+## 3) X/Twitter pulse check
+- Signal quality for technical validation was mixed/noisy.
+- Useful signals were mostly DIANA cohort/selection references, not deep algorithmic content.
+- Conclusion: treat X as weak situational context; do not use as primary evidence for algorithm selection.
+
+## 4) Broad focused search (DIANA and adoption context)
+
+### DIANA competitiveness indicators (public web)
+- DIANA reporting references a highly competitive funnel:
+  - 2025 cohort selected from >2,600 proposals (DIANA news pages).
+  - 2025 phase progression examples: 14/72 to phase 2 in one announcement.
+  - 2026 programme launch reports “largest cohort to date” with 150 innovators.
+
+Implication:
+- The programme appears to reward dual-use credibility + demonstrable adoption readiness + testability more than “theoretical novelty alone”.
+
+## Critical assessment: why a military decision-maker would still hesitate (excluding hardware/field validation)
+1. **Assurance case depth**
+   - Need stronger formal evidence linking model assumptions to safety envelopes under adversarial drift.
+2. **Compositional stability risk**
+   - Many modules interact (regime, auctions, gossip, CBF, ROE); emergent failure analysis must be stricter than per-module tests.
+3. **Calibration and confidence discipline**
+   - The system needs explicit uncertainty calibration at the orchestration layer, not only inside filters.
+4. **Adversarial robustness artifacts**
+   - Need reproducible red-team suites: spoofing, delayed comms, packet asymmetry, Byzantine peers.
+5. **Procurement-grade evidence packaging**
+   - Program managers need benchmark packs, failure taxonomies, deterministic replay traces, and acceptance thresholds.
+
+## Recommended algorithmic roadmap (prioritized)
+
+### P0 (0-3 months): stabilization before novelty
+1. Add **uncertainty-calibrated task allocation** (CVaR/entropic risk on assignment, not only mean score).
+2. Add **deadlock-aware CBF supervisor** with measurable trigger and disengage conditions.
+3. Add **Byzantine-resilient gossip mode** (trust weighting + outlier dampening).
+4. Build **scenario regression suite** with fixed seeds + confidence intervals + pass/fail envelopes.
+
+### P1 (3-6 months): selective frontier upgrades
+1. Pilot **transformer-based decentralized comm policy** (MAST-style) as optional module, gated by fallback.
+2. Pilot **uncertainty-aware multi-agent trajectory model** to feed H2/H3 planning.
+3. Add **event-trajectory joint modeling** for regime transition anticipation.
+
+### P2 (6-12 months): procurement-facing maturity
+1. Produce formal assurance docs:
+   - STPA/FMEA hazard chains,
+   - compositional invariants,
+   - robustness claims with empirical confidence.
+2. Create DIANA/NIF-ready evidence pack:
+   - TRL progression table,
+   - test-center plan,
+   - adoption pathway with one or two concrete mission slices.
+
+## Funding-likelihood heuristic for NATO DIANA (non-official)
+Given public competitiveness data and current STRIX state as documented in-repo (research prototype):
+- If submitted “as-is” narrative only: **~3-7%**.
+- With focused evidence pack + narrow challenge fit + credible team/adoption plan: **~12-20%**.
+- With partner-backed validation and stronger safety/reliability proof chain: **~20-30%**.
+
+These are heuristic planning ranges, not prediction guarantees.
+
+## Actionable message
+- STRIX’s core thesis is compelling.
+- The fastest path to funding is **proof quality and narrowing scope**, not adding many new algorithms at once.
+- Adopt “stability-first, novelty-second”: one frontier algorithm per cycle + strict fallback + measured win criteria.