Program: HYPERLINK (Hypersonic Intercept Performance Evaluation — Relay Navigation, Engagement, Kinematics & Link)
Codename: Project SHEAR (Sensor-degraded Hypersonic Engagement And Reachability)
Working title: Boost-to-Glide Transition Uncertainty & Feasible-Set Prediction under Degradation and Adversarial Maneuvering
Project SHEAR develops real-time, physics-constrained tracking, uncertainty quantification, and feasible-set prediction for hypersonic threats during boost through early atmospheric glide initiation, under sensor degradation, sparse/noisy measurements, and adversarial maneuvering within physical limits.
What SHEAR enables that existing methods don't: a single auditable intercept disk
See docs/what-this-enables.md for the one-page reviewer brief.
Scope: Boost ascent → MECO → coast → pull-up → ~30–60 s early glide. Handoff to sibling project BOUND at equilibrium glide entry.
| Phase | Status |
|---|---|
| Foundation (Phase 0) | Detailed Design / Implemented — problem definition, architecture, UQ approach |
| Boost ROM (Phase 1) | Implemented & Tested (synthetic) — ParametricSample, V&V report, pytest coverage |
| Transition + Glide (Phase 2) | MVP — glide_rom.py, aero.py, transition.py, event log |
| UQ + Pipeline (Phase 3) | MVP (synthetic V&V) — conformal, latency tiers, adversarial stress |
| Plasma observation (Phase 3b) | MVP surrogate — live pipeline + PlasmaGapSpec on RevisitConfig; not CEM-calibrated |
| Hybrid estimator (Phase 4) | MVP foundation — opt-in estimator_enabled |
| Phase branches (COAST/PULLUP) | MVP — phase_propagation.py, opt-in |
| Truth-in-Tube evidence (Phase 6) | Implemented & Tested (synthetic, n=500) — conformal 0.902 · Wilson [0.873, 0.925] · defense suite PASS · RunPod final pass 2026-05-28; not Validated |
| Adversarial stress grid (Phase 7) | Implemented & Tested (synthetic, 36-cell) — REQ-ADV-001 closure 36/36 · baseline closure 24/36 · P3 nominal truth 36/36 (revisit-centroid; prior 2/36 was x_hat proxy bug) · plasma on · not Validated |
| BOUND handoff (Phase 8) | Implemented & Tested (synthetic) — schema + emitter; REQ-HDO-001 smoke PASS 2026-05-28; CI handoff step wired (not Validated) |
| Audit records (REQ-AUD-001) | Schema + emitter — audit.py, pipeline validation, optional emit_audit_path |
| Stage | Meaning |
|---|---|
| Concept | Problem/architecture only |
| Detailed Design | Spec + interfaces, limited code |
| Implemented | Code exists, minimal tests |
| Implemented & Tested | Unit/integration tests pass (scope noted: synthetic, schema, E2E pending) |
| Validated | Independent or flight-truth evidence |
Status table rows use scoped labels — e.g. Integrated MVP (synthetic V&V) — not operational Validated unless explicitly stated.
SHEAR is a research prototype, not operationally accredited. Gate G1 (partner outreach) evidence is packaged for government partnership discussions — not production deployment or DoD accreditation:
validation/reports/g1-partner-briefing.md— stakeholder briefingvalidation/reports/g1-readiness-checklist.md— G1 exit criteriavalidation/reports/defense-readiness-gap-analysis.md— G0→G1 gap analysis
python validation/vv/run_defense_readiness_suite.py| Project | Regime | Primary output |
|---|---|---|
| SHEAR | Boost → early glide | $\mathcal{R}{\text{honest}}(\tau)$, $R{\text{shear}}(\tau)$, handoff packet |
| BOUND | Equilibrium glide |
|
SHEAR addresses the upstream hard problem BOUND explicitly excludes. See references/bound-sibling.md.
cd implementation
pip install -e ".[dev]"
pytest tests/ -vSHEAR/
├── README.md
├── docs/ # Concept note
├── research/ # Problem definition, literature, equations, open questions
├── modeling/ # Architecture, UQ approach, algorithm specs
├── implementation/ # Python package (shear)
│ ├── src/shear/
│ ├── tests/
│ └── data/config/
├── validation/ # Acceptance criteria, V&V harnesses, runpod campaigns
├── certification/ # IUS, traceability, external review packet
├── operator/ # CLI revisit loop
├── runpod/ # Compute profiles and pod helpers
├── references/
└── export/
At each sensor revisit:
| Output | Symbol | Description |
|---|---|---|
| State estimate | Phase-conditioned, physics-constrained | |
| Honest reachable set | Decomposed param + phase + obs_gap + adversarial | |
| Bounded radius | Ground-plane disk at coverage |
|
| Validity mode |
FULL_TRACK … INFEASIBLE_STATE
|
Fail-closed gating |
| Handoff packet | BOUND ingest | At stable EARLY_GLIDE
|
On this problem, impressive ≠ low RMSE:
- Honest tubes through transition — n=500 stressed conformal pinned (overall 0.902, Wilson [0.873, 0.925], synthetic ROM); 36-cell adversarial grid: REQ-ADV-001 closure 36/36, P3 nominal truth 36/36 (revisit-centroid)
- No covariance collapse after pull-up vs. UKF-only baseline (partial synthetic evidence)
- Adversarial closure — DE worst-case endpoint inside inflated disk on cert grid; not the same as nominal truth coverage
- Physics beats kinematics when constrained; wider when degraded (synthetic V&V)
- Phase honesty —
PHASE_UNCERTAINduring transition, not false confidence - BOUND handoff — pod E2E PASS; CI step wired; double-counting review not Validated
- Latency — fast tier adversarial P50 ~389 ms on elite pod (31 vCPU); tier- and config-dependent; REQ-LAT-001 500 ms target met on pinned prod preset only
- No classified ground truth — synthetic validation only
- Plasma surrogate may miss band-specific blackout — must stay conservative
- Phase detection latency → collapsed multi-modal posterior
- Scope creep into BOUND equilibrium glide territory
- Conformal calibration on nominal-only holdout → optimistic coverage under adversarial stress
- Research prototype only — no operational accreditation or flight-truth validation
- Plasma model — phenomenological surrogate; not CEM/radar-calibrated
- Adversarial stress — 36-cell expanded grid pinned (
cbc808d): closure 36/36 (redteam) / 24/36 (baseline harness); P3 nominal truth 36/36 (revisit-centroid; prior 2/36 was post-process proxy bug) — not flight-validated - Conformal holdout — n=500 stressed conditional pinned (overall 0.902); prior n=100 in git history; calibration still
calibrate_without_adversarial: trueby default - BOUND handoff — pod smoke PASS; CI E2E wired; not operationally Validated
- Hybrid estimator / phase branches — opt-in MVP; not default production path
- First green GitHub CI run with public BOUND + handoff artifact
- Fix post-process P3 metrics export (store revisit centroid in grid CSV to avoid re-run)
- Optional
--with-adversarialconformal calibration pin (separate from n=100 stressed holdout) - Partner sanitized track ingest (G2)
- Fire-control-quality birth-to-death salvo tracking
- Real-time full 6-DOF game-theoretic reachable sets for unknown adversary
- Guaranteed intercept against arbitrary adversarial HGV
- Single-modality radar through worst-case plasma blackout
Working materials are UNCLASSIFIED for portfolio and technical submission use.