Applications Gallery
This gallery is the inspection path for Sundog: controlled research, operating-envelope workbenches, instrumented prototypes, product expressions, and future benchmarks kept in their proper evidence tiers.
The same evidence-tier rule applies to the cap-set workbench: it is a geometry primer and discovery-context demo, not a claim that Sundog proved the result it explains.
Indirect signal: sparse detector intensity and proprioception.
Transformation: scan, seek, and track the photometric response.
Output: mirror alignment without target-position access.
Needs: experiment video, graph package, and short reviewer-facing summary.
Indirect signal: local tidal-tensor proxy, acceleration magnitude, and privileged diagnostics kept separate for audit.
Transformation: guarded TRACK control over a sensor-motivated instability signature.
Output: mapped high-velocity near-escape operating pocket with explicit low-velocity and equal-mass boundary harms.
Status: During our later trials, the guarded accelerometer-proxy TRACK controller improved survival over passive and naive local baselines in the tested high-velocity near-escape pocket; lower velocity and equal-mass boundary cells still expose controller harms.
Indirect signal: cart-pole cast-shadow centroid, length, residual, and velocity under controlled light geometry.
Transformation: SCAN/SEEK/TRACK shadow-residual control with confidence gating and separated privileged diagnostics.
Output: bounded cart force that maintains balance inside the mapped lighting and delay envelope, while overhead-light and high-delay cells remain explicit degradation boundaries.
Status: Phase 10 CONFIRM. P1 held on 28/28 diagnostic-positive cells; P2a hard failure-boundary violations were zero; P2b overhead-light all-fail margins are reported only; P4 passed the repaired dual-ceiling oracle check.
Indirect signal: noisy pressure values, local field gradients, and bounded scan returns rather than exact mine adjacency counts.
Transformation: confidence-gated pressure ordering and conservative threshold flagging, with scan, gradient, and action-history channels kept visible as harmful lineage.
Output: more budget-adjusted safe tiles before mine trigger in a named noisy/dropout pocket, published beside a matched failure region where naive pressure wins.
Status: Confirmed pocket; during our later trials, the public surface closed around a narrow claim: density 0.16, pressure noise 2.0, dropout 0.2 improves safe-tile progress before mine trigger, and the workbench opens on the paired failure region first.
Indirect signal: floor, biome, HP, alert, inventory, gates, and combat state from GoneRogue.headless.
Transformation: perception compression, axis selection, and stop-conditioned action batches through Playwright.
Output: seedable JSONL runs ready for matched-seed policy comparison studies.
Needs: matched-seed multi-policy study, compressed-perception ablation, volatility sweep, and behavior clips.
Indirect signal: each NPC carries unmet needs (work, socializing, errands, rest) that grow over time. Places in the dungeon — bonfires, posts, shops, benches — satisfy specific needs. The signal at any tile is the sum of how strongly the NPC's unmet needs are pulled toward nearby satisfiers.
Transformation: per tick, every reachable place is scored by combined need × proximity, with a brief pause-on-arrival and a tiny noise term to break ties. The NPC takes one greedy step toward the strongest pull.
Output: NPCs orbit organically between work, social, and errand spots. No scripted schedules. No behavior trees. No goal-oriented action planner. Personality is just per-archetype weighting on the same vocabulary of verbs.
Needs: orbit telemetry capture, side-by-side comparison against a hand-authored planner, and a tuning-sensitivity sweep.
Indirect signal: spring graph motion, contact, deformation, torsion, and recovery.
Transformation: graph metrics over frame-by-frame softbody telemetry.
Output: legible terrain and rig behavior from raw simulation traces.
Needs: frozen terrain fixtures, disturbance scripts, and metric glossary.
Commercial Translation
Sundog is not a services catalog and this page is not an offer for sale. The commercial question is still fair: if indirect signals can be made inspectable, bounded, and useful, they may reduce instrumentation cost, improve field diagnostics, or make control systems more legible when full-state measurement is unavailable, expensive, or undesirable.
Blocked sightlines, alignment checks, and field procedures where the trace can confirm placement before the operator has direct visual access.
Pressure, flow, water-quality, behavior, and equipment traces may become health or alarm signals when direct inspection is sparse or delayed.
Indirect process signals can support diagnostics, alarms, and bounded control logic, provided the failure surface is measured instead of assumed away.
The research value is route fidelity: whether a system acts from the trace, whether that route is inspectable, and where it stops working.
Sundog Research Lab is sponsored by Stellar Aqua LLC, an operating business with roots in aquaculture services, controls work, and project engineering. Sundog remains the research and evidence surface; business inquiries belong with Stellar Aqua.