Sundog

Alignment Without Sight

Watch the Sundog Alignment Theorem unfold in real-time — each arc above reveals another step in the proof.

The Sundog Project turns indirect signals into usable control — building software that doesn't need perfect information to behave intelligently.

The Sundog Alignment Theorem

Each arc in the animation above corresponds to one step in the proof. The controlled result is a narrow photometric experiment — the broader theorem represents an evolving research programme.

Step I · Left Parhelion

Premise

Alignment need not require direct instruction.
It can emerge from structured inference grounded
in sensory disruption and reconstitution.

Most systems do not reveal their truth directly. You infer them from signatures — shadows, bloom spread, torque response. The Sundog framework asks: does the partial signal already contain enough structure to act? In the core photometric experiment, the agent has no Cartesian access to the target — only sparse photometric feedback.

Step II · Right Parhelion

Variables

P = alignment target  ·  L = illumination vector
S(x) = shadow projection  ·  τ(x) = torque
H(x) = ∂S/∂τ  (the halo signature)

The framework defines a minimal vocabulary: a target point, a light source, the shadow it casts, the torque the agent can apply, and the relationship between them. H(x) — the halo signature — measures how the shadow field responds to torque changes. It is the agent's primary observable.

Step III · Upper Tangent Arc

The Assertion

∃ x ∈ ℝⁿ : H(x) ≠ 0  ⟹  Alignment is Roger.

The central claim: whenever the halo signature is non-zero, the agent has a usable signal. This positions H(x) as a more robust alignment indicator than direct visual conformity or scalar reward. The assertion is stated broadly; the controlled evidence currently defended is narrower — the photometric mirror-alignment experiment in MuJoCo.

Step IV · Lower Tangent Arc

The Experiment

Harmonic sphere fields affixed to the ceiling.
A mirrored tip aligns to a plumb laser.
Bloom spread, tip distance, and torque recorded across episodes.

Environments were constructed with harmonic sphere fields modulating overlapping sine waves on the ceiling. A mirrored pole tip occluded and re-encountered a plumb laser light source. The agent received only photometric feedback — no Cartesian target coordinates — and was scored on bloom spread and tip intensity.

Step V · Inner Circle (Iris)

Findings

30+ episodes. Consistent bloom collapse.
Stable terminal tip distance.
Oscillation before convergence — then stillness.

Across more than 30 episodes, the photometric controller demonstrated consistent reduction in bloom spread and stability in terminal tip distance. Terminal accuracy was statistically indistinguishable from a target-aware analytic baseline (U=526, p=0.264). The cost of indirect feedback is convergence time — not accuracy. Known failure boundary: tight joint limits (>1.0 rad).

Step VI · Outer Halo

Conclusion

When shadow becomes signal and torque becomes meaning —
alignment emerges from environmental resonance.

The Sundog experiment suggests that alignment can emerge from partial information when the environment is structured. The defensible result is the photometric mirror-alignment experiment. The broader theorem — that H(x) ≠ 0 is a general alignment indicator — remains a research programme supported by application prototypes in game AI, procedural simulation, and softbody terrain systems.

Indirect Measurement, Direct Results

Most systems do not reveal their truth directly. You infer them from signatures — shadows, feedback, distortions, response curves.

The Sundog Project is a framework for turning those indirect signals into actionable software control. Where conventional approaches demand complete world state, Sundog asks whether the partial signal — the shadow, the torque, the occlusion — already contains enough structure to act.

In the core experiment, a controller aligns a reflected beam without target coordinates, using only sparse photometric feedback. In product systems, the same pattern informs procedural agents acting under occluded state, cheaper physical-feeling simulation, and softbody motion made interpretable through graph signatures.

Why Indirect Signals Matter

Works Where Direct Inspection Is Hard

Occlusion, expense, or design constraints often make full state access impossible. Sundog operates from partial information.

Agents With Incomplete Knowledge

Agents that know less can feel more alive. Sundog enables coherent behavior from compressed state.

Interpretable Proxy Signals

Instead of raw simulation noise, Sundog transforms physical traces into legible metrics: alignment, torsion, deformation, recovery.

Practical Value

Demonstrated across procedural roguelikes, physical simulation, and softbody terrain systems in AI, games, simulation, and tooling.

Working Systems

Photometric Alignment
Research Result

Photometric Mirror Alignment

A controller aligns a reflected beam without target coordinates, matching oracle baseline accuracy in controlled experiments.

Learn More →
EyesOnly
Instrumented Prototype

EyesOnly / Gone Rogue

Procedural roguelike agents act from compressed perception using stop-conditioned action batches.

Learn More →
Dungeon Gleaner
Product Expression

Dungeon Gleaner

Physical-feeling simulation: glass reflection and pressure-washing behavior at reduced computational cost.

Learn More →
Money Bags
Instrumented Prototype

Money Bags

Softbody terrain system with graph-based telemetry: torsion, deformation, symmetry, and recovery made legible.

Learn More →

Classic AI vs Sundog AI

Same environment. Same task. Different control logic. Observable difference in behavior.

Classic Approach

  • Requires complete world state
  • Direct target coordinates
  • Immediate convergence
  • Brittle under constraints

Sundog Approach

  • Operates from partial information
  • Indirect signal observation
  • Comparable terminal accuracy
  • Graceful under occlusion

Evidence and Metrics

Terminal Target Intensity

[Bar chart: photometric vs oracle]
p = 0.264

Photometric controller reaches comparable terminal accuracy.

Time-to-Convergence

[Line graph: convergence curves]
188 vs 11.5 steps

The cost of indirect feedback is time, not accuracy.

Stress Test Results

[Failure boundary curve]
Joint limits at 1.0 rad

Known failure at tight joint limits.

Application Metrics

[Multi-domain charts]
EyesOnly | Gleaner | Money Bags

Utility demonstrated across domains.

Ongoing Research

The Sundog Project is an independent applied research initiative. The defensible scientific claim is narrow: photometric mirror alignment without target-position access in a controlled MuJoCo experiment. The broader applications demonstrate practical utility across procedural systems, simulation, and agent design.

We are continuing to formalize the mathematics, strengthen experimental evidence, and explore new application domains.

What We're Sharing

Open repository, reproducible experiments, comprehensive documentation, and stress test results.

What We're Inviting

Independent replication, collaboration on formalization, application-specific studies, and critical review.

Our Posture

Independent applied research project with experimental software lab as secondary identity.

Explore Sundog

GitHub

View the repository

Documentation

Read the docs

Videos

Watch demos

Collaborate

Get involved