A traceability harness for useful partial information.

Sundog is a traceability harness for indirect-inference alignment: a way to test whether a system can act from indirect signals, whether the route is inspectable, and where the coupling fails.

The year-one result is not a universal theorem. It is an apparatus: small systems, explicit baselines, public stress tests, and named failure boundaries around the cases where partial information becomes useful.

That same apparatus posture now extends to public math primers such as the cap-set workbench: a bounded way to inspect external geometry breakthroughs without pretending the primer is the proof.

The public site now leads with evidence surfaces that can be inspected independently. The photometric result remains the narrow controlled claim. Three load-bearing pillars keep the broader apparatus honest: write the falsifier first, map the operating envelope, and keep proof-status gates visible.

Full state is often a fantasy. Real systems are occluded, noisy, delayed, expensive to instrument, or deliberately hidden. A useful system may not get to inspect the target, and sometimes it should not be given privileged state just to make the demo easy.

The target may be hidden, but the world responds. Light shifts. Torque changes. A field bends. A rig deforms. A player hesitates. A procedural system exposes just enough state to act without authorial omniscience.

We call that surrounding structure the halo: the signal around the signal. The lab's work is to make the halo inspectable.

Sundog began as a field alignment problem, not as an abstract theorem. The original procedure came from aligning acoustic ceiling anchors for suspended HVAC equipment above a quantum-computing environment.

A plumb laser, a fastener head, and an occluded line of sight forced a practical question: could alignment be confirmed through the loss and return of a signal, rather than by direct inspection of the seated fastener?

That origin keeps the lab grounded. Sundog is about work done under constraints: blocked sightlines, imperfect communication, physical tolerances, and the need to know when a procedure has actually landed.

Sundog workbenches and product systems still follow one recurring pattern:

1. The decisive state is hidden.
2. The world leaks structure through an indirect signal.
3. That signal is compressed into a control-relevant form.
4. The system acts from the transformed signal.
5. The operating envelope and failure boundary are measured.

The important language is surfaces, not proofs. Each system shows where the idea can be made practical, but not every system carries the same evidence weight.

The current scientific claim is deliberately narrow: in a controlled MuJoCo mirror-alignment experiment, a controller with no Cartesian target access aligns from sparse photometric feedback and reaches terminal accuracy comparable to a target-aware analytic baseline, while taking longer to acquire the target.

The broader program is staged, and the theorem posture has been deliberately demoted. Applications motivate controlled studies. Controlled studies earn stronger language. Failure-boundary coincidence, causal steerability, and preregistered nulls decide whether the language grows or shrinks.

We publish the repository, docs, workbenches, stress tests, and claim boundaries because the idea is only useful if it can survive inspection. A clean debunk path is not a threat to the project; it is the project doing its job.

Humility, for Sundog, is methodological. We do not assume full sight. We do not promote a broad theorem before the evidence earns it. We do not hide the boundary where the signal stops working.

• Sundog is not a port, fork, or p-system continuation of SunDog: Frozen Legacy. The Atari game and its preservation ecosystem are separate historical projects.
• Sundog is not a crypto project.
• Sundog is not a claim that indirect signals always beat direct state.
• Sundog is not a claim that agents should be denied information for aesthetic reasons.
• Sundog is not "better game AI" or deeper search. Game-facing workbenches matter only when they express hidden state, indirect signal, transformation, action, and failure boundary.
• Sundog is not a universal alignment proof. The defensible contribution today is an inspectable apparatus plus bounded results.

A sundog is not the sun. It is a displaced bright spot produced by geometry.

That is the point. The useful signal is not always the source itself. Sometimes the system cannot inspect the target directly, but the surrounding geometry still reveals something actionable.

The name is not nostalgia and not mysticism. It is a reminder that indirect light can still be information.