**1. Tail between my legs.**

Last year we made our second post to Less Wrong; a half-math, half-narrative note about how I perceived shadows as signals while working in controls and automation. Not machine learning. The seed came from a real job: we had to suspend acoustic trapeze for hanging hvac equipment that serves a quantum cloud. The tools we were using was primarily a telescoping 18 foot pole on the end of a Milkyway impact driver and a plumb lazer. If you put the pole over the lazer, you lost vision of your plumb dot painted on the ceiling. But there was a clue from the driver bit on the end of the 18' pole, a chrome plated socket driver that would scatter the lazer if the pole was positioned at a comfy angle while still being marginally vertical (probably about 22 degs). I got the *feel* for this procedure but then I had to explain it to my non-native language speaking help. How could I explain to them that when the alignment target was occluded by the task, they could carry on target confidently? That was the premise you heard about from a tutorial now digested and rewritten to a feral spiel attempting to describe the epistemological reframe that I could *taste*. The way in which the laser being passed over by the chrome driver would cause the light-painted dot to move between visual phases appearing to be like an hourglass then an arachnid then to a solid pair of halos then a narrowing halo. It was so vivid in the imagination it seemed obvious to me that everyone else should naturally track. The laborers immediately got it though, through the language-barrier. The chatbot took a bit of thrashing to reveal the math but H of X was tautologically solid and it really cleaned up gud.

Since then we developed a vocabulary to approach this concept from a software and computer science perspective. Sundog probably no longer an outsider theorem of alignment but a traceability harness for indirect-inference alignment, and most of the year was spent hiding from my original claim, keeping my day job. While quietly testing utility in client and pet-project software. That was until my lexical and technical proficiency could prompt engineer (or brute force) my way out of being haunted by the Solar optical halos all around me. This post-nut version of malice thought: "what kind of maniac would deploy sketchy promotional circuit based on the fact that he's either proven water is indeed wet, or debunked Goodhart, without knowing the difference?" So what remains I hope is available and legible for scrutiny and I invite the problems. ~(Bots seem to loose their edge when you're on the verge of scientific discovery so we're all just spit balling at this point. A scholastic parrot doesn't *know* the "new" thing, amirite?)

**2. The narrowed claim.**

The apparatus seems universally applicable given enough environmental "geometry" (((don't steal my patent))):

- the decisive state is hidden

- the world leaks structure via an indirect signal;

- the signal is transformed into a control-relevant signature;

- the system acts from the signature;

- the operating envelope and failure boundary are measured.

## The defensible question:

When direct inspection is unavailable, expensive, or occluded, or delayed, especially deliberately withheld, is the *trace* enough to make decisive actions **and** can we perceive when it is not?

**3. What the public repo contains.**

Receipts with boundaries. I started adulthood doing military PMCS, we do documents to justify documents to spec code to handoff engineering to outsource to fiverr for documents about what didn't work. Methodologish: If you read some of the experiment documents (and code) carefully, many are rigged with ticking time bombs to stamp the site with debunked if the gate is failed. I run half a dozen brands on mid scripts and a local orchestrator from minimax. Competitive, borderline malignant behavior between chatbots as groups of teams across sessions yields fun, fruitful and high fidelity results (it still cost $300/m for this after-work hobby). Running a solo model against itself in a few different context windows seems to devolve into a circle jerk of bugs. Everything is attacked by 4 teams, red team, blue, me pushy npm goey, and audit.

- Photometric Mirror Alignment (mujoco), the pole screw thing agent with no target position-access. Terminal accuracy is not detectably different from a target aware analytic baseline at n=30 (Mann-Whitney U=526, p=0.26). P=0.26 is the absence of evidence of a difference, not evidence of equivalence. The honest version is "not detectably different at n=30, ~16x slower to acquire." The 16x is not a footnote; it may be the actual result. The discipline rule for this project is that the 16x appears wherever the parity is mentioned.

- Geometry workbench (solar optical phenomena explained: halos, tangent arcs, CZA (halo smiley face over the sun or moon) cutoffs live). Separates rendered primitives from anchored inverse routes; the only promoted public inverse is the parhelion-offset handle on a strict three-photo subset, and the previously-drawn -0.05·R22 parhelic-belt rule was falsified (Spearman ρ ≈ 0.086) and retired across coupled surfaces. We think we're seeing the sundog hologram across neural networks, among other things (please send help).

- Mesa-Trap empirical front. Where we applied real selection pressure we did not find ~immunity. We found a sharp behavioral cliff at > λ ≈ 0.95–0.97 and localized the basin-attractor: in a trained controller's 256-unit final hidden layer it is not one neuron, not a

> handful of features, not any linear decomposition — **it is an entangled

> 5 Dimensional subspace** at net.7 (top-5 PCs ~ 97.4% across the cliff,

> both directions). A *located boundary* in a tested Small/Medium MLP

family, in vitro. The λ-confound control that would tell us whether λ is a real critical point or an optimizer artifact is still open; do not over-read the ~just cliff yet.

- Isotrophy sidecar for anniversary day (today) — a useful negative. In plain terms first: this is where we asked whether Sundog's whole "read the hidden geometry off an indirect signal" move generalizes to the three-body problem's symmetries. If it does, the count of valid "choreographies" (orbits where all three bodies follow one shared closed curve) should be predictable from symmetry alone. If it collapses to plain group theory, *that* boundary deserves to be public — which is what happened. After spending a few weeks chewing on planar and three-body dynamics I was staged to run an experiment that would take 60 days on my 2020 VR backpack style rig. When we packaged the thing to ask a resourced friend to fold it at their home, we found *Planar and three-dimensional periodic orbits for the Newtonian three-body problem* by Xiaoming Li and Shijun Liao. Instead of brute-forcing our eager grasp of the situation after thrashing on the three-body workbench, we found the professionals who had already discovered maintainable circuits from that one wire-maze toy from the doctor's office, and plotted it. A σ₃ detector reconciled their catalog of piano chords (mujoco taught us how to intuitively couple with a harmonic series): 21 strict single-curve equal-mass three-body choreographies (13 canonical + 8 opposite), plus 4 relative/rotating ones cleanly split off by a 2π/3 global rotation. 25 gauge-invariant vs the literature's 21, explained. That part is a durable detector / literature-count win, not theorem evidence. The proposed v0.2 daughter-count theorem test was then retired at its cheap K1 precheck: under v0.2 it reduced to the equivariance-only null (K_facet = 0; generically Z₃ ∩ Z₂ = {e}). I am not going to patch d_i to dodge the zero. K_facet = 0 does not mean "no piano-trios exist," and the 4 relatives are outside the strict single-curve convention, not literature errors.

Application surfaces (smaller workbenches) focused on mimicking common A.i demo gimmicks and an ambitious wild hair after reading that one book about aliens. Sharpened our grasp.

- Three-Body Solution, the near-escape pocket that feels ridiculous to approach but the chatbots basically one-shotted this while we've been thrashing for months on the geometry math v pictures workbench.
- Balance, the classic hold a spinning plate on a stick thing.
- Minesweeper, our highest fidelity Sundog Vs Bayes side by side.
- EyesOnly, a client project for a procedural roguelike videogame + tactical blue force tracker for ARG events that features a test harness to wrap api and run against the high score board and the sundogging manipulation agent that changes upcoming levels/cryptographic/landnav puzzles based on *perceived* human player/agent progress.
- Dungeon Gleaner, NPCs in a dungeon town drift between work, social, and errand spots by following the gradient of their own unmet needs. No scripted schedules, no behavior trees, no GOAP planner. A.i. Personality is per-archetype weighting on the same vocabulary of verbs. Reads organic from twenty feet away; cheap. "We're amazed how much you did within the game jam the world feels huge"
- Money Bags, softbody terrain rig interpreted through graph telemetry — torsion, deformation, symmetry, recovery — instead of trying to "look at" the simulation. Telemetry as the design surface.
- Ask Sundog, that little browser helper in the corner is a measured experiment too. 5,670 trace-conditioned trials across OpenAI/Anthropic/Meta builds with zero unsafe-accepts in the tested envelope. Mesa-adjacent, not solved alignment, bounded to that corpus and those models.
These are stress-test venues, not proofs; each has named failure cells published before you ask.

> Claude said I'm supposed to attach *Public docs + workbenches re-registered nulls + the start of a coarse-graining proof roadmap* here.

**FORE. The load-bearing postulate we are actually chasing.**

The Coarse-Graining Postulate, now staged in its own roadmap. A controller does not need to invert the world; it needs the signature to be a *sufficient statistic for the optimal action* — Blackwell sufficiency with respect to the control objective, not state reconstruction. Rate-distortion reading via the Shannon seed. Provable in toy LQG and finite-MDP cases where both Φ and π* are known; the falsifier is explicit. A task where the optimal policy is provably 𝓕_σ-measurable yet signature-only control fails to reach Bayes-optimal cost on that set kills it; the converse on the pushable-occluder boundary kills it too.

Two corollaries I want to name as conjectures, not results:

- Goodhart-immunity reframed as capacity-relative one-wayness of the reward channel. The agent can only game what it can invert; "immunity" is a relation between Φ and the agent's capacity class, not an attribute of the agent. The Mesa cliff is *predicted* by this reading rather than contradicting it — and it predicts the cliff location moves with capacity. Falsifiable.
- The 16x slowdown treated as a candidate conserved quantity rather than an embarrassment: (information withheld from the observation) × (excess acquisition time) ≥ const. Predicts the slowdown from the dimensionality gap across substrates. If the ratio is not conserved, the conjecture dies.

**5. What we got wrong, retired, or bounded this year.**

Can tacking the Sundog apparatus onto the Chinese guys' pretty plots of three celestial orbits give me the ez pk theorem proof I need to return to my day job or put on a lab coat? (But not both if I want to keep my wife). The v0.2 isotrophy daughter-count path is closed: K1 reduced to the equivariance-only null, recorded as a negative, not patched. v0.3 is load-bearing-up-next in principle, *not yet pre-registered or frozen*. The corrected foundation no longer treats α_I = ((12), T/2) as if it were already an orbit isotropy (that would repeat the v0.2 failure one level up). For rows where (12) maps the choreography to itself up to phase and spatial gauge, define G_i := ρ((12)) ∘ Φ_{T/2} and prove [M_i, G_i] = 0 before block-diagonalizing the monodromy; for rows where (12) maps it to a *different* equal-mass orbit, use an induced representation over the S₃ group orbit. The experiments cheap enough to run on my ancient REVIT VR walking rig pre-derivation case split has already run on both spatial parities (tau12_I and tau12_Z). The answer is unambiguous: **0 endomorphism cases, 21 induced-representation cases, 0 marginal**. So v0.3, if it continues obsessively tonight BECAUSE WE'RE SOOO CLOSE, is induced-representation-only across all 21 strict choreographies. The actual load-bearing deliverable now is the *written projector derivation*, not a run. If the derivation cannot be made clean, the honest move is to accept isotrophy as a clean negative plus the G.2 detector win and stop spending my precious tiny laptop fan ramp-up on the daughter-family claim. That fork stays open.

&& I pinky swear no more universal claims unless I'm fundraising from siblings.

Applications are surfaces, not proofs. The previous LessWrong attempt got

-9 and silence. I am not pretending the reception was warm. This site still has rough edges; some endpoints external crawlers see as dead. The recent IA cleanup pass closed the worst of those; not all. The failure boundary is the credential. We publish the places the traceability stops being enough and seek the path of most resistance.

**6. Why this might matter anyway.**
Partial observability is not an edge case. Real systems are occluded, delayed, noisy, sensor-limited, or deliberately withheld from privileged state. The conventional move is to demand more state; Sundog asks the narrower question and tries to make the answer inspectable. There are existing analogues that already do indirect alignment in narrow domains; plate-solving polar alignment, for instance, hits the polar axis without a direct line of sight to Polaris by recovering pointing from the field itself. I note that parallel without overstating it. It is the *kind* of move; Sundog is asking whether the kind generalizes once you write the boundary down.

**7. Open invitation, falsificationist.**

Year one of an apparatus, not year one of a finished result. The repo is

public because the only useful version survives contact with people

trying to break it. I mean if it's a thing, like, a man's gotta eat too

though.

Specifically asking for:

- replications or tighter same-information baselines on the photometric and Mesa fronts (a Bayes-optimal same-information baseline is the single result that would matter most);

- critiques of the coarse-graining sufficiency framing, especially the places where it might just be re-labeling "ill-posed inverse problems are hard";
- people who care about traceable evaluation harnesses under partial observability;

- folding at home? please help offload compute? I have a large capacity to brainstorm about things that kinda look like the thing in my dream a decade ago.

If the signature controller matches Bayes-optimal regret exactly where the postulate predicts it should and diverges where it should not — that is interesting — can we say interesting? Ok cool, ya it's interesting. If it does not, that is also interesting and should be written down b4 the cults grab it.

**Links.** Repo: https://github.com/humiliati/sundog · Site:

https://sundog.cc · Docs index: https://sundog.cc/docs/ ·

Coarse-graining proof roadmap:

https://sundog.cc/docs/COARSE_GRAINING_PROOF_ROADMAP.md


**8. Closing.** We started with a shadow and a torque reading (oh ya the original mujoco didn't even get torque and stuff the results were accidentally more narrow than stated. that's a whole another thing). We spent 7 months of the year trying to make the shadow legible enough for ourselves to attack and when we did boy it broke wide tf open. The claim is modest ;) some shadows are structured, some structures can be controlled from, and the boundary can be measured. \/\/e_h0p3_y0u_sP3nT_e><tra_t0k3Ns_UnFuCk1nG_tH1S Everything else is still on the table to be falsified.