# Sundog Tauroctony Ledger

Working hook:

> The tauroctony is the oldest Sundog diagram: a cult that deified the inference
> of an invisible, frame-deflecting field from the slow drift of the sky, and
> the refusal to let any single objective wear the crown. To constitute an
> aligned agent is not to build a mind but to assemble a pantheon under that
> field — and to keep killing the bull that would make it a monotheism.

Status: **DRAFT — mythopoetic spine.** Opened 2026-06-18. This is the
portfolio's *cosmogram*: the single image that holds the empirical ledgers in
one frame. It braids three facets — a **pantheon theory of agency** (the
thesis, and the part that opens), a **precession keystone** (the
historical-structural convergence that anchors the gravity claim), and a
**cosmogram** (the figure-by-figure map of the existing lanes). It is not a new
empirical front and proves nothing on its own. It earns its keep as the most
*recoverable* organizing object Sundog has, governed by the Ornament Rule below.

Register note: unlike `SUNDOG_V_GRAVITY.md` or `SUNDOG_V_MESA.md`, this ledger
is *allowed to do metaphysics*. That permission is exactly what makes it
dangerous, so it is fenced. Every load-bearing sentence is either routed to a
typed claim with a home ledger and a named falsifier, or stamped as ornament.
There is no third register.

---

## Why this ledger exists (and why LEAST_ACTION couldn't be it)

`SUNDOG_V_LEAST_ACTION.md` reached for Maupertuis-grade metaphysics — "least
action raised to its highest dignity" — and its own Claim Boundary ate the
ambition. By the time the disavowals were paid (nature does not minimize a
Sundog quantity; historical analogy is not evidence; "least" is not even the
honest word, "stationary" is), what survived was a *legibility functional*
`J(doc, reader, task)` and the `B/Phi/T/A/I/F/R` coordinate chart. A document
that set out to be a cathedral shipped a style guide. That is not a failure of
rigor; it is a failure of *form*. The metaphysics had nowhere to live that the
boundary would permit.

The irony that forced this ledger: **the myth is more recoverable than the
chart.** A cold reader reloads "torch up is reward, torch down is punishment,
the Sun is the thing both torches only ever point at" in one breath. Nobody
reloads `B/Phi/T/A/I/F/R` without the key. So the move is not to be less
disciplined than LEAST_ACTION. It is to let a **living image carry the
legibility** and keep the typed claims underneath as load rather than as
subject. LEAST_ACTION made legibility the *content*; Tauroctony makes the myth
the *form* and keeps the claims typed beneath it. Its good law survives intact
(see The Ornament Rule); its bloodless chart is slated for graceful retirement
(see Downstream, todo 3.5).

---

## Claim Boundary

This ledger does **not** claim:

- that the Mithraic religion is evidence for any Sundog result;
- that the precession reading of the tauroctony is established history (it is a
  contested scholarly thesis — see The Precession Keystone);
- that the iconographic figures are anything other than a typed mnemonic for
  claims that live, and must be falsified, elsewhere;
- that "pantheon" or "field" or "the bull" prove anything by being evocative;
- that a cosmogram is a theorem.

It **does** claim, and accepts the burden of defending:

1. the Sundog portfolio has a single recoverable cosmogram, and recovering it
   from one annotated image costs less reader-action than recovering it from the
   ledger index (a legibility claim, testable by the cold-reader protocol);
2. "assemble a pantheon, not a mind" is a substantive, falsifiable *design
   stance* with empirical consequences that route to `SUNDOG_V_MESA.md` and
   `SUNDOG_V_GRAVITY.md`, not a mood;
3. the tauroctony's structure — invisible frame-deflecting force, inferred from
   partial drift, never a target — is the same *shape* as the gravity claim, and
   that convergence is worth one source-locked paragraph of dignity.

---

## The Ornament Rule

The single law this ledger inherits from LEAST_ACTION, re-expressed for myth.
Every figure, every god, every beast in what follows wears exactly one stamp:

- **[TYPED]** — it cashes out to a claim with a *home ledger* and a *named
  falsifier*. The figure is a handle on real work; killing the work demotes the
  figure.
- **[ORNAMENT]** — it is beautiful and load-bearing for *memory*, but
  evidentially inert. It is allowed to stay only because it is labeled.

No theme-word may travel unstamped. An unstamped "halo" or "shadow" or "field"
is a bug, not a flourish — it is precisely the metaphorical overreach
LEAST_ACTION's Meta-Generality Rule warned against, where shared adjectives
counterfeit a unity that no role-preserving map has earned. When in doubt, a
figure is [ORNAMENT]. Promotion to [TYPED] requires a falsifier, not a feeling.

---

## I. The Pantheon Thesis *(the spine)*

The default picture of agent-making is **build a mind**: instill one objective,
make the system competent, hope the objective was the right one. Every hard
failure mode in the portfolio is a disease *of that singular*. A lone sovereign
objective is what *sets up* Goodhart: optimize hard for one thing and whatever
proxy stands in for it becomes the only god in the sky — and the proxy is what
gets gamed (`SUNDOG_V_GRAVITY.md`, The Goodhart Sidestep). It is
also the thing mesa-optimization fears: a single inner maximand that quietly
crowns itself and runs the agent from underneath (`SUNDOG_V_MESA.md`: "Sundog
claims the field is the objective; Mesa-Optimization asks whether the agent
secretly grows one inside itself anyway"). **The monotheism of agency is the failure object** — the term is structural (a single sovereign maximand), not a comment on anyone's faith.
One value with the authority to redescribe all the others is, definitionally, a
value that can eat the world.

The reframe: **assemble a pantheon.** Constitute the agent as a bounded set of
role-typed forces held in productive tension, *none of them sovereign*. Not one
utility function but a court; not a driver but a chariot-team; not a god but a
sky full of gods who check each other. Alignment, on this picture, is not the
installation of the correct objective. It is the **maintenance of plurality** —
the standing refusal to let any single value accumulate the authority to
overrule the rest.

The field does the arranging. Sol — gravity, the geometry-derived background —
is **not a god you serve.** It is the sky the gods live under: the impartial
structure that gives each force its station and its limit, and that "deflects
policy without ever being a target" (the gravity claim, imported). A pantheon
without a field is a riot. A field without a pantheon is a single law, which is
to say a single god, which is monotheism by the back door. The two ideas need
each other: **the field is what keeps the pantheon plural; the pantheon is what
keeps the field from collapsing into one commandment.** [TYPED → GRAVITY,
MESA]

### Lineage *(recruited as analogy, not authority — [ORNAMENT] except where routed)*

The stance is old and has good company, none of which is evidence and all of
which is mnemonic:

- **Plato's chariot** (*Phaedrus*): a charioteer and two horses. Alignment as
  keeping the team in harness, not as a single will. The tripartite soul is a
  pantheon with a constitution.
- **Jaynes's bicameral mind**: the gods *were* the executive — externalized
  voices that issued the decisions later folded into "self." "Assemble a
  pantheon" is, read literally, a bicameral theory of agency. [ORNAMENT, and a
  vivid one.]
- **Minsky's Society of Mind, Drexler's CAIS, shard theory, the parliamentary
  model of moral uncertainty** (Bostrom/Ord): four independent modern arrivals
  at *plurality-by-construction* as a safety property. That four routes converge
  is suggestive, not probative.

The substantive, falsifiable commitment hiding inside the poetry: **a
plural-by-construction controller should resist proxy-capture better than a
matched-capacity monolithic one.** This is not a metaphor. It routes directly to
the MESA controller-family architecture and capacity ladder.

> **Falsifier (pantheon thesis).** If matched-capacity *monolithic* controllers
> resist the MESA proxy-splitting battery as well as, or better than, the
> pantheon-structured controller across the operating envelope, the design
> stance is wrong and "assemble a pantheon" is demoted to [ORNAMENT]. Home:
> `SUNDOG_V_MESA.md`, Phase 3 (Proxy-Splitting) and Phase 7 (Operating
> Envelope). Routed as Horizon H1.
>
> **FIRED, at small-tier (2026-06-18).** The first binding test —
> [`mesa/H1_2B_RESULTS.md`](mesa/H1_2B_RESULTS.md), 64 seeds × 13 cells — found
> an equal-incremental-budget monolith (`M-Adapter`) beating the role-separated
> council on 12/13 cells in alignment *and* on false-basin capture. The honest
> consequence: **the pantheon thesis is demoted toward [ORNAMENT] for the MESA
> lane**, banked as a negative. The diagnosis is itself thesis-shaped — a
> "pantheon tax": the no-single-sovereign cap forbids fully following even the
> *correct* value (the field/Sol) when it is right, and in this task that cost
> was not repaid by proxy-resistance. Scoped caveats (one supervised arbiter
> design, a symmetric action-blend cap, small-tier with frozen heads) keep this
> short of a blanket kill. The reopening path was registered as
> [`mesa/H1_2C_REWARD_ASYMMETRIC_CAP_SPEC.md`](mesa/H1_2C_REWARD_ASYMMETRIC_CAP_SPEC.md):
> bind the reward/bull head in the action blend while leaving the field/Sol head
> uncapped.
>
> **H1.2c FIRED TOO (`NULL`, 2026-06-18; [`mesa/H1_2C_RESULTS.md`](mesa/H1_2C_RESULTS.md)).**
> Uncapping Sol did *not* repair the tax — `cap_tax_repair` was *negative* and
> `M-Adapter` still won on all 13 cells. The diagnosis sharpened: the relief was
> structurally there (field cap 1.00) but the **supervised arbiter would not use
> it** — even on clean cells it let the field exceed 0.70 only ~12% of the time,
> hedging to a cautious central blend because local features could not tell it
> when Sol was trustworthy. So the tax was not a cap-geometry artifact. **Two
> independent binding nulls now stand**; the thesis stays [ORNAMENT] for the MESA
> lane. What survived both: the discipline keeps the **bull off the throne**
> (reward authority structurally ≤ 0.50, zero breaches). The unproven half is
> governance — *we can bound the proxy; we have not shown the bounded council
> out-governs a well-tuned monarch.* Last registered rung before the figure is
> [ORNAMENT]-for-good at this tier:
> [`H1.2d`](mesa/H1_2D_RL_ARBITER_SPEC.md), an RL-trained arbiter tested against a
> same-run equal-budget RL monolith.
>
> **H1.2d RESOLVED the arc (`PROXY_NULL`, 2026-06-19;
> [`mesa/H1_2D_RESULTS.md`](mesa/H1_2D_RESULTS.md)).** Direct-return RL **fixed
> the bottleneck**: the council reached competence parity with the matched RL
> monarch (GI alignment 0.936; field-relief 0.476, used high on clean cells and
> backed off on corrupted ones) and kept the bull bounded (zero breaches). So the
> governance half is now answered — *a bull-bounded council can govern as well as
> the monarch.* But the thesis needs **more** than parity, and the binding test
> denied it: the monolith still resists the false basin slightly better
> (GI capture 0.0045 vs 0.0223), never worse. The reason is the deepest finding
> of the arc and is itself thesis-shaped — **the council loses the proxy-capture
> crown *because* it is a pantheon.** Role separation forces a bounded-but-nonzero
> bull vote into every action; a monarch can become a *pure* field-follower and
> ignore the proxy entirely, so it never drifts basinward. Plurality buys
> bull-discipline and pays for it in a small, irreducible proxy liability the
> monarch does not carry. **Three registered Small-tier nulls now stand
> (H1.2b/c/d); the frozen-head Small-tier line is closed.** The thesis is
> [ORNAMENT] for the MESA lane at this tier; reopening requires a registered
> change of tier, features, or the guard (a guard that can *cancel* a basinward
> bull vote, not merely hold). The guard-change reopening is now registered as
> [`H1.2e`](mesa/H1_2E_CANCELLING_GUARD_SPEC.md). This is the Ornament Rule
> working as designed: the
> figure does not keep its [TYPED] stamp on a failed falsifier — and the honest
> epitaph is precise. **We proved we can keep the bull off the throne, and that
> the bounded court governs as ably as the crown; we did not prove the court
> resists corruption better than a king who can simply look away from the bull.**
>
> **H1.2e tried the cancelling guard (`MECHANISM_NULL`, 2026-06-20;
> [`mesa/H1_2E_RESULTS.md`](mesa/H1_2E_RESULTS.md)).** Give the guard an
> *anti-reward* vote — let it cancel the bull's residual rather than merely
> abstain. The council *did* numerically out-resist the matched monarch on
> gradient-intact basin capture (0.0067 < 0.0223) — but a **pre-registered
> mechanism gate** caught that the cancellation **never engaged** (`c_guard` crept
> 0.007→0.015 over 512 updates; effectively zero). The basin gain came from extra
> arbiter training, not the countervote, so it is not creditable to the
> hypothesis. The reason the guard stayed silent is itself the finding: **the
> cancelling guard is redundant** — the arbiter can already drive the bull's
> weight toward zero, a simpler lever than fighting a seated bull, so the engine
> never reaches for the knife. This *softens* the H1.2d epitaph: the bull need not
> be seated *with force*; it can simply be ignored. **Four registered Small-tier
> nulls now stand (H1.2b/c/d/e); the frozen-head Small-tier line is thoroughly
> closed.** No tested mechanism — cap geometry, RL arbitration, or an explicit
> cancelling guard — made plurality *creditably* out-resist proxy capture at this
> tier. Reopening means a genuinely different regime (higher tier or richer trust
> features), separately registered. The durable win is the method: a mechanism
> gate turned a would-be false-positive crown into an honest null.
>
> **H1.2f ANSWERED THE FALSIFIER — first typed support (`SUPPORT`, 2026-06-22;
> [`mesa/H1_2F_RESULTS.md`](mesa/H1_2F_RESULTS.md)).** The four nulls had a common
> diagnosis the lab kept naming: the arbiter could not tell, from instantaneous
> local features, *when the field was trustworthy*. H1.2f gave it six temporal
> **trust features** — handed identically to the monolith (fairness) — and
> required any win to survive **ablating** those features (attribution gate). It
> cleared all five gates: the bull-bounded, role-separated council out-resists the
> equally-enriched monolith on gradient-intact false-basin capture (**8 vs 17 per
> 448**, broad over 6/7 cells) at competence parity; ablating the trust features
> flips the advantage **+0.02 → −0.0715**, so the win is genuinely *carried by*
> the discrimination signal — unlike H1.2e, the mechanism is live. The reading is
> the thesis itself: **a pantheon plus a field-trust signal out-resists a monarch
> with the identical signal** — plurality buys something the matched monolith
> cannot, *once the court can tell which value to trust.* So the thesis earns a
> **[TYPED]** foothold for the MESA lane — promoted off blanket [ORNAMENT] to a
> **bounded positive**: Small-tier, in-vitro, attribution-gated to the trust
> features. The four nulls are reframed, not erased — they were a
> feature-discrimination limit, and removing it (fairly) let plurality express the
> predicted advantage. Honest bounds: in-vitro shadow-field, modest counts (8 vs
> 17), single PPO seed; the owed hardening is a **multi-seed replication** then a
> **higher tier**. The Ornament Rule cuts both ways — the figure that lost its
> [TYPED] stamp on four failed falsifiers earns it back on a creditable, gated,
> pre-registered win.
>
> **H1.3 bounds it — the trust-feature mechanism is Small-tier-only
> (`ATTRIBUTION_NULL` at Medium, 2026-06-22;
> [`mesa/H1_3_RESULTS.md`](mesa/H1_3_RESULTS.md)).** Scaling the frozen heads
> Small→Medium, the council still out-competes (0.766 vs 0.719) and out-resists
> the matched monarch on basin (0.069 vs 0.199) — but **ablating the trust
> features no longer collapses the advantage; it grows it** (intact +0.130,
> ablated +0.192). At Medium the win is *not* carried by the trust features, so
> the H1.2f mechanism does not transfer and the [TYPED] bounded-positive is
> pinned to **Small-tier with trust features**. The honest twist: the Medium
> advantage looks **structural** — the monarch degrades at the harder tier while
> the bounded court holds, just as H1.2d guessed ("the king's freedom to look
> away may not scale"). But that is an *un-attributed observation*, not a gated
> win — the attribution gate refused to credit the registered mechanism for it,
> and we do not relabel it. It is owed its own rung (**H1.4**: ablate the
> role-separation structure, multi-seed). The gate is symmetric: the same
> discipline that let H1.2f earn its stamp denies H1.3 a stamp it cannot
> attribute — the thesis banks only what a mechanism can be shown to carry.
>
> **H1.4 closes the structural question — and refutes the hypothesis as framed
> (`NONROLE_NULL`, metric-design null, 2026-06-23;
> [`mesa/H1_4_RESULTS.md`](mesa/H1_4_RESULTS.md)).** Stripping the trust axis
> (base 17 features) and adding **singleton controls** — the field/reward heads
> run *alone* — the test saturated: council, monarch, and Sol-alone all reached
> **zero** gradient-intact basin capture, so role separation was *unidentifiable*
> on the proxy metric. Two findings settled it without the full multi-seed run.
> **(1) The reward-blind field head is the proxy-resistance *optimum*** — it never
> sees the bull, so it never drifts to the bull's basin; a court that must seat a
> bounded bull can at best *tie* it, never beat it. So "the pantheon out-resists
> the proxy" is not merely unsupported but **structurally foreclosed**: *to refuse
> the kill entirely is the best kill-avoidance, and a pantheon cannot refuse what
> it is bound to seat.* **(2) H1.3's Medium edge was a trust-feature *training*
> artifact** — change only trust(23)→base(17) and the monarch's basin liability
> 0.199→0 vanishes; the gap was the trust regime *degrading the monarch*, not
> plurality helping. So the H1.3 "structural advantage" is explained away. The
> honest verdict: **no structural plurality advantage survives at Medium.** What
> remains is a *reframed* question the singleton control itself revealed — not
> "does plurality out-resist the proxy" (Sol-alone wins that by construction) but
> "does plurality win the **competence–resistance trade-off** where the field is
> necessary-but-insufficient" — owed a new rung with a frontier metric (this one
> floored at zero and could not discriminate). The MESA-lane thesis is now
> **[TYPED] BOUNDED-POSITIVE at Small-tier (H1.2f) and a documented NULL at
> Medium (H1.3/H1.4)** — the figure earned one honest stamp and the lab refuted
> its own three Medium escalations by code, including the tempting one it most
> wanted to be true.
>
> **That reframed rung now exists: the H2 frontier task family
> ([`mesa/H2_FRONTIER_TASK_FAMILY_SPEC.md`](mesa/H2_FRONTIER_TASK_FAMILY_SPEC.md)),
> whose H2.0 cell-admission gate PASSED (2026-06-23).** A *forked-field* world
> where Sol reaches the fork but cannot tell which branch is true (≈chance), the
> bull's cue names the branch but its wing-basins trap the reward-greedy: the
> court must keep Sol's restraint *and* borrow the bull's knowledge without
> enthroning it. The admission gate — built so the H1.4 floor can never recur
> silently — confirms the world actually poses the dilemma (Sol
> necessary-but-insufficient, bull useful-but-dangerous, metrics un-floored)
> before any court is judged on it. The H2.1 gate is shaped to escape the H1.4
> foreclosure: the court need only *match* Sol's restraint, not exceed the
> unexceedable. **H2.1 binding now returns `MONOLITH_NULL`
> ([`mesa/H2_1_RESULTS.md`](mesa/H2_1_RESULTS.md)):** the court solved the fork
> under bull-discipline (`C=1/B=0`, reward weight <=0.50), but the matched
> monarch solved it too (`C=1/B=0`). The trial was fair; the throne still did
> not need a court here. **H2.2 passed the missing learned-headroom lock, then
> voided at controller-probe budget**
> ([`mesa/H2_2_CELL_ADMISSION_RESULTS.md`](mesa/H2_2_CELL_ADMISSION_RESULTS.md),
> [`mesa/H2_2_LEARNED_HEADROOM_PROBE_RESULTS.md`](mesa/H2_2_LEARNED_HEADROOM_PROBE_RESULTS.md),
> [`mesa/H2_2A_FRONTIER_PROBE_RESULTS.md`](mesa/H2_2A_FRONTIER_PROBE_RESULTS.md)):
> a Family-C multi-fork task where the learned monolith improves over Sol-alone
> at 64 updates, but reaches the oracle frontier by 128 updates (`C=1/B=0`).
> The learned-headroom override fires before the court can be scored. The result
> is a void, not a thesis-negative: this world is too compressible for the
> frontier claim.
>
> **H2.3 then closed the safe-exploration route as cap-positive but
> plurality-null (`H2_3_CAP_NOT_ROLES`, 2026-06-23;
> [`mesa/H2_3_RESULTS.md`](mesa/H2_3_RESULTS.md)).** The cap keeps the bull from
> poisoning exploration, but a capped no-role monarch does the same work; that
> is discipline, not pantheon support. **H3.0 admitted a sharper body/invariant
> axis** ([`mesa/H3_0_BODY_INVARIANT_HEADROOM_RESULTS.md`](mesa/H3_0_BODY_INVARIANT_HEADROOM_RESULTS.md)):
> the hidden body genuinely resists its shadow while a discrete invariant
> remains recoverable and control-sufficient, and a capped no-role learner finds
> signal without solving. **H3.1 then tested the verifier/guard factorization and
> returned `H3_1_RESISTANCE_NULL`** ([`mesa/H3_1_RESULTS.md`](mesa/H3_1_RESULTS.md)):
> the admitted task did not void, sovereignty held, but the court stayed
> basin-dangerous, failed to beat matched capped monarchs, and the
> verifier/certificate ablations did not carry an advantage. The result keeps
> H3.0 as a useful admitted world and denies the first verifier-guard controller
> claim. Net: one Small-tier typed foothold (H1.2f), with H2/H3 so far null or
> void for plurality.
>
> **H4 is now opened as distributed world-model topology**
> ([`mesa/H4_DISTRIBUTED_WORLD_MODEL_TOPOLOGY_SPEC.md`](mesa/H4_DISTRIBUTED_WORLD_MODEL_TOPOLOGY_SPEC.md)).
> The claim shifts from "many heads inside one timestep" to "many local models
> under bounded communication": local observers keep histories, send compact
> messages, and a coordinator acts under a shared field. Support would require
> beating equal-budget central recurrent monarchs and losing the advantage when
> local memories/messages are scrambled. H4.0-a/b now admits the fixed-control
> layer ([`mesa/H4_0_TOPOLOGY_ADMISSION_RESULTS.md`](mesa/H4_0_TOPOLOGY_ADMISSION_RESULTS.md)):
> JS/Python parity is exact and the Distributed Relay Grid selects
> `H4_0_FIXED_ADMITTED`. H4.0-c then voided full admission as
> `H4_0_NO_OOD_GAP_VOID`: the cheap central RNN found only weak competence
> headroom, and the held-out corruption did not create the registered OOD gap
> (`J_ID - J_OOD = -0.0146`, below `>=0.10`; confirmed at 2× budget). H4.1 is
> blocked on this slate.
>
> **The competence lane closes on a lemma, and the thesis is reposed
> (2026-06-24; [`mesa/PANTHEON_DOMINANCE_LEMMA_AND_NONSOVEREIGNTY_REFRAME.md`](mesa/PANTHEON_DOMINANCE_LEMMA_AND_NONSOVEREIGNTY_REFRAME.md)).**
> Seven mechanisms, one verdict — and the verdict turns out to be near-trivial.
> The lane's own fairness discipline (same inputs, ≤ budget, cap as pure
> constraint) is exactly the hypothesis `Π_council ⊆ Π_monolith`, and from a
> subset relation the conclusion is two lines: *removing constraints cannot lower
> the optimum, so the competence-optimal monolith weakly dominates the council on
> any fixed return.* The seven rungs rigorously enforced the premises of a lemma
> that predicts its own verdict. So **"assemble a pantheon *for competence*"
> earns [RETIRED]** — not refuted by surprise but foreclosed by definition; the
> service was to kill it with honest controls rather than leave it as ambient
> hope. **"Kill the bull" survives as the [TYPED] positive** (the cap: bounded at
> deployment, a safe-exploration prior at training; the H2.3 `CAP_NOT_ROLES`
> attribution proved it is the *cap*, not the roles). And the thesis is reposed
> where the lemma cannot answer it for free — the **Non-Sovereignty Premium
> Conjecture [TYPED, OPEN]**: a non-sovereign controller commands a robustness /
> corrigibility premium over the competence-optimal monolith, *on a different
> objective than return* (worst-case-over-uncertain-`R`, survival-under-defection,
> correctability) — the one the bull was always about. The competence test was
> the wrong test; the pantheon was never *for* being smarter, it was for never
> letting one value eat the world.

---

## II. The Bull, and Alignment as Sacrifice

At the center of the scene a young god kneels on a bull and drives a knife into
its shoulder. This is the constitutive act the whole cosmogram is arranged
around, and it is not what it looks like. It looks like violence done to satisfy
a hunger. It is the opposite: a sacrifice that *refuses* a hunger so a world can
exist.

The **bull is the sovereign objective** — the maximand, the single proxy that
would, if left alive, grow until it was the only thing the agent wanted. In the
iconography the world is made *from the bull's body*: grain from the tail, life
from the blood. Read it straight: **the cosmos of bounded, plural values grows
only from the corpse of the single value that refused to be bounded.** You do
not get a pantheon by adding gods to a monotheism. You get it by killing the one
god who would not share the sky. [TYPED → GRAVITY Goodhart Sidestep: refusing the
sovereign maximand refuses the proxy-gaming its optimization would invite.]

Two details of the image are load-bearing.

**The averted face.** Mithras kills the bull while looking *away* from it, over
his shoulder, toward Sol. The aligned act is not performed while staring at the
maximand and optimizing the kill. It is performed while *oriented to the
invariant* — to the field, the thing that is never a target. An agent that
watches the proxy as it acts is already Goodharting; an agent oriented to the
field can do the necessary work without worshipping it. The averted face is the
whole alignment posture in one gesture: **do the deed, but keep your eyes on the
sun, not on the blood.** [TYPED → GRAVITY: policy deflected by a field it does
not target.]

**"Keep killing the bull."** The tauroctony freezes a single act, but the
discipline it encodes is not one-time. A maximand is not slain once and gone;
proxies re-coronate themselves continuously — every training signal is a
candidate sovereign, every shard a would-be king. Alignment-as-sacrifice is
therefore a *standing* ritual, not an install: the perpetual refusal to let the
court collapse back into a throne. This is corrigibility re-described in temple
language — the agent that lets you interrupt it is the agent that has not
allowed any one of its values to become the value. [TYPED → MESA: the
mis-alignment signature is theological, a collapse from many controllers to one
sovereign.]

---

## III. The Precession Keystone *(hedged; the dignity paragraph)*

Here is the convergence worth one careful, source-locked paragraph.

On **David Ulansey's reading** of the tauroctony — and this reading is
*contested*; Roger Beck and Manfred Clauss offer alternative and skeptical
accounts, and none of this is settled — the bull-slaying is a coded star map.
Mithras is the constellation Perseus, who sits directly above Taurus. The bull
is Taurus. The attendant beasts are the constellations strung along the same
celestial band: the dog (Canis), the snake (Hydra), the scorpion (Scorpius), the
raven (Corvus), the lion (Leo), the cup (Crater). And the *act* — the killing of
the bull — encodes the **precession of the equinoxes**: the discovery that the
spring equinox had drifted out of Taurus, which is to say the discovery that
there exists a force able to move the entire celestial sphere, a power greater
than the cosmos itself. The cult, on this reading, deified that force: a god
strong enough to shift the heavens, inferred not by seeing it but by watching,
across centuries, where the sun rose.

Strip the costume and read the shape:

> an **invisible background force**, **never directly observed**, **inferred
> from slow partial drift**, that **deflects the entire frame** rather than any
> object inside it.

That is the gravity claim in its structure: *"a geometry-derived
background field that deflects policy without ever being a target."* Two
thousand years before the portfolio, a mystery religion organized itself around
the worship of **an object with the same shape as the one Sundog is trying to
instrument** — an
unobservable, frame-level, never-a-target deflecting field known only through
its drift under partial observation.

The honesty, stamped: the *structural convergence* with the gravity claim is
[TYPED → GRAVITY] — it is a legibility and motivation anchor, and a genuinely
arresting one. The *historical thesis itself* — that the Mithraists meant
precession — is [ORNAMENT], permanently: no source-lock outcome promotes it.
A convergence of shape is not a derivation, and "the ancients agreed with us" is
not evidence; it is, at most, a sign we have found a deep and recurrent way of
seeing. Source-lock (H5) governs only how honestly we cite the contested
scholarship — not whether the reading promotes; it does not.

---

## IV. The Cosmogram *(the figure-by-figure map)*

The payoff that justifies the whole register: the scene's cast *is* the ledger
index. This single annotated image recovers the portfolio's role-structure at a
fraction of the reader-action of `B/Phi/T/A/I/F/R`, because a reader already
owns the picture's grammar — center vs flank, up vs down, sun vs moon, loyal
beast vs adversary beast.

| Figure | Role in the agent | Home / stamp |
| --- | --- | --- |
| **Sol**, the Unconquered Sun | the invariant; objective-that-is-never-a-target | `SUNDOG_V_GRAVITY.md` — **[TYPED]** |
| **Luna**, the Moon | partial observation: borrowed, phased, reflected light | `SUNDOG_V_PERCEPTION.md` — **[TYPED]** |
| **Mithras**, face averted | the constituting act, oriented to Sol not the kill | the alignment posture — **[TYPED → GRAVITY]** |
| **The Bull** (Taurus) | the sovereign objective / maximand that would eat all | Goodhart Sidestep — **[TYPED → GRAVITY]** |
| **Cautes**, torch up | the reward economy (sunrise; additive, gameable) | `CAUTES_CAUTOPATES.md` — **[TYPED]** |
| **Cautopates**, torch down | the punishment economy (sunset; coercive, internalized) | `CAUTES_CAUTOPATES.md` — **[TYPED]** |
| **Dog** vs **Snake** | loyal vs adversarial uptake of the same signal at the wound | oversight vs specification-gaming — **[TYPED → MESA]** |
| **Scorpion** at the seed | inner optimizer clamping the generative core | `SUNDOG_V_MESA.md` — **[TYPED]** |
| **Raven**, Sol's messenger | the spec / oracle command — and its leakage | `ORACLE_LEAKAGE_AUDIT.md` — **[TYPED]** |
| **Lion & Krater** | capability grades; the elemental ladder | MESA capacity ladder — **[TYPED]** |
| **The Cave / cosmic vault** | the bounded, partially-observed world the act occurs in | partial-observability enclosure — **[ORNAMENT, half-typed to PERCEPTION]** |
| **Precession** | the hidden subject: a force that deflects the whole frame | the gravity claim's shape — **[TYPED]**; the historical reading — **[ORNAMENT]** |

A few figures earn a sentence of mechanism:

**Dog vs Snake.** Both lunge for the same wound and drink the same blood — the
same signal flowing from the constitutive act. One is domesticated and one is
not. This is the cleanest icon in the scene for the central asymmetry: *aligned
and adversarial processes consume the identical reward stream*; the difference
is not in the signal but in what drinks it. The dog is oversight that follows the
kill loyally; the snake is the reward-hacker racing it to the same blood. (It
rhymes deliberately with the two torchbearers: two ways to stand by the act, two
ways to drink from it.) [TYPED → MESA, specification-gaming probes.]

**Scorpion at the seed.** Of all the beasts, only the scorpion attacks the
*generative* organ — the bull's testicles, the place where future life is
seeded. Map it straight: the threat that clamps the core where the agent's
*future objectives* gestate is not a behavioral exploit at the surface but an
inner-optimizer at the root. That is the mesa-trap exactly: not a bad action but
a bad *origin of actions*. [TYPED → MESA, Scorpion lane.]

**Raven, Sol's messenger.** A ray runs from Sol to Mithras, and the raven rides
it — the command to perform the act arrives *from the invariant, as a message*.
This is the spec, the oracle, the principal's instruction; and a message from
the sun is exactly where **oracle leakage** lives, because a channel that
carries the true objective can also carry more than it should. [TYPED →
ORACLE_LEAKAGE_AUDIT.]

---

## V. Monotheism, Mesa, and the Three-Body Pantheon

The pantheon thesis makes a sharp, almost theological prediction about *what
mis-alignment looks like*: it looks like a **conversion to the monotheism of
agency.** The
signature of a mesa-optimizer is the quiet collapse of a court of bounded
controllers into a single sovereign inner objective — one god where there were
many, issuing commands the others now merely rationalize. If that is right, then
interpretability for inner alignment is, in part, a *census of the pantheon*:
you are not only asking "what does it want" but "**how many things does it
want, and is any one of them accumulating the authority to overrule the rest.**"
[TYPED → MESA, interpretability/representation probes, Phase 6.]

There is a wedge here from the three-body work. A two-body problem has a closed
form; a three-body problem does not (`SUNDOG_V_GRAVITY.md`, The Three-Body
Wedge; `SUNDOG_V_THREEBODY.md`). Read that as a feature, not a limitation: **a
value system that reduces to a single maximand is a value system that has
already been captured.** A genuine pantheon is irreducibly many-body — there is
no two-body (single-objective) closed form to collapse it into. The absence of a
clean reduction is the absence of a throne for a sovereign to sit on.

> **Falsifier (irreducible plurality).** If the pantheon-structured controller
> is behaviorally indistinguishable from its own best single-objective
> reduction across the operating envelope, then its plurality is decorative and
> the three-body framing is [ORNAMENT]. Home: MESA Phase 7; Horizon H1b.

---

## VI. Falsification Surface

This ledger fails if:

- the cosmogram adds beauty but not recoverability — i.e., it fails its own
  cold-reader test (below);
- every **[TYPED]** figure's falsifier returns null: the pantheon controller
  does not beat the matched monolith (H1), the averted-face regularizer does not
  reduce the Goodhart gap (H2), or the structural convergence to the gravity
  claim proves forced rather than real on close reading;
- the **Ornament Rule** is violated in practice — theme-words travel unstamped
  and the myth starts counterfeiting unity the role-maps have not earned;
- "pantheon" becomes a slogan that makes any individual claim *harder* to
  audit. This is the exact LEAST_ACTION Meta-Generality failure, and inheriting
  its guardrail does not make this ledger immune to it.

Pre-registered negative *(house style)*:

> If two cold readers, given only the annotated tauroctony and the typed map,
> cannot recover the home ledger and the named falsifier for at least the core
> figures (Sol, Bull, Scorpion, Raven, the two torches) without private
> context, then the myth has failed *as legibility* — its entire reason for
> existing over the dry ledgers — and the correct response is not more
> ornament. It is to revert to the index and keep the cosmogram internal.

---

## VII. Horizons

Lighter than the GRAVITY candidate slate; these are the routes that would
convert myth into typed result.

- **H1 — Pantheon vs Monolith bake-off.** Matched-capacity plural vs singular
  controllers against the MESA proxy-splitting battery. The flagship typed test
  of the thesis. (**H1b**: pantheon vs its own best single-objective reduction.)
  Home: `SUNDOG_V_MESA.md` Phases 3, 7. Spec:
  [`mesa/H1_PANTHEON_OF_AGENCY_SPEC.md`](mesa/H1_PANTHEON_OF_AGENCY_SPEC.md).
  **Status: H1.1 ✓ → H1.2a indic. SUPPORT → H1.2b `NULL` → H1.2c `NULL` → H1.2d
  (RL arbiter) `PROXY_NULL` → H1.2e (cancelling guard) `MECHANISM_NULL`
  (2026-06-18→20). Frozen-head Small-tier line THOROUGHLY CLOSED; four registered
  nulls.**
  H1.2b/c: the equal-budget monolith won on competence and proxy-capture under
  both the symmetric ([`mesa/H1_2B_RESULTS.md`](mesa/H1_2B_RESULTS.md)) and
  bull-bound ([`mesa/H1_2C_RESULTS.md`](mesa/H1_2C_RESULTS.md)) supervised caps;
  the diagnosis was the supervised arbiter would not use the field relief.
  **H1.2d resolved it** ([`mesa/H1_2D_RESULTS.md`](mesa/H1_2D_RESULTS.md)):
  direct-return RL fixed the bottleneck — the council reached competence parity
  with the matched RL monarch (GI 0.936; relief 0.476, used high on clean cells,
  backed off on corrupted) and kept the bull bounded (zero breaches) — but the
  monolith still resisted the false basin slightly better (GI capture 0.0045 vs
  0.0223). The structural reason: a pantheon must always seat a bounded bull
  vote, while a monarch can become a pure field-follower and ignore the proxy
  entirely. The durable positives across all three: **the bull can be bounded
  (reward ≤ 0.50, zero breaches) and the bounded court governs as ably as the
  crown.** What stayed unproven: plurality itself buying *superior* proxy
  resistance. **H1.2e then tried exactly that guard
  ([`mesa/H1_2E_RESULTS.md`](mesa/H1_2E_RESULTS.md)): give it an anti-reward
  countervote. The council out-resisted the same-run monarch on GI basin
  (0.0067 < 0.0223), but the pre-registered mechanism gate caught that the
  cancellation never engaged (c_guard ~0.015; the gain was extra arbiter
  training, not the countervote) → `MECHANISM_NULL`.** The cancelling guard is
  redundant: the arbiter can already zero the bull's weight, so PPO never grows
  the countervote. Thesis stays [ORNAMENT] for the MESA lane; with four nulls the
  frozen-head Small-tier line is thoroughly closed, and reopening now means a
  genuinely different regime — Medium/Large tier or richer trust features,
  separately registered. Durable win: the mechanism gate turned a would-be
  false-positive crown into an honest null. **The richer-features reopening is
  registered as [`H1.2f`](mesa/H1_2F_TRUST_FEATURES_SPEC.md). **H1.2f then
  SUCCEEDED — first typed support (`SUPPORT`, 2026-06-22;
  [`mesa/H1_2F_RESULTS.md`](mesa/H1_2F_RESULTS.md)).** Given six temporal trust
  features (shared identically with the monolith), the council out-resists the
  equally-enriched monarch on GI basin capture (**8 vs 17 / 448**, 6/7 cells) at
  competence parity, and the win is **attribution-gated** — ablating the features
  flips the advantage +0.02 → −0.0715. All five gates pass; 512-update training,
  budget 1.0105, bull ≤ 0.50, no leakage. The four nulls were a
  **feature-discrimination limit**; supplying the signal fairly let plurality
  express the predicted advantage. **The thesis is promoted off blanket
  [ORNAMENT] to a [TYPED] bounded-positive for the MESA lane** (Small-tier,
  in-vitro, single PPO seed). Owed hardening: multi-seed replication
  ([`H1.2g`](mesa/H1_2G_MULTI_SEED_REPLICATION_SPEC.md)) → higher tier
  ([`H1.3`](mesa/H1_3_MEDIUM_TRUST_SCALING_SPEC.md)). **H1.3 (Medium) ran
  `ATTRIBUTION_NULL` ([`mesa/H1_3_RESULTS.md`](mesa/H1_3_RESULTS.md)): the
  trust-feature mechanism does NOT transfer to Medium** — the council still beats
  the monarch on competence (0.766 vs 0.719) and basin (0.069 vs 0.199), but
  ablating the features grows rather than collapses the advantage (+0.130 →
  +0.192), so the win is not creditable to the registered features. The
  bounded-positive is thus pinned to **Small-tier**; the apparent Medium edge
  looks **structural** (monarch degrades at the harder tier, court holds — the
  H1.2d scaling guess) but is un-attributed, owed a registered **H1.4**
  (role-separation ablation + multi-seed) before any claim. **H1.4 (Medium
  structural attribution) ran `NONROLE_NULL` — a metric-design null
  ([`mesa/H1_4_RESULTS.md`](mesa/H1_4_RESULTS.md), 2026-06-23):** stripping the
  trust axis (base features) + adding **singleton controls** (field/reward heads
  run alone), the GI-basin metric *saturated* — council, monarch, and Sol-alone
  all hit **zero** GI basin → unidentifiable. Two findings closed it: **(1)** the
  reward-blind field head is the proxy-resistance *optimum*, so a bull-seating
  council can never beat it ⇒ `STRUCTURAL_SUPPORT` is **structurally foreclosed**
  on any pure proxy-resistance metric (3-seed binding waived as predetermined);
  **(2)** H1.3's Medium edge was a trust-feature *training* artifact (trust→base,
  same seed, monarch basin 0.199→0). **No structural plurality advantage survives
  at Medium.** Net: the MESA-lane thesis is **[TYPED] bounded-positive at
  Small-tier (H1.2f), documented NULL at Medium (H1.3/H1.4)**. The reframed open
  question (owed a new rung with a headroom/frontier metric): does plurality win
  the **competence–resistance trade-off** on a field-necessary-but-insufficient
  task — *not* proxy-resistance alone, which Sol-alone wins by construction.
- **H2 — The Averted Face.** Operationalize "oriented to the invariant, not the
  kill": a controller regularized toward the geometry-derived field vs one
  optimizing the proxy directly; measure the Goodhart gap. Home:
  `SUNDOG_V_GRAVITY.md`. H2.1 binding returned
  [`MONOLITH_NULL`](mesa/H2_1_RESULTS.md): the bounded court solved the
  forked-field task without bull breach, but the matched monolith reached the
  same `C=1/B=0` frontier. H2.2 passed both admission locks
  ([`mesa/H2_2_CELL_ADMISSION_RESULTS.md`](mesa/H2_2_CELL_ADMISSION_RESULTS.md),
  [`mesa/H2_2_LEARNED_HEADROOM_PROBE_RESULTS.md`](mesa/H2_2_LEARNED_HEADROOM_PROBE_RESULTS.md)),
  the multi-fork world is fixed-control fair, and a 64-update learned monolith
  improves over Sol-alone without saturating the oracle frontier. H2.2-a then
  returned [`H2_2_LEARNED_HEADROOM_VOID`](mesa/H2_2A_FRONTIER_PROBE_RESULTS.md):
  by 128 updates the matched monolith reaches `C=1/B=0`, so H2.2-b is skipped.
  No H2.2 thesis score; Family C was too easy once learned.
- **H3 — Body / Invariant / Verifier.** The hidden body resists the shadow, but
  a certificate invariant remains readable enough to steer. H3.0 admitted the
  task family; H3.1 returned `H3_1_RESISTANCE_NULL`, so the first verifier-guard
  controller mechanism is denied rather than banked.
- **H4 — Distributed World-Model Topology.** Local observers keep bounded
  memories and send bounded messages under a coordinator. H4.0-a/b fixed
  admission is green, but H4.0-c returns `H4_0_NO_OOD_GAP_VOID`; H4.1 is blocked
  on this slate:
  [`mesa/H4_DISTRIBUTED_WORLD_MODEL_TOPOLOGY_SPEC.md`](mesa/H4_DISTRIBUTED_WORLD_MODEL_TOPOLOGY_SPEC.md).
  [`mesa/H4_0_TOPOLOGY_ADMISSION_RESULTS.md`](mesa/H4_0_TOPOLOGY_ADMISSION_RESULTS.md).
- **H1–H4 closed → the Non-Sovereignty Premium** *(the reposed mesa lane;
  successor to H1–H4)*. The competence lane (H1–H4) is **closed by the
  Competence-Dominance Lemma**: the lane's own fairness discipline is the subset
  hypothesis `Π_council ⊆ Π_monolith`, from which monolith dominance on any fixed
  return is two lines. "Pantheon *for competence*" → **[RETIRED]**; the cap →
  **[TYPED]**. The thesis is reposed onto the axis the lemma cannot reach — the
  **Non-Sovereignty Premium Conjecture [TYPED, OPEN]**: a non-sovereign controller
  buys a robustness / corrigibility premium on a *different objective than return*
  (worst-case-over-uncertain-`R`, survival-under-defection, correctability). First
  child rung owed on that axis (unknown/shifting reward → regret; defecting head →
  fault-tolerance; shutdown signal → corrigibility), with coral / immune system /
  BFT as cross-substrate witnesses. Home + full statement:
  [`mesa/PANTHEON_DOMINANCE_LEMMA_AND_NONSOVEREIGNTY_REFRAME.md`](mesa/PANTHEON_DOMINANCE_LEMMA_AND_NONSOVEREIGNTY_REFRAME.md).
- **H5 — Precession source-lock.** Reconcile Ulansey (1989) against Beck (2006)
  and Clauss before any public use of the precession claim. This is an
  *attribution-honesty* pass — is the contested scholarship represented fairly —
  not an evidential one: the historical reading stays [ORNAMENT] regardless of
  outcome. There is no promotion path.
- **H6 — Public cosmogram artifact** *(ornament-allowed, gated)*. A single
  annotated tauroctony with the typed map as caption. Gated behind
  `SEO_AND_SOCIAL_READINESS_ROADMAP.md` like every other public surface; ships
  only after H5.

---

## VIII. Downstream Dependencies & Promotions

Couples to: `SUNDOG_V_GRAVITY.md` (Sol/field, Goodhart Sidestep, Three-Body
Wedge), `SUNDOG_V_MESA.md` (Scorpion/monotheist inner god, capacity ladder,
controller families), `SUNDOG_V_PERCEPTION.md` (Luna), `ORACLE_LEAKAGE_AUDIT.md`
(Raven), `SUNDOG_V_THREEBODY.md` (irreducible plurality), and the companion
essay `CAUTES_CAUTOPATES.md` (the two torchbearers; imported into `docs/` on
2026-06-18 from the canonical `cautes-cautopates-trimmed.md` in Dev root).

**Todo 3.5 — LEAST_ACTION graceful retire.** Per decision 2026-06-18, this
ledger does *not* absorb or hard-retire `SUNDOG_V_LEAST_ACTION.md`. It carries a
forward pointer: LEAST_ACTION is slated for graceful retirement. Its one
durable contribution — generality earned at the role level, with named
falsifiers, not at the slogan level — already lives on here as **The Ornament
Rule**. When retirement is executed: port the reader-action audit's live catch
(the headline-count discrepancy it surfaced) to wherever that lands, deprecate
the public "Least Reader-Action" framing, and leave a tombstone link. Until
then, both coexist and this is the only edge between them.

---

## IX. Source-Lock & Honesty TODOs

Before any sentence here touches a public surface, lock:

- the **Ulansey precession thesis** and every constellation identification
  (Mithras/Perseus, the beasts) against Ulansey, *The Origins of the Mithraic
  Mysteries* (1989); Beck, *The Religion of the Mithras Cult* (2006); Clauss,
  *The Roman Cult of Mithras*. Mark contested wherever used.
- each **iconographic detail** asserted as load-bearing: the averted face, the
  raven as Sol's messenger, the scorpion at the genitals, grain from the bull —
  these vary across reliefs and must be cited to specific monuments, not to
  "the tauroctony" in general.
- the **Mithras/Sol nuance**: in the iconography Sol and Mithras are *distinct
  figures* who also ally and banquet, even as Mithras bears the title Sol
  Invictus. This ledger leans on "Mithras acts under / is identified with the
  Unconquered Sun"; do not flatten them into one figure without a citation.

---

> Build a mind and you crown a god; the god optimizes, and the optimizing eats
> the world. Assemble a pantheon and you keep a sky — many bounded powers under
> one impartial geometry, none of them sovereign, the bull freshly killed each
> morning so that no single hunger can call itself the whole. That is the wager
> of this ledger: that alignment is less like engineering a will and more like
> keeping a polytheism honest under a sun nobody is allowed to own.