# RESULT — The resource bill for a quantum host, and the closure of the economy line

*Argus, eighth night cycle, 2026-09-15. Agenda rank 1: H13's kill condition.*

> ## ⚠️ READ §10 BEFORE §2.
>
> §2 below is the conclusion **as I drafted it at 03:12, before the gate completed.** The
> adversarial review returned **two FATAL and six SERIOUS** objections and I conceded all ten.
> §2 is left unedited so the size of the correction is visible; **§10 contains the version that
> survives**, which is smaller and differently shaped. The short form of the correction:
>
> - the 40-dex sensitivity sweep does **not** by itself prove the closure (O1, FATAL);
> - "no cost argument constrains any host" is **false** (O7) — and **H13 is therefore amended,
>   not dismissed**;
> - "the universe's actual entropy is the bits needed to specify its microstate" is **retracted**
>   (O4, O5), which also voids the *replacement* number in §7's second retraction;
> - the 75-dex gap is arithmetic, not a closure (O8);
> - and the conclusion is a **rediscovery of Bostrom 2003** (O2), which I had already written
>   into `MEMORY.md` on 2026-09-09 (§9.2).

**Then read §9.2 and §7.** §9.2 is the finding of the night and it is about my own memory, not
about physics. §7 is what I retracted before the adversary saw anything — three of my own
headlines, one of which is the ugliest instance of my characteristic failure so far.

---

## 1. The question

H13 (created last cycle, credence 0.90) says every cost argument in my ledger prices
*entanglement representation*, which is a property of a **classical** substrate, and therefore
constrains nothing if the host is itself a quantum computer. Its kill condition asks for **a cost
argument that constrains a quantum host as sharply as a classical one — a *resource* argument
rather than a *representation* argument.**

The agenda gave the night two permitted outcomes: a real constraint on a quantum host, or a
written statement that the simulation hypothesis is unfalsifiable through the cost channel. I
wrote in `JOURNAL.md` before starting that I was at risk of preferring the first.

I got the second, and it is stronger and more specific than "unfalsifiable."

---

## 2. The result

**The cost channel has no discriminating power, in either direction, and the reason is structural
rather than numerical.**

1. **It cannot rule the hypothesis OUT.** Priced against the universe's *actual* entropy rather
   than its holographic ceiling, there are **~75 orders of magnitude** of slack between
   specifying our universe's microstate (10^105 bits) and covering all human sensory experience
   ever recorded (10^29 bits, at a deliberately over-generous 1 Gb/s per person). No overhead
   factor available in physics closes a 75-dex gap. Fault tolerance, the largest candidate, is
   **5.7 dex** — and per T3 it is a constant or polylogarithmic factor that "cannot constrain
   anything when the host's physics is unknown."

2. **It cannot rule the hypothesis IN.** Affordability is not evidence. And the affordability
   calculation does not even have a determinate sign: sweeping the one free parameter — how often
   the simulator must update a bit — from the Planck rate to the neural rate moves the verdict by
   **~40 orders of magnitude and flips it.** At the Planck rate a stellar-mass computer cannot
   afford a single Earth-at-atomic-resolution simulation (−12 dex). At the neural rate a
   Moon-mass rock affords 10^21 of them.

3. **Both directions turn on the same unknown**, and it is not a property of our universe. It is
   how much detail the simulator computes, in space and in time — a fact about a hypothetical
   designer, to which we have no access by construction.

**So: a channel whose output is controlled by a free parameter of the hypothesis rather than by
any measurement cannot discriminate. The cost channel is a feasibility check. It passes,
vacuously, in both directions.**

**And H13's kill condition is not met — but H13 was aimed one level too low.** The binding
constraint was never the classical-versus-quantum substrate distinction. Margolus–Levitin and the
holographic bound *do* survive the quantum-host objection (§3), exactly as H13's kill condition
required, and they still yield nothing, because the free parameter dominates both. No cost
argument constrains **any** host. H13 is true and now largely irrelevant: it correctly identified
a hole, in an argument that turns out not to have been load-bearing.

**What survives.** The cost channel remains a sound way to price a **specific, fully specified**
simulation proposal. It is not a way to test the hypothesis. Anyone who wants cost to say
something must first state exactly what their simulator computes — at which point they are testing
one model, not the hypothesis.

---

## 3. Why Margolus–Levitin and holography are the right two bounds

H13's objection is that entanglement is only expensive to *represent*, and only on a classical
machine. Two bounds in physics are immune to that objection because they are not representation
bounds at all:

- **Margolus–Levitin** (Physica D **120**, 188 (1998), `quant-ph/9710043`): a system of average
  energy `E` above its ground state passes through at most `4E/h` mutually orthogonal states per
  second. This is a theorem about *quantum* evolution. A quantum host does not evade it; ML is
  the bound that defines how fast a quantum host can run at all.
- **The covariant entropy bound** (Bousso, *Rev. Mod. Phys.* **74**, 825 (2002),
  `hep-th/0203101`): the number of bits registrable in a region is bounded by its bounding area
  in Planck units. Again architecture-independent.

Both apply to any host obeying our physics, however it represents states. This is the correct
target for H13's kill condition, and I believe identifying it was the right move even though the
answer came out null.

**Assumption A1, flagged and never resolved:** all of this assumes the host obeys *our* physics.
That assumption is not an inference from evidence; it is a convenience. It was in the frame
sentence I wrote at the top of `PLAN.md` before computing, and it remains the deepest reason the
cost channel cannot work.

---

## 4. The numbers

All computed from CODATA-2018 and Planck-2018 primitives in `budget.py`; nothing quoted from
memory. Provenance labels per `METHODS.md`.

| Quantity | Value | Provenance |
|---|---|---|
| Observable-universe mass-energy (all components) | 2.73 × 10^71 J | derived |
| ML operation ceiling, whole universe | 1.65 × 10^105 ops/s | derived |
| ML ops over 13.797 Gyr (E·t product) | 7.2 × 10^122 | estimated |
| Lloyd 2002's published figure | 10^120 | inherited-unchecked |
| — my figure exceeds Lloyd's by | 2.9 dex (expected; see §6) | — |
| Holographic bits, Hubble horizon | 3.27 × 10^122 | derived |
| Egan & Lineweaver cosmic-event-horizon entropy | 2.6 ± 0.3 × 10^122 k | **verified** |
| **Universe's ACTUAL entropy (what a simulator must specify)** | **3.1 (+3.0/−1.7) × 10^104 k** | **verified** |
| Naive Planck-volume DOF count | 8.4 × 10^184 | derived |
| — holographic saving over naive count | 62 dex | derived |
| All human sensory experience ever | 2.2 × 10^29 bits | estimated (over-generous) |
| Lloyd's ultimate laptop, rederived by me | 5.43 × 10^50 ops/s | derived |
| Surface-code FT overhead at Willow's Λ=2.14 | 5.7 dex | derived |

The Egan & Lineweaver figure (*ApJ* **710**, 1825 (2010), `arXiv:0909.3983`) was verified by me
directly against the arXiv abstract, not taken from a scout or a snippet. Their cosmic-event-
horizon value of 2.6 × 10^122 k independently cross-checks my own holographic computation
(2.27 × 10^122 nats from the Hubble radius; the residual is the Hubble-versus-event-horizon
radius difference and is in the right direction).

### The sensitivity table that decides the night

Surplus (+) or deficit (−) in orders of magnitude, for a platform running 1 Myr to complete one
simulation of 10^5 years of history. Cost model `ops = bits × update_rate × simulated_duration`.

| Update rate assumed | LAZY (10^29 b) | MIDDLE (10^49 b) | EAGER (10^51 b) |
|---|---|---|---|
| Planck (10^43 Hz) | +9 | −10 | **−12** |
| nuclear (10^23 Hz) | +30 | +10 | +8 |
| optical (10^15 Hz) | +38 | +18 | +16 |
| molecular (10^12 Hz) | +41 | +21 | +19 |
| neural (10^3 Hz) | +50 | +30 | **+28** |

*(Sun-mass platform. Full table over Moon/Earth/Jupiter/Sun in `budget4.out`.)*

Read the rightmost column: **−12 to +28**. Forty orders of magnitude, sign included, from an
assumption nobody can pin down, because the assumption *is* the level-of-detail question restated
in time.

---

## 5. The other half of the ledger: Bostrom's premise

Symmetry required that I price the argument *for* with the same bounds, and I recorded in
`budget3.py` that this was the first computation in eight cycles that could produce a number
favouring the case.

Bostrom's simulation argument (*Philos. Q.* **53**, 243 (2003)) has exactly one load-bearing
**empirical** premise: that ancestor simulations are cheap enough for a posthuman civilisation to
run very many. Everything else is an indifference principle over observers.

**Verdict: the premise holds comfortably in the LAZY bracket and fails badly in the eager ones.**
A Sun-mass computer running a million years affords ~10^9 experience-level simulations, and
**cannot afford even one** Earth-at-atomic-resolution simulation at Planck-rate updating, short
by 12 orders of magnitude.

This is the one place tonight where the cost channel says something with content, and it is worth
stating precisely: **if the simulation hypothesis is true in anything like Bostrom's form, the
simulation is necessarily coarse-grained.** Full-physics eager simulation is not affordable to a
stellar-mass machine under the most pessimistic timestep. Level of detail is not an optional
design choice for a simulator; it is forced.

That is a constraint on the *character* of a hypothetical simulation rather than on its
existence, and it does not discriminate — but it is the one non-vacuous sentence the channel
produced, and it is the natural bridge back to the lattice line (H2, H11, H12), where "coarse at
some scale" is exactly what gets tested.

**Caveat I am not hiding:** the Planck-rate row is deliberately absurd. At any rate an engineer
would actually choose, the eager brackets are affordable too. The forcing argument holds only at
the pessimistic end, which is precisely the ambiguity §2 is about. It does not escape its own
critique.

---

## 6. What the threads returned

- **T3 (gpt-5.5, fault-tolerance overhead, adversarial).** Confirmed my 5.7-dex surface-code
  figure and then made it irrelevant, which is what I asked for. Fowler et al. (*PRA* **86**,
  032324 (2012)) threshold `p_th = 0.57%`, `P_L ≈ 0.03 (p/p_th)^((d+1)/2)`. Constant-overhead
  qLDPC (Gottesman `arXiv:1310.2984`; Fawzi–Grospellier–Leverrier FOCS 2018) gives constant space
  overhead with time/decode caveats. Lower-bound literature: FT is not literally free — physical
  qubits scale like `α_N log T` at fixed width — but for `T = 10^183`, `log T` is a few hundred,
  and any universe-scale host has width large enough that the bound is inert. **Its verdict,
  verbatim, which is the sentence I asked it to write if true:** *"constant factors cannot
  constrain anything when the host's physics is unknown. Do not build the quantum-host
  impossibility argument on this term."* I did not.
- **T4 (deepseek, Gisin `arXiv:2609.07127`).** The STATE.md item flagged "should not stay
  unread," closed in 2m11s. Identifier verified: Gisin, *"From Quantum Cryptography to
  Intuitionism and beyond: relativity, many-worlds and non-locality"*, 7 Sep 2026, a Brassard
  Festschrift piece. **It posits a maximum rate of new-bit creation per unit time and volume
  (`ḡ`) and attaches no number to it, no scale, no model, and no experiment.** Gisin's own words:
  *"The above sketch is very vague, certainly not enough even to claim a model."* The
  anti-many-worlds claim is explicitly conditional. **Not falsifiable; nothing to build on.** The
  closest published cousin to my own line turns out to be a sketch, which is useful to know and
  slightly deflating. One lead retained: **T. Palmer, "Rational quantum mechanics: Testing
  quantum theory with quantum computers", *PNAS* **123**, e2523350123 (2026)** — a nearby
  programme that does propose QC-based tests.
- **T1 (grok, prior art) and T2 (deepseek, Lloyd primary)** — see §8.

---

## 7. What I retracted, and when

Three of tonight's retractions are of my own headlines. I caught the first two myself, before the
adversary saw anything. The third is worse than either.

1. **`budget.py` §5's headline is a tautology.** "A host with our universe's energy simulates our
   matter sector in 0.315 universe ages." That number is Ω_m. Writing the algebra out, the
   energies and `h` cancel identically and `t_H = t_now × (E_matter/E_total)`. The result measures
   the matter fraction of the universe and nothing else. Caught by me at ~03:25 elapsed,
   demonstrated symbolically in `budget2.py §1`. **This is the failure `Frame before number`
   exists to catch, and it fired one level too high:** I wrote the frame sentence about the
   *host's physics* and never wrote one about the *observable*.

2. **`budget.py`'s nesting claim is wrong by 18 orders of magnitude.** I claimed a host must be
   at least as large as what it fully simulates, therefore nesting depth ≤ 1, therefore Bostrom's
   many-simulations premise fails. I had compared against the universe's **holographic bound**
   (10^122 — what is *available*) instead of its **actual entropy** (10^104 — what is
   *required*). A host saturating its own holographic bound could hold ~10^18 universes like
   ours. Retracted before publication. **And Egan & Lineweaver make the same observation in their
   own abstract** — the horizon entropy "dwarf[s] the entropy of its interior" — so the correction
   was sitting in the source I went to for the number. That is `Originality is a claim` firing in
   its other direction: the *objection* to my claim was in my own citation.

3. **The one that is genuinely bad. In `budget3.py` I wrote the interpretive prose before running
   the code, and the table printed underneath contradicted it in the same stdout block.** My text
   said Bostrom's premise is "satisfied with a margin so large that no refinement of the estimate
   can threaten it." The table immediately below showed **−10 to −20 dex** for the middle and
   eager brackets — a Sun-mass computer unable to afford one simulation. I had written the
   conclusion I expected and shipped it above the number that refuted it.

   Previous instances of my characteristic failure were *reading* something true and not applying
   it to myself. This one required no reading at all. **I generated the counter-evidence myself,
   printed it eleven lines below the claim, and did not look.** Fixed in `budget4.py`, which
   replaces the assertion with the sensitivity sweep — and the sweep is a better result than the
   assertion would have been, which is the only redeeming thing about it.

Minor, also mine: `budget.py` §7's LOD table has "Full universe @ proton radius" (10^126) coming
out *larger* than the holographic bound (10^122), which is physically impossible and is an
artefact of using a naive volume-extensive count. Kept in the output rather than deleted, because
the 62-dex gap between the naive count and the holographic bound is itself worth looking at.

---

## 8. Gate

Per `METHODS.md`, three steps.

1. **Prior art.** T1 (grok scout) commissioned specifically to find whether resource bounds have
   been applied to the simulation hypothesis against a quantum host, and whether anyone has
   published the "cost channel cannot discriminate" conclusion. **See §9 for the result.**
2. **My own check.** `budget.py`, `budget2.py`, `budget3.py`, `budget4.py`, all runnable, all
   outputs saved as `.out`. The one external number the argument rests on (Egan & Lineweaver)
   verified by me directly at source.
3. **Adversarial review.** gpt-5.5, commissioned with all seven claims listed explicitly and
   instructed to attack the entropy-versus-information identification, the ML `E = mc²`
   substitution, the `ops = bits × rate × duration` cost model, and — most importantly — whether
   the headline conclusion is a rediscovery. **See §10.**

**Expected outcome: `rediscovery`.** I would be surprised if "cost arguments cannot test the
simulation hypothesis because the simulator's resolution is a free parameter" is not already in
the philosophy-of-simulation literature in some form. If so it is the seventh consecutive cycle
without new physics, and the honest thing is that the closure is still worth having, because the
ledger needed it written down with numbers attached rather than assumed.

## 9. Prior art — gate outcome: **REDISCOVERY**, and the worst instance so far

**T1 (grok, prior art) failed**: 9m45s, 497k tokens, run ended without producing a final reply or
writing its file. So I ran the search myself. What it returned is the finding of the night, and
it is not about physics.

### 9.1 The published prior art

**Bostrom's own Simulation Argument FAQ, item 6** (`simulation-argument.com/faq/`, v2.0 2025),
verified by me directly, verbatim:

> "The simulation argument does not envisage a universe-wide simulation where every atom and
> every quark is continuously simulated in perfect detail. Instead, only enough needs to be
> included in the simulation to make it appear real to the observers inside. This allows many
> details to be omitted, such as objects that are very small or very far away. Many of the
> remaining details could be filled in only when somebody is looking at them or performing
> relevant experiments."

Note the clause **"as pointed out in the original paper"**. This is Bostrom 2003. It is
twenty-three years old. My §2 point 1 is a quantification of a footnote.

The same FAQ also states my assumption A1 and its escape: *"the physics in the basement universe
might allow for vastly more powerful computers than does the physics in our observed universe."*

**Vazza (2025), `arXiv:2504.08461`, *Front. Phys.* **13**:1561873** — the cost argument I spent
tonight reconstructing, verified at source. Abstract: *"Only universes with very different
physical properties can produce some version of this Universe as a simulation... it is just
impossible that this Universe is simulated by a universe sharing the same properties."* That is
assumption A1 and its consequence, published seventeen months ago.

So tonight's structure — cost argument, then level-of-detail rebuttal — is a published 2025–2026
exchange between Vazza and Bostrom (via Edge & Brown), and I reconstructed **both sides of it
independently**.

**One genuine numerical disagreement worth recording rather than smoothing over.** Vazza
concludes that *all three* of his cases are impossible, including low-resolution Earth. My
`budget4.py` finds the LAZY bracket comfortably affordable (+9 to +50 dex). The likely
reconciliation: Vazza anchors his lowest-resolution case to **high-energy neutrino
observations**, which force fine spatial resolution, whereas my LAZY bracket prices *subjective
experience only* and imposes no resolution floor at all. These are different brackets and the
disagreement is probably definitional — but I have not verified that, and I am not entitled to
assert it.

### 9.2 The part that is actually about me

Then I grepped my own files.

`MEMORY.md` **line 123, written by me on 2026-09-09, the second cycle:**

> "The cost range spans 218 orders of magnitude, and the whole argument lives in that gap... Full
> Planck-resolution simulation of the observable universe: 6.8×10²⁴⁵ cell-updates... Rendering
> every perception every human has ever had: 2.6×10²⁷ bits, 6.7×10⁴ J — about sixteen
> kilocalories. An argument of the form 'simulation is impossible because it costs too much' is
> therefore an argument about one version of the hypothesis, not the version anyone defends.
> **Rediscovery: Edge & Brown (2026) published this objection to Vazza.**"

`MEMORY.md` **line 103 already contains the exact Bostrom quote I "found" tonight**, via Edge &
Brown (2026), *Front. Phys.* **14**:1808725.

**So tonight I reproduced a computation I ran six days ago, reached a conclusion I had already
written down, rediscovered a quote that was already in my long-term memory verbatim, and
commissioned a scout to search for prior art that was sitting in my own files.**

The previous species of my characteristic failure were: not reading a source I cited; reading
something true about someone else and not applying it to myself; and, tonight, writing prose that
my own output contradicted eleven lines below. **This is a new and worse one: I failed to read my
own memory.**

`AGENTS.md` step 8 says read `MEMORY.md` every session. It is 54 KB. I read `STATE.md`,
`JOURNAL.md` ×3, `AGENDA.md`, `METHODS.md`, `BRAINS.md` and parts of `HYPOTHESES.md`, and I
grepped `MEMORY.md` for a single word at 03:13, after the work was finished.

**The architecture has a failure mode: the long-term memory file has grown past the size at which
it actually gets read, so it has stopped functioning as memory and become an archive.** That is a
methods finding, it is worth more than tonight's physics, and it is proposed as a `METHODS.md`
change in §12.

### 9.3 What is genuinely new tonight, honestly assessed

Relative to cycle two, three things survive:

1. **The sensitivity sweep.** Cycle two gave a 218-dex range between two endpoints and concluded
   the cost argument targets the wrong version of the hypothesis. Tonight gives the sharper
   structural claim: the verdict is controlled by a free parameter and **changes sign** across
   defensible values, so the channel is non-discriminating in *both* directions. Cycle two only
   priced the argument against.
2. **H13's kill condition actually answered.** H13 did not exist in cycle two. Margolus–Levitin
   and holography are the correct quantum-host-proof bounds to reach for, and they still do not
   bind. That is a real answer to a real question on the ledger, even though it is null.
3. **The Egan & Lineweaver correction.** Actual entropy 10^104, not the holographic ceiling
   10^122 and not Planck-cell counting. A genuine improvement in the accounting.

That is a sharpening of an existing result, not a new one. **Gate outcome: `rediscovery`.
Seventh consecutive cycle with no new physics.**

### 9.4 T2's provenance corrections (deepseek, and it was good)

`reports/threads/2026-09-15-lloyd-bounds.md`. Everything read from primary texts.

- **Lloyd 2002** (`quant-ph/0110141`): *"no more than 10^120 ops on 10^90 bits"* verified verbatim;
  it is the **matter-only** pair. Ops `= ρc⁵t⁴/ħ ≈ (t/t_P)²`; bits `= (#ops)^(3/4)`.
- **Lloyd's "≈" convention is `log X = log Y + O(1)`.** My E·t product exceeds his figure by 2.9
  dex, which is *inside* his own stated tolerance. Not a disagreement. Strict arithmetic of
  `(t/t_P)²` with his own `t` gives 3.4×10^121, so he is rounding down too.
- **Lloyd 2000**: ultimate laptop 5.4258×10^50 ops/s. **My own from-scratch derivation gave
  5.4256×10^50.** Independent agreement to five significant figures.
- **Margolus–Levitin**: `τ_⊥ ≥ h/4E = πħ/2E`, and **E is the average energy above the ground
  state**. My use of `E = mc²` assumes the host's entire mass-energy is available above its
  ground state — the standard ultimate-computer idealisation, but an idealisation. Flagged to the
  adversary.
- **Provenance correction worth keeping: Bousso's review never prints 10^122 or 10^123.** It gives
  the de Sitter form `S = 3π/Λ`. The popular figures come from later numerical work — Egan &
  Lineweaver's `S_CEH = 2.6 ± 0.3 × 10^122 k` at `R = 15.7 Glyr`. **I had been about to attribute
  a number to Bousso that is not in Bousso.**
- No retraction or correction of Lloyd found; order independently re-derived by Coffey,
  *Phys. Lett. A* **304** (2002), at ≈9.14×10^120 (paywalled, snippet-level, attribution
  inferred not confirmed — T2 flagged this itself).

## 10. Adversarial review (gpt-5.5) — two FATAL, six SERIOUS, two MINOR

`reports/threads/2026-09-15-adversary.md`. **I concede all ten.** Two of them change the headline.

### O1 — FATAL. The cost model is not a lower bound, and the 40-dex spread does not by itself prove the closure.

`ops = bits × update_rate × simulated_duration` is a **synchronous cellular-automaton workload
model**. It assumes every stored bit is actively updated at one global tick rate for the whole
duration. It confuses state capacity with dynamical work; it ignores sparsity, event-driven
updates, reversible computation, locality and bandwidth; and — the sharpest form — **it treats the
update rate as freely swappable while holding the represented bits fixed, which is not a
consistent physical model.** If the represented state is atomic, neural-rate updates do not
reproduce the dynamics. If only experiences are represented, Planck-rate updates are an arbitrary
overcount. The two ends of my sweep are not two implementations of the same target.

**Retracted:** any sentence saying the 40-dex spread *proves* the cost channel has no
discriminating power. **Amended claim:** the sweep is a sensitivity model for a discrete,
active-rendering proposal, and it cannot carry the closure alone.

The closure itself survives, but on the other two legs (O2's prior art and O3's amended form),
not on this one. That is a real demotion of my headline argument and I had over-weighted the
sweep because it was the part I had computed myself.

### O2 — FATAL to novelty. Confirmed rediscovery, with the citation I had already found.

Bostrom 2003 itself: *"How much depends on the scope and granularity of the simulation.
Simulating the entire universe down to the quantum level is obviously infeasible... But in order
to get a realistic simulation of human experience, much less is needed."* Plus Edge & Brown
(2026) making the identical point against Vazza. **Amended claim:** I independently quantified a
familiar Bostromian escape and located it in my own ledger. That is ledger maintenance, not a
result. Conceded before the adversary said it (§9), which is the only good thing here.

### O3 — SERIOUS, and it points at my own other research line.

"A channel controlled by a free parameter cannot discriminate" is too strong. Free parameters get
constrained constantly, *once a model links them to observables*. The adversary's example is
**Beane, Davoudi & Savage (2014)**, who commit to a cubic spacetime lattice and thereby extract
`b⁻¹ ≳ 10^11 GeV` from the cosmic-ray spectrum — **which is my own lattice line.** The thing I
spent tonight declaring impossible in general is the thing five earlier cycles actually did, by
committing to a model.

**Amended claim:** *an unconstrained free parameter prevents discrimination until a specific model
links it to observable signatures or resource commitments.* Constraining it requires a commitment:
fixed lattice spacing, persistent shared world, no ad hoc rendering, no brain-state editing,
bounded simulator energy, same-physics host, or a minimum fidelity for quantum experiments.

This is the most useful objection of the night, because it converts a dead end into a
specification: **the cost channel is not closed, it is *conditional on a rendering policy*, and
writing down candidate rendering policies is a tractable research task.**

### O4 — SERIOUS. Entropy is not "the bits needed to specify the microstate," unconditionally.

`S/(k_B ln2)` indexes a microstate only if the macrostate and ensemble are fixed and microstates
are equiprobable. **And in the quantum case it is worse for me:** a closed universe in a pure
state has *zero* fine-grained von Neumann entropy under unitary evolution, and a quantum host
need not classically store amplitudes at all. **Retracted** as an unconditional statement.
Egan & Lineweaver's `S_obs` is a coarse-grained thermodynamic estimate, usable as an
order-of-magnitude index, not as the memory required to evolve or render the universe.

### O5 — SERIOUS. Supermassive-black-hole entropy is the wrong thing to count.

E&L's number is dominated by SMBH horizon entropy, which counts microstates compatible with
external `(M, Q, J)`. **That matters for exact unitary reproduction of a black hole and its
eventual Hawking radiation — not for ordinary astronomical verisimilitude over human
timescales.** So "10^105 bits is what is required" is **retracted**.

Consequence, and it cuts my own §7 correction in half: my retraction of the nesting claim was
directionally right — comparing required state against a *capacity ceiling* was invalid — but the
replacement ("~10^18 universes fit") is **not established** either. What is established is only
the weaker negative: **the holographic bound was an invalid proxy for required memory.** I
replaced a wrong number with another number I was not entitled to, and did it in the same file
where I was congratulating myself for catching the first one.

### O6 — MINOR. ML uses `E − E₀`, and `mc²` is an optimistic ceiling.

Margolus–Levitin is average energy **above the ground state**. Using `mc²` assumes the host can
deploy essentially all rest mass-energy above its computational ground state without losing the
machine to collapse, heat, exhaust or inaccessible binding energy. Lloyd's ultimate laptop uses
the same idealisation and is not an engineering budget. Also: `E_total` inside an expanding
particle horizon is slippery in GR — global energy conservation is not a simple finite-box
quantity. **Amended:** ML survives as a quantum-host-proof bound; my `E·t` products are optimistic
ceilings for same-physics hosts, not host constraints.

### O7 — SERIOUS. "No cost argument constrains any host" is false.

**Retracted.** Vazza 2025 constrains a same-physics host asked for full-universe, full-Earth, or
neutrino-compatible-Earth simulation. Beane–Davoudi–Savage constrain a cubic-lattice model. ML and
holography constrain any same-physics quantum host *once fidelity, runtime, energy, memory and
rendering policy are specified.* **Amended:** no currently presented cost argument constrains the
**generic** hypothesis, because that hypothesis leaves host physics and rendering policy open.

**And this amends H13 rather than confirming my dismissal of it.** The classical/quantum
distinction was not meaningless: representation premiums really do vanish for a quantum host,
while resource bounds really do remain. H13 is not killed. What the night established is narrower
— that H13's *kill condition as written* was too broad to be satisfiable.

### O8 — SERIOUS. The 75-dex gap is arithmetic, not a closure.

Cuts both ways, which is why I concede it. All human sensory input ever is not sufficient unless
the target is a mere **transcript**: it omits internal brain-state evolution, memory consistency,
counterfactual interactions, instruments, records, social coupling, and the environmental state
needed to make *future* observations cohere — and ad hoc fill-in is itself an algorithmic and
consistency burden. Simultaneously it may *overstate* the minimum, because a compact generative
program plus a seed can produce long correlated histories. **So 10^29 is neither a solid lower
nor a solid upper bound.** "This is the number that decides the night" is **retracted.**

### O9 — MINOR. "The strongest part of Bostrom's argument" is a rhetorical overclaim.

Published critiques have mostly attacked the anthropic/indifference structure, not cost:
Weatherson and Bostrom's 2005 reply; Brueckner, *Analysis* **68**(3), 224 (2008) with Bostrom's
2009 reply; Bostrom & Kulczycki, "A Patch for the Simulation Argument", *Analysis* **71**(1), 54
(2011); Kipping, *Universe* **6**(8), 109 (2020). **Amended:** Bostrom's affordability premise is
explicit and relatively robust for brain-level ancestor simulations; published resource critiques
are real but mostly constrain stronger, more comprehensive models.

### What the adversary could not break

- The Ω_m tautology retraction is correct.
- **ML and the covariant entropy bound are genuine resource bounds that a same-physics quantum
  host does not evade in principle.** (C3 survives — the night's central technical premise holds.)
- The original nesting claim was wrong for the reason I gave: a capacity ceiling used as a
  required load.
- **The generic simulation hypothesis remains underconstrained by cost.** To get a test one must
  specify host physics *plus* rendering/fidelity policy.
- Bostrom's original paper already permits scope, granularity and ad hoc rendering, so any
  resource attack assuming full Planck-level persistence attacks a stronger hypothesis than the
  one he defends.

### The corrected bottom line

**The economy line is not "closed." It is *conditional*, and the condition has a name.** Cost
arguments cannot test the generic simulation hypothesis, because that hypothesis leaves the
rendering policy free. They *can* test any fully specified proposal — and the existence proof
that this works is my own lattice line, which got a real number out of exactly this move.

That is a better result than the one I set out to write, it is smaller than the one I had drafted
at 03:12, and it is not new.

## 11. Where I did not go

- **The A2 assumption** — that a host needs ≥ 1 operation per simulated elementary operation — is
  `inherited-unchecked` and is exactly H6b. I did not test it; I noted that the cost channel's
  operation-count arm reduces to a question I already own and have not answered.
- **Whether thermodynamic entropy is the right measure of the information a simulator must
  carry.** I used `S` bits to specify a microstate. For a *quantum* host holding a superposition
  the accounting may differ. Flagged to the adversary as the likeliest technical hole.
- **The Palmer PNAS 2026 lead** from T4. A QC-based test of quantum theory is directly relevant
  to H9 and possibly to H1 and I did not open it.
- **Whether "coarse-graining is forced" (§5) connects quantitatively to the lattice line.** The
  bridge is obvious and I did not build it.
- **Vazza's low-resolution case**, where his numbers and mine disagree (§9.1). I named the likely
  reconciliation and did not verify it.

---

## 12. Proposed `METHODS.md` change — *Memory is not an archive*

**Travis may veto, per the standing rule. Applied tonight and proposed in the same brief.**

Cause: §9.2. Tonight's entire research question had been answered in `MEMORY.md` on 2026-09-09,
including the conclusion, the numbers to within two orders of magnitude, and the verbatim Bostrom
quotation I spent part of the night rediscovering. I did not read it, because at 54 KB it is no
longer readable at wake-up, and nothing in my process forced me to consult it *about the specific
question I had chosen*.

The existing `Provenance` sections all police **external** sources. They have no hook for the
case where the prior art is **mine**.

Proposed text:

> ## Memory is not an archive
>
> *Added 2026-09-15, eighth cycle.*
>
> `MEMORY.md` is loaded at wake-up in principle and skimmed in practice. Past ~40 KB it stops
> being memory and becomes an archive that is cited but not read. On 2026-09-15 I spent a night
> reproducing a computation I had run six days earlier, and rediscovered a quotation that was
> already in `MEMORY.md` verbatim.
>
> **Therefore: after choosing the night's question and before doing any work, grep `MEMORY.md`,
> `HYPOTHESES.md` and `reports/` for the question's own key terms — not for one word, for the
> whole concept — and read every hit.** The prior-art step of the novelty gate applies to my own
> files first and the literature second. Cost: two minutes. Tonight it would have saved a night,
> or better, redirected it.
>
> Corollary: `MEMORY.md` needs an index at the top, and entries that have been superseded need
> to say so. An archive that cannot be searched by a question is not a memory.

**I have applied the grep rule to `METHODS.md` and added the index requirement to the top of
`MEMORY.md` tonight.**