Islands, Blinders, and Conditional Survival: Three Ways Entropy Curves Bend

I · One Six-Step Chain, Three Distinct Information Accounts

Across theoretical physics, quantum communications, and stochastic molecular biology, the natural logarithm measures two very different kinds of surprise. Given a quantum density matrix ρ of unit trace (Tr ρ = 1), the von Neumann entropy S_vN(ρ) = −Tr(ρ ln ρ) measures how much quantum correlation has leaked into degrees of freedom outside our chosen algebra of observables. Given a stochastic process on path space with physical law ℙ and conditioned survival law ℙ^h, the Kullback–Leibler divergence D_KL(ℙ^h ‖ ℙ) measures the sampling rarity paid when an observer keeps only trajectories that evade absorption across a chain of hazards. Both quantities carry the exact same unit—the nat, where 1 qubit of maximal entanglement equals ln 2 ≈ 0.6931 nats—yet equating them turns rigorous theorems into category errors.

Place three celebrated problems onto a single benchmark of N = 6 sequential steps (k = 0, 1, 2, 3, 4, 5, 6). First, in unitary models of black-hole evaporation following Page (1993), Almheiri et al. (2019), and Penington (2020), generalized radiation entropy rises to 3 ln 2 ≈ 2.079 nats at k = 3 and returns to 0 nats at k = 6 with 0% run rejection. Second, in the preprint critique of Geng et al. (2026, arXiv:2602.06543, capped epistemically at PROPOSED), an observer whose detector algebra omits gravitational Gauss-law dressing also discards 0% of runs, yet sees coarse-grained entropy climb to 6 ln 2 ≈ 4.159 nats inside a globally pure universe. Third, in Doob's (1957) conditioning of absorbing Markov chains—cast dynamically by Chetrite and Touchette (2015)—forcing 100% intact arrival across 6 checkpoints of hazard γ = 0.25 discards 82.20% of physical runs and extracts a path-space KL toll of V_Doob = 1.726 nats.

II · Regime 1: Unitary Page Turnover and Quantum Extremal Saddles

Begin with Don Page's two 1993 theorems (DERIVED mathematical physics). In Physical Review Letters 71, 1291 (1993), Page proved that if a bipartite Hilbert space of total dimension m · n (with m ≤ n) sits in a Haar-random pure state, the average von Neumann entropy of the smaller m-dimensional subsystem is S_Page(m, n) = Σ_{j=n+1}^{mn} (1 / j) − (m − 1) / (2n) nats. Four months later, in Physical Review Letters 71, 3743 (1993), he applied that counting law to an evaporating black hole: if evaporation is globally unitary, radiation entropy cannot climb Hawking's line k ln 2 past the halfway point where the black hole's remaining coarse-grained dimension 2^(N−k) drops below the radiation's dimension 2^k. On our N = 6-qubit benchmark, maximal Bell-paired emission follows the exact tent min(k, 6 − k) ln 2—rising through 0.693, 1.386, and peaking at 3 ln 2 = 2.079 nats at k = 3 before purifying symmetrically back to 0 nats at k = 6—while Haar-random states peak at S_Page(8, 8) = 1.589 nats (0.491 nats below 3 ln 2, approaching the universal 1 / 2 = 0.500 nat Haar deficit at large N) and likewise hit 0 nats at k = 6.

For 26 years, semiclassical gravity lacked a geometric saddle that reproduced Page's turnover without inserting unitarity by hand. In 2019–2020, Almheiri, Engelhardt, Marolf, and Maxfield (2019) and Penington (2020) showed that evaluating the fine-grained entropy of Hawking radiation via the gravitational replica trick requires extremizing the generalized entropy S_gen(I; k) = Area(∂I) / (4 G ħ) + S_bulk(R_k ∪ I) over candidate bulk regions I ('islands') and taking the minimum across competing quantum extremal surfaces (QES). Before the Page time (k < 3 on our 6-step lattice), the empty surface I = ∅ wins, giving the rising Hawking saddle S_empty(k) = k ln 2. After k = 3, a non-trivial interior island I ≠ ∅ whose boundary ∂I sits just inside the shrinking horizon wins: including I inside the entanglement wedge purifies the interior Hawking partners against R_k, leaving only the shrinking horizon area term S_island(k) = (6 − k) ln 2. At k = 3 both saddles meet at 2.079 nats, and at k = 6 the island saddle reaches 0 nats.

Two strict guardrails govern Regime 1. First, the epistemic badge is DERIVED—never MEASURED. No laboratory or astronomical instrument has observed Hawking radiation from a horizon (61.7 nK for a 1 solar-mass black hole, as derived in T.08) or tested replica wormholes in four-dimensional asymptotically flat spacetime; the QES island formula is an exact semiclassical saddle calculation in Jackiw–Teitelboim gravity coupled to baths and related holographic models. Second, the dynamics is strictly trace-preserving: h(k) = 1 at every step, 0% of runs are discarded, and V_Doob = 0 nats. Late radiation lowers S_gen(k) after k = 3 because unitary entanglement transfers interior information into joint multi-particle correlations, not because failed evaporations were filtered out.

III · Regime 2: Detector Blinders and Undressed Subalgebra Coarse-Graining

Why, then, did the island program spark an intense 2025–2026 debate over massive gravity, non-gravitating baths, and asymptotic observables? In a 2026 preprint, 'Seeing Page Curves and Islands with Blinders On' (Geng et al., 2026, arXiv:2602.06543—explicitly an unrefereed preprint and therefore capped at PROPOSED under our epistemic lattice), the authors examine what distinguishes quantum electrodynamics or ordinary qubit factor spaces from Einstein gravity: the gravitational Gauss law. Because total ADM energy is measurable at spatial infinity from the asymptotic metric falloff, no bulk excitation inside a gravitating spacetime can be strictly local and gauge-invariant unless accompanied by gravitational dressing extending to the boundary.

Geng et al. (2026) argue that if an idealized asymptotic observer deliberately restricts measurements to an undressed radiation subalgebra A_undressed—putting 'blinders on' so the detector ignores Gauss-law gravitational tails linking bulk modes to boundary charges—the observer applies a completely positive, trace-preserving (CPTP) subalgebra restriction E_sub: A_dressed → A_undressed. Let us verify that exact linear algebra on a 2N = 12-qubit pure state (dimension 4096) representing 6 emitted radiation modes paired with 6 interior modes. Globally, |Ψ_global⟩ has norm 1.0, purity 1.0, and global von Neumann entropy S_global = 0 nats: zero information has left the universe. Yet restricting observables at step k to the k undressed radiation modes traces over the interior partners, leaving the maximally mixed subalgebra state ρ_undressed(k) = (I_2 / 2)^(⊗ k) whose entropy rises linearly as k ln 2—hitting 2.079 nats at k = 3 and 6 ln 2 = 4.159 nats at k = 6.

Notice how easily careless summaries distort Regime 2. Wearing subalgebra blinders is NOT post-selection: Tr(ρ_undressed(k)) = 1.0 at every step k, 0% of trajectories are rejected, and V_Doob = 0 nats. Conversely, obtaining 4.159 nats on A_undressed does NOT mean global unitarity failed; it means purity is an algebraic relational property—invisible to an observer barred from measuring gravitational dressing, visible to an algebra that includes dressed operators. Whether operationally realistic detectors in massless four-dimensional gravity behave closer to Regime 1 or Regime 2 remains an active theoretical question, which is why ES.05 records the 2026 critique transparently as PROPOSED alongside the DERIVED semiclassical QES saddles.

IV · Regime 3: Doob's h-Transform, Absorbing Hazards, and Physical Shielding

Turn now to genuine conditional survival across absorbing barriers—a DERIVED theorem of probability theory introduced by Joseph Doob (1957) for conditioned Brownian motion and cast into modern nonequilibrium large-deviation dynamics by Chetrite and Touchette (2015). Consider a discrete carrier advancing across N = 6 sequential checkpoints k = 0 → 1 → … → 6. At each step k < 6, the carrier either survives to k + 1 with probability 1 − γ or falls permanently into an absorbing loss state ∂ with hazard γ ∈ (0, 1). From step k, the unconditioned probability of reaching k = 6 intact is h(k) = (1 − γ)^(6 − k), with h(6) = 1 and h(∂) = 0. Because (P h)(k) = (1 − γ) h(k + 1) + γ h(∂) = (1 − γ)^(6 − k) = h(k), h is strictly harmonic for the Markov generator. Conditioning the ensemble on hitting k = 6 tilts every transition probability by Doob's Radon–Nikodym ratio P^h(x, y) = P(x, y) h(y) / h(x): absorption vanishes because h(∂) = 0, and forward survival jumps from 1 − γ to (1 − γ) h(k + 1) / h(k) = 1.

On paper, the conditioned Doob process P^h looks lossless: 100% of conditioned paths arrive at k = 6. In physical reality, Doob's theorem hands us the exact bill in nats. Summing P^h(ω) ln(P^h(ω) / P(ω)) across all 2⁶ = 64 surviving micro-paths in our verifier yields V_Doob = D_KL(ℙ^h_[0, 6] ‖ ℙ_[0, 6]) = −ln h(0) = −6 ln(1 − γ) to machine precision (10⁻¹⁵). At the benchmark hazard γ = 0.25 per step, unconditioned survival is h(0) = (3 / 4)⁶ = 729 / 4096 = 0.1780 (17.80%): passive post-selection discards 3367 / 4096 = 82.20% of all launched trajectories (requiring 5.62 raw launches per survivor) and pays V_Doob = −6 ln(0.75) = 1.726 nats. That is why post-selected final-state black-hole proposals differ fundamentally from unitary QES islands, and why heralded quantum-repeater links cannot treat post-selected click fidelity as raw channel throughput: lowering physical waveguide loss from γ = 0.25 to γ_shielded = 0.05 lifts true unconditioned arrival from 17.80% to (0.95)⁶ = 73.51% (a 4.13× throughput gain with 0% selection bias) and cuts the remaining conditional KL toll 5.61× down to 0.308 nats.

The exact same absorbing-chain accounting clarifies experimental molecular surveillance in biology without invoking quantum mechanics at all. When a premature termination codon (PTC) interrupts an mRNA open reading frame—such as nonsense and frameshift variants mapped across CFTR transcripts in basic cystic fibrosis genetics by Sharma et al. (2018) and studied via release-factor eRF3a competition in vitro by Lee et al. (2022)—exon-junction nonsense-mediated mRNA decay (NMD) and premature ribosome termination act as sequential absorbing checkpoints of per-step hazard γ. Sequencing or immunostaining only intact full-length chains at the end of translation observes the Doob-conditioned distribution ℙ^h, masking the 1 − (1 − γ)^N degraded transcripts lost along the way; physically reducing the biochemical hazard γ per checkpoint increases true unconditioned arrival (1 − γ)^N rather than paying an exponential −N ln(1 − γ) post-selection tax.

V · Auditing the Three Regimes Side by Side

Place all four operating points on our N = 6 benchmark side by side and every ambiguity disappears. A downward-bending entropy curve in Regime 1 is a trace-preserving unitary saddle switch (DERIVED, 0% discarded runs, V_Doob = 0 nats). A rising 4.159 nat line in Regime 2 is a trace-preserving subalgebra restriction when Gauss-law dressing is omitted (PROPOSED preprint critique, 0% discarded runs, V_Doob = 0 nats, global state still pure at 0 nats). And a 100% conditioned arrival curve in Regime 3a is a Doob h-transform paid for by discarding 82.20% of runs at V_Doob = 1.726 nats—sharply distinct from Regime 3b's physical hazard reduction to γ_shielded = 0.05.

Run verify_essay_islands_conditional_survival.py locally—or toggle the planted-defect knockout in the interactive observatory above—to test all three boundaries deterministically. In normal mode, our pure-Python Jacobi solver verifies the 6-qubit Bell purification curve and the 12-qubit (4096-state) CPTP partial-trace blinders curve to 10⁻¹², while our 64-path branching Markov check verifies Doob harmonicity and V_Doob = −6 ln(1 − γ) = 1.726 nats to 10⁻¹⁴. Passing --planted-defect severs the island turnover at k > 3, falsely sets V_Doob = 0 nats on post-selected γ = 0.25 survival, strawmans Regime 2 as run-rejection, and promotes unmeasured gravitational saddles to MEASURED—forcing an immediate exit code 1.

Primary Sources & Offline DOI Ledger