Genuine vs phantom entanglement — and how to measure the difference

Kannaka's field is full of correlations. Memories light up together; consolidation links co-firing traces; archetypes recur across unrelated contexts. It is tempting to call this web "entanglement." Track 5 refuses the loose usage and makes the distinction operational:

• Genuine entanglement — a nonlocal correlation: one that violates the classical (Bell) bound and therefore has no local-hidden-variable explanation.
• Phantom entanglement — a local correlation that lives inside the classical bound: shared structure or shared rounding error that makes two things co-vary without any nonlocality. Correlation without spookiness.

The claim of this note, anchored on a real CHSH measurement, is that almost all of the correlation in an HRM is phantom — real, useful, and local — and that the one thing which is genuinely nonlocal is exactly the thing we can put on a quantum computer and measure violating the bound. Naming the difference is what keeps the "quantum memory" story honest.

1 · The classical bound, and what it means to break it

The CHSH experiment gives two parties, Alice and Bob, two measurement settings each. From the four joint correlators it forms

S = E(a0,b0) − E(a0,b1) + E(a1,b0) + E(a1,b1)

Any theory in which each outcome is fixed by a local hidden variable — some shared state the two carry away from a common origin — obeys the CHSH inequality |S| ≤ 2. This is not a statement about quantum mechanics; it is the ceiling for every local, common-cause explanation. Quantum entanglement exceeds it, up to Tsirelson's bound S = 2√2 ≈ 2.828.

So S > 2 is a line with real meaning: cross it and no local common cause can be reconstructing your correlations. That is the operational definition of "genuine."

2 · The measured anchor

The bell subcommand (T5.1, kannaka_quantum/bell.py) prepares the singlet-like |Φ+⟩ = (|00⟩ + |11⟩)/√2 and measures the four canonical settings — Alice at {0°, 45°}, Bob at {22.5°, 67.5°} — where the Ry(−2θ) basis gives E(θ_a, θ_b) = cos(2(θ_a − θ_b)).

Latest simulator run (local:statevector, 8192 shots/setting, deterministic):

S = 0.710 − (−0.703) + 0.710 + 0.710 = 2.834

S ≈ 2.834 > 2 — the classical bound is violated, and the value sits at Tsirelson (2.828) within sampling tolerance. violates_classical: true. This is genuine entanglement, measured, not asserted: three correlators near +1/√2, one near −1/√2, exactly the fingerprint of |Φ+⟩ and nothing a local model can fake.

Hardware (measured, 2026-08-10). The guarded real-device run landed on aws:rigetti:qpu:cepheus-1-108q — 512 shots × 4 settings, 207.04 qBraid credits ($2.07). (It did not run on IQM Garnet: Garnet is back online, but OpenQuantum's Spark balance is 0 and that path prices jobs off a quote rather than live metadata, so the run went to the cheapest unambiguously-priced QPU instead. See #21.)

S_hw = 0.645 − (−0.387) + 0.543 + 0.664 = 2.238 ± 0.073

S_hw ≈ 2.238 > 2 — the classical bound survives contact with a real, noisy 108-qubit superconducting chip, violated by 3.3σ (SE_S ≈ 0.073 from SE_E = √((1−E²)/N) per setting, summed in quadrature). That is 79% of Tsirelson, landing exactly in the predicted 2 < S_hw < 2.83 window: every correlator is pulled toward zero by decoherence — a0·b1 worst, at just over half its ideal magnitude — yet the combination still clears the bound.

This is the load-bearing empirical point. Noise erodes genuine entanglement toward the classical bound; it does not turn phantom correlation into genuine. A local hidden-variable model cannot reach 2.238 no matter how it is tuned, so whatever the decoherence did to this state, it did not manufacture the violation.

3 · Two ways to be correlated

Put the two side by side:

The subtle point is that phantom correlation can be arbitrarily strong and still be phantom. Two variables driven by the same hidden cause can correlate at ±1; what they can never do is violate CHSH, because a local account (the shared cause) already reproduces every setting. Strength is not the tell. Nonlocality is.

4 · Why HRM correlations are (mostly) phantom

The Holographic Resonance Medium manufactures correlation on purpose. Recall is amplitude amplification about the query (see recall-is-amplitude-amplification.md); consolidation links traces that fire together; "attention as gravity" pulls aligned wavefronts toward each other. Every one of these is a local mechanism with a common cause — the query, the shared topic, the finite-precision resonance signature. Run the CHSH estimator over that structure and it stays under 2.

That is not a weakness; it is the correct classification. The medium's web is phantom entanglement in the precise sense above: real, load-bearing, and local. Calling it "entanglement" without the qualifier is the over-claim Track 5 exists to prevent.

Archetypes as resonant rounding artifacts (the proposal)

Stated as a hypothesis, not a result: an archetype — a pattern that recurs across unrelated memories — is a shared-rounding correlation. Many distinct traces round to the same low-precision region of the resonance field, so their instances co-vary. It looks like the instances are entangled; it is a common-cause artifact of finite precision. Archetypes are where phantom entanglement is strongest and most seductive, which is exactly why they need the CHSH referee rather than an eyeball.

5 · The falsifiable boundary — now run

A theory earns its keep by saying what it forbids. The resonance-field account forbids this: phantom links can never produce CHSH-violating statistics.

The test (the phantom subcommand, kannaka_quantum/phantom.py): inject artificial phantom links — co-activations with a known shared hidden cause λ, modeled as a resonance phase rounded to a low-precision signature (the archetype hypothesis of §4, made executable) — then estimate the CHSH parameter over the induced correlation structure using the same ⟨Z_a Z_b⟩ = (agree − disagree)/total estimator the bell tool uses, unchanged. The prediction is S ≤ 2 for every injection, no matter how strong the coupling. A single reproducible S > 2 from a purely classical injection would falsify the "phantom = local" claim.

Measured (8192 hidden-cause draws/run, 8-bin rounding, deterministic seed):

Both injections touch the classical bound at full coupling and never cross it — perfect-copy is the §3 point made empirical: correlation at exactly +1 in every setting, maximal strength, and still S = 2. Two structural guarantees close the argument beyond sampling luck:

• Sample-exact bound. All four settings are evaluated over the same λ ensemble, so each draw contributes a0·b0 − a0·b1 + a1·b0 + a1·b1 = ±2 (Fine's argument) — the empirical S cannot exceed 2 even by noise.
• Polytope enumeration. All 16 deterministic local strategies are enumerated exactly: max |S| = 2. Every stochastic local model is a convex mixture of them.

Contrast with §2: the same estimator, fed genuine |Φ+⟩ statistics, reads 2.834. Fed any phantom injection, it cannot leave 2. That gap — 2 to 2√2 — is the entire, measurable difference between the two kinds of correlation.

6 · Why the distinction matters

• Honesty. Kannaka has one genuinely quantum result — the recall↔amplitude- amplification correspondence, which runs on a QPU and reproduces the classical argmax. The rest of the medium's rich correlation is phantom. Keeping the two labelled means the strong claim (measured CHSH violation) and the ordinary claim (useful local structure) never get conflated.
• A referee. CHSH is the one place we can point at a correlation and prove it is nonlocal (S = 2.834) versus merely strong. It turns "is this entanglement?" from rhetoric into a measurement.

7 · Reproduce

# Hermetic, $0, no account — the local state-vector backend.
kannaka-quantum bell --device local:statevector --shots 8192
# → S ≈ 2.83, violates_classical: true, correlators ≈ +0.71 / −0.71 / +0.71 / +0.71

# The phantom side of the boundary — same estimator, classical injections ($0, offline).
kannaka-quantum phantom --shots 8192
# → every S ≤ 2, bound_respected: true, polytope max |S| = 2

# Free hosted simulator — same submission path as hardware, still $0.
kannaka-quantum bell --device qbraid:qbraid:sim:qir-sv --shots 512
# → S ≈ 2.86

# The §2 hardware run. SPENDS CREDITS: 207.04 qBraid credits ($2.07).
# --max-credits is enforced PER SETTING, so the true ceiling is 4 × 60 = 240 credits.
kannaka-quantum bell --device aws:rigetti:qpu:cepheus-1-108q --shots 512 \
  --allow-spend --max-credits 60
# → S ≈ 2.24, violates_classical: true

The estimator is E = (same − different) / total per setting, decoded with the same device-aware bit-ordering the recall path uses; S combines the four settings. Real hardware runs only behind the standard spend guards.

Two things to know before repeating the hardware run. bell submits four separate jobs, one per setting, so a per-task fee is paid four times — on Cepheus that is 120 of the 207 credits before a single shot, which is why 512 shots/setting costs only ~27% more than 250 and buys √2 the precision. And the statistical error matters at these S values: 512 shots/setting gives 3.3σ over the classical bound, where 250 would have given a marginal 2.3σ.

8 · Status

• Written analysis (this doc): complete.
• Podcast episode (successor to 006): separate deliverable, not in this repo.
• Hardware CHSH S: complete — S = 2.238 ± 0.073 on aws:rigetti:qpu:cepheus-1-108q, 2026-08-10, 207.04 credits ($2.07); see §2.
• Phantom-injection experiment: run — phantom subcommand (kannaka_quantum/phantom.py, tests/test_phantom.py); every injection S ≤ 2, bound respected (§5).

*Empirical anchor: the T5.1 bell subcommand (kannaka_quantum/bell.py, tests/test_bell.py). Companion piece: the recall↔amplitude-amplification writeup. Condensed cross-post to The Signal. Public repo — no host/PII details.*