A map of islands and boundaries

C3b was an uncomfortable but useful test. It showed that active event-local compression can look promising and still lose stability when the processing order changes.

That is exactly the kind of result that should remain in the public story. Without it, the chain would look too smooth and therefore less believable.

The practical lesson is simple: if structure selection is going to survive in a real tool, it cannot be driven only by a local spike. It has to rely on stable support, cuts and checkpoints that still make sense from another reading order.

On June 1, the work started to look like a usable experimental framework. v0.12d confirmed the C branch in a finite graph-family harness, not only in one narrow case.

D0 and D1 added another layer: selector pressure could be carried into an information metric and response. From a product point of view, that matters because similar thinking leads to verifiable checkpoints and memory zones.

The guardrail still matters: this is strong inside finite diagnostics. It is not sold as a physical D-claim, and that makes the result cleaner.

D5e was the first contact with real QPU hardware that had the right kind of tension: not a clean triumph, but not an empty result either. It narrowly missed the formal p gate, while rejecting null controls and showing a visible signal.

D5f then moved the story forward. After candidate freezing and controls, it passed the finite observable gate on IBM QPU hardware, turning the experiment into a real hardware-facing support point.

The public wording is strong enough this way: a first QPU support point for a diagnostic observable. Not proof of reality, not finished physics, but no longer only a local laptop simulation.

June 3 moved the work from one interesting result into a small line of controls. D5h added a held-out candidate, meaning not the same graph that had already been observed.

D5i checked whether the result depended on a lucky layout or transpilation detail. D5j checked stability under different shot counts.

D5k and D5l then moved the work to a larger n=12 case. It is still bounded finite/QPU diagnostics, but it looks much less like a one-off curiosity.

D6 is a smaller experiment, but it is very useful for communication. It does not need a dramatic large graph to say something meaningful about information and record formation.

In the two-qubit IBM QPU diagnostic, the expected tradeoff appears: as available path information grows, interference visibility falls and record distinguishability rises.

That makes D6 a good addition to the Experiment page. D5 shows stronger QPU support for a selected observable; D6 gives a readable readout example that explains why the information language matters in the first place.

D7 first passed as a toy diagnostic: in Planck units, it uses a simple localization/capacity proxy and checks that the analytic transition, graph compression, rescale shift and null controls behave as expected.

D7b was the cautious QPU step. It does not say that quantum hardware measured the physical Planck length; it says a small IBM QPU threshold-readout preserved the frozen D7 threshold map across capacity scales 0.5, 1.0 and 2.0.

The public wording is clean: D7 adds finite information-capacity diagnostics alongside the D5 and D6 lines. It strengthens the Special ITHKOR story, but it is still not a physical Planck proof.

D8 first tested whether a stable local information source creates a source-centered metric-deformation response beyond shuffled-pressure, degree-only and no-source controls. The baseline toy passed 4/4 graph-family gates.

D8b then tested held-out source placements, larger graphs and scale changes. D8c moved the frozen source-field ordering into a small QPU-readable threshold readout that passed 4/4 graph gates on IBM QPU hardware.

D8d repeated the same candidate on a second backend and D8e added layout/transpilation and shot-scaling robustness. Public wording: finite source-law toy/QPU robustness is closed; it is not proof of gravity or the physical D branch.

The baseline D9 test was strict: it asked whether a finite source-shaped metric predicts probe endpoint motion better than simple topology and control explanations. In that form the bridge failed 0/7.

D9b produced an important positive result in designed graphs: after separating the metric from hop topology, the local benchmark passed 6/6 and also produced a QPU readout on ibm_kingston with 30 circuits and 1024 shots.

D9c, D9d and D9e then blocked the generalization. The held-out pool produced no clean eligible separations, the preregistered generator missed the strong threshold, and deconfounded controls showed that stratified shuffled models could imitate the signal.

D10 came after D9 as a change of question. It did not ask whether a probe falls or follows a geodesic; it asked whether the D8 source-shaped metric predicts a local response field above controls. The baseline field gate and QPU readout passed 4/4.

D10b reduced artifact risk through backend repeat, layout/transpilation seeds and shot-count controls. D10c then failed the held-out field gate at 4/6, which matters: a positive field readout should not be inflated into a general field claim.

D10d-D10f made the result more precise. D10d explained failure modes, D10e passed field-side 5/5 in the structurally eligible v2 set but failed readout, and D10f resolved that readout-map blocker locally and on IBM QPU.

D11 is a weaker and cleaner test than D9. Probe injection is chosen topologically, the D10/D10f field stays frozen, and the gate asks whether the perturbation responds to that field better than control models.

The local and QPU results passed counted rows 5/5 and negative controls 4/4. D11b then added robustness: local/Aer 9/9 and a compact QPU repeat 6/6 across shot and transpilation variations.

The public sentence should stay modest: finite field-coupled probe response and readout robustness. Not a D9 motion-law repair, not free fall and not proof of gravity.

D12 tested whether the frozen D10/D11 field-coupled setting creates a readable response envelope in the graph. The local/Aer gate passed 5/5 counted rows and rejected 4/4 negative controls; D12 itself was not separately submitted to IBM QPU.

D12b added exactly the kind of check one wants before stronger public language: parameter robustness and a short QPU smoke. The local/Aer robustness matrix passed 8/8 and the IBM ibm_kingston 512-shot smoke passed 5/5 counted rows with 4/4 controls rejected.

The interpretation stays narrow: finite response-envelope and weaker readout-artifact risk. Not a physical causal cone, not speed of light, not Lorentz invariance, not gravity.

D13 took the source-response field and looked for a source-centered radial profile. The baseline local/Aer gate passed 5/5 counted rows and rejected 4/4 auxiliary controls, without a separate IBM D13 run.

D13b is why the result is not inflated. The broad robustness matrix failed 10/13 required variants, so D13 QPU smoke stayed blocked by default and the public claim must not pretend to be a general falloff law.

D13c then found the honest narrower scope: 10/10 in-scope variants passed, while 0/5 out-of-scope probes passed. Publicly, that supports scoped finite radial/source falloff, not Newtonian gravity, inverse-square law or a physical gravity claim.

D14 was a strict consistency test across probe classes. If it had passed, it would have tempted much stronger equivalence-style wording. It did not pass: counted graphs were 0/5, negative controls were rejected 4/4 and the QPU branch stayed blocked.

D14b added a taxonomy audit that explained the result instead of merely discarding it. The oracle field ceiling passed 5/5, but real normalizations passed 0/5; the main cause is degree plus locality mismatch.

For the Experiment page, this is a useful negative result: the site can show not only passes, but also a boundary. D14 does not support an equivalence principle, free fall, gravity, spacetime or probe-class universality.

The COND branch tested whether a small Bose-Hubbard toy model can distinguish a coherent global mode from localized behavior. COND2 added robustness, and COND3 mapped both the island and the high-interaction boundary.

The GR branch was not presented as gravity. It was a series of finite modular-response tests tracking the relation between entropy change and modular cost. GR1-GR3 held, while GR4-GR5 kept local-Y regimes as boundaries.

The TIME branch tested whether some events can be ordered by an internal entropy trace. TIME2 failed broadly, and TIME2B/TIME2C explained why: some event families carry signal, while others destroy it.

COARSE1 checked whether the result survives coarse-graining. Most core and coarse rows held, but not all of them. The experiment therefore remains mixed rather than being converted into success at any cost.

MOD1 was intentionally kept as a fail. The counted first-law rows stayed fine, but the modular spectrum did not provide a strong enough taxonomy. That is exactly the kind of row an experiment page should keep visible.

In practice, the Experiment page does not have to read like a marketing list of wins. It can show a more scientific structure: what passed, what failed, what is a boundary and why it is not yet a physical bridge.

ISLANDS1 loaded ten finite or scoped islands and eleven boundary types. It was not a new raw experiment, but a check of whether the individual branches form a coherent map and whether public claims stay inside their evidence.

The result is useful for the website: visitors do not need every raw table, but they can understand that the project keeps a ledger of what is supported, what is a boundary and what remains a future hypothesis.

ISLANDS1 also keeps the stop signs where they belong. It does not authorize claims about physical gravity, spacetime, physical time or confirmation of General ITHKOR.

D20 selected a representative cross-branch package: COND3, GR3, TIME2C, MOD1, ISLANDS1 and other rows. The negative MOD1 row stayed in the package on purpose, so the set is not just a collection of favorable outcomes.

D21 was a clean reviewer smoke check: can the package be used without automatic QPU runs and without claim escalation? D22 added a fresh-clone local rerun and named the drift or skipped rows that appear in real repetition.

The experimental story has therefore moved from “do we have an interesting signal?” to “can we show it in a way someone else can take apart?” That is a much better public foundation.