# One missed pulse flips a placement-dependent phase, not its completion count

14 September2026. **Executed conditional abstract-tissue experiment; not biological measurements, R104Q calibration or intervention efficacy.** All six prospectively bounded coarse cases completed, followed by independent endpoint, geometry/conductance and full-state verification. Original model and historical outputs remain unchanged.

## New result

With the exact existing spatial masks, omit only the sixth scheduled stimulus while continuously integrating the same400ms window. Compare the same six later stimulated windows7–12:

|Allocation|Regular pacing, windows7–12|Omit pulse6, windows7–12|Completion count|New observation|
|---|---|---|---|---|
|Near mask,x0p5|1 1 1 1 1 1|1 1 1 1 1 1|6/6→6/6|Reliable over this tested interruption/horizon|
|Far mask,x0p6|0 1 0 1 0 1|1 0 1 0 1 0|3/6→3/6|Strict alternating phase inversion; opposite final-window label|
|Equal-mean uniform|1 0 1 0 1 0|1 0 1 0 1 0|3/6→3/6|Same intermittent sequence|

Here1means all31distal checkpoint nodes activated;0means that completion criterion was not met. Actual per-window launch, activated-node and distal-node counts remain in the saved records. Every deliberately omitted window had0actual crossings across active nodes in these runs; zero was **measured from the model outputs, not assigned by the omission label**.

The new causal comparison shows that the farther-mask final-beat classification is sensitive to a single missed pulse, even with unchanged mask and unchanged post-interruption completion count. The regular-pacing near-mask advantage survives this particular interruption. It does not show that the near allocation is universally superior, that biological delivery is tolerant to interruption, or that the total three-of-six response improved.

![Verified beat-sequence comparison](delivery_history_v1.png)

Figure generation, exact source rows, hash gate and visual QA were independently handled after the numerical gates passed. See [caption](DELIVERY_HISTORY_FIGURE_CAPTION_V1.md). The plot is a deterministic output summary, not uncertainty/error bars.

## What was actually changed and preserved

The prospective [novelty specification](DELIVERY_HISTORY_NOVELTY_SPECIFICATION.md) fixed three allocations×two schedules, one CPUworker,600seconds per case and100MiBtotal output cap. No denser placement sweep, new anatomy, biological conductance calibration, GPU or remote job was used.

The isolated C derivative retains the original ionic kernel, geometry, diffusion, mask identities, stimulus footprint/amplitude, dt0.01ms and200isolated400ms conditioning beats. The schedule gates only the coupled stimulus. No pulse6does not skip its integration window or change conditioning to800ms. Each pair's complete end-window5state and first five full node maps match byte-for-byte before divergence. Across different allocations, the original allocation-specific conditioning assumptions remain inherited; this experiment does not isolate those as a new factor.

All-state finiteness is checked at every active-cell **coupled** integration step. This is not an explicit every-step check inside isolated conditioning. Full saved states5/12also passed independent shape and finiteness checks. Omitted-pulse status preserves actual events/counts and is not called failed capture.

## Independent acceptance checks

- All three regular controls reproduce every field of their previously saved first ten beat records exactly.
- Independent Python parsing of all72node maps reconstructs activation, proximal launch and distal counts, checks finite active-node observables, crossing-time ranges and earliest-distal-after-earliest-proximal ordering.
- Separate post-campaign verifier requires exactly one row for each of1860grid coordinates, independently reconstructs the1280active-node geometry, verifies coordinates, and checks every cell's exact mask/value assignment—not merely the mean conductance.
- All six mosaics retain intended mean0.5402and exact128cell treatment masks where applicable. Twelve full-state arrays have1860×27double components, exact length and finite values.
- All three paired pre-perturbation states and first five node maps are identical. A fresh independent pair reconstruction exactly matches final `results.json`.
- Six synthetic verifier tests pass: missing-pulse reporting; corrupted counts; premature distal arrival; duplicate coordinates despite preserved row count; wrong active geometry; and compensating per-cell conductance errors. Three malformed CLI schedules reject before scientific output creation.
- Parent inspected the exact Cdiff: no unintended ionic-kernel/geometry modification was found. The stronger grid checker was added separately without modifying running pinned sources or restarting the campaign.

These checks validate the implemented finite comparison. Earliest checkpoint ordering is not a complete every-neighbor causal-wave proof; finite state is not physiological calibration.

## Actual compute and preservation

|Case|Measured subprocess wall seconds|
|---|---:|
|Near regular|196.714|
|Near omit6|195.891|
|Far regular|197.947|
|Far omit6|198.678|
|Uniform regular|195.653|
|Uniform omit6|195.617|

Total case time1180.500s, approximately19.68minutes, excluding separate compilation/verifier/figure overhead. The campaign terminated with exit0; no case hit600seconds. Its116retained files total11,931,041bytes including final results, below100MiB. These are measured local serial CPU times, not GPU throughput or a future timing guarantee.

`DELIVERY_HISTORY_EXECUTION_MANIFEST_V1.json` hashes all116campaign files plus code, tests and numerical-verification records. `DELIVERY_HISTORY_GRID_VERIFICATION_V1.json` is the separate strong PASS record. Original source hashes stayed fixed during all six runs. A preliminary Python bytecode-cache check encountered a macOS sandbox cache-permission error; actual imports, Ccompilation, tests and simulations subsequently ran without changing permissions or installing software. An optional hardware-info query was denied; no unverified CPUmodel is assigned to these timings.

## Scope and stopping decision

Original item57now has a genuinely new, independently verified at-home delivery-history endpoint comparison, not a reused geometry result or merely a plan. It adds a recovery-history counterexample to final-window classification and a finite robustness observation for the near mask.

Stop at the six authorized cases. The new omitted-pulse contrast is **coarse-grid only**; prior regular-pacing refinement does not establish post-interruption convergence. The specification permits a separately reviewed matched fine-grid pair if that specific phase result warrants a new decision. None was automatically launched. Thirty-beat ionic drift, general initialization independence, alternative schedules, actual S10s mixed-genotype action and clinical translation remain unestablished. Those limits do not erase the completed finite result or justify calling it treatment evidence.