# Protocol 023's early stop: public summary of the exploratory diagnosis

Dated 8 November 2026. Simulation data only. Nothing here is about cat cortex.

Protocol 023 ended by its declared early stop, and its declared outcome is **NOT INTERPRETABLE** (`../results/results.json`). The analysis wrote that file at 01:55:52 UTC on 8 November, and the diagnosis began after it: its folder was made at 02:00:03 UTC. This summary describes that diagnosis. It does not change, qualify or repeat the outcome, and nothing in it is a protocol result. It is not a software regression, a numerical agreement test of a new build, a paper reproduction, a biological validation or a controller test. It is about an execution fault on the laptop that ran the model.

Each claim carries a label:
- **measured**: read from a file or a run;
- **inferred**: reasoned from measurements;
- **untested**: a candidate or plan that no test has checked.

A skeptical review read the diagnosis and its records afterwards. Its verdict was "holds with corrections". The corrections are applied here, and the places they changed are marked "(corrected after review)".

## A procedural gap

The spec's fences say: "Any exploratory reading of 023 data gets its own manifest, frozen before it reads any 023 export." This diagnosis had no such manifest. It read the s7 export, the model's output folder for s7, the queue's receipts and the protocol's records without one. **measured**: no manifest exists in the diagnosis folder or beside it.

Its records say that it computed no plaid quantity (no XOS, Dc or ds), and the outcome was written before it began. That limits what the gap can have touched, but it does not close it. A new declared revision has to say how it treats this diagnosis.

## 1. What stopped the protocol

- **measured.** Run s7 (pynn_seed 7, run name `m23-s7-intact-a1`) ran on 8 November from 00:47:21 to 01:53:28 UTC and exited 0. Its memory peak was 3.186 GiB under the 3.7 GiB cap, and it built 14,712,684 connections. Export SHA-256 `736c2b3c…41a2` (`../exports/s7.json.gz`).
- **measured.** The server check, which runs on the laptop right after each run, refused the export on one kind only, the LGN lead-in identity. X_ON: 1,800 of 1,800 trains equal the reference. X_OFF: 1,799 of 1,800. The one differing cell is 2738. The run, seed, weight, kernel and LGN band checks passed. The local check, after the fetch, refused the same cell, with identical check values.
- **measured.** Cell 2738's spikes in the 504 ms lead-in: 72.7, 77.2, 116.4, 152.5 and 270.2 ms in s7, against 72.9, 119.5 and 270.6 ms in the reference.

## 2. Repeats with s7's exact inputs

| Run | What it repeats | Run name | Exit | Wall time | Memory peak | Lead-in check | Compared with | Result |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| d1 | s7's first 5 steps | `m23d-s7-s7first5` | 0 | 10 min 39 s | 2.746 GiB | 3,600 of 3,600 | s7, 0 to 2,289 ms | only cell 2738 differs among the LGN cells |
| d5 | all of s7 | `m23-s7-intact-a1`, reused | 0 | 1 h 02 min 36 s | 3.179 GiB | 3,600 of 3,600 | s7, 0 to 35,133 ms | only cell 2738 differs among the LGN cells |
| rate-r1 | all of s7 | `m23d-s7-rate1` | 0 | 1 h 32 min 21 s | 3.194 GiB | equal to d5 | d5, 0 to 35,133 ms | all 18,600 trains equal |
| rate-r2 | all of s7 | `m23d-s7-rate2` | 98 | stopped at 18 min | 2.717 GiB | none | none | stopped at the disk floor |

- **measured.** d1 and d5 agree on all 18,600 trains over 0 to 2,289 ms.
- **measured.** d5 reused the protocol's run name, `m23-s7-intact-a1`, to repeat s7's arguments exactly. It is a diagnostic run and never a protocol attempt (corrected after review). d1 and the rate runs used `m23d` names.
- **measured.** rate-r2 stopped because another program on the laptop filled the disk below the 3 GiB floor. It is not a model result. rate-r3 was not run.
- **measured.** The engine files, the model folder and the 315 hashed runtime files were byte-identical to s7's for the repeats. The recorded environment variables equal s7's except the export and sequence paths. s7's receipt does not record PATH, so PATH equality with s7 is not verified (corrected after review).
- **measured.** Some things a repeat cannot hold fixed: the working folder, Python's hash seed (unset in every run), the layout of memory, and the machine's load. The load average was 6.51 at s7's start, 2.9 at d1's and 3.4 at d5's.

## 3. Cell 2738 in s7 against the repeats

- **measured.** Over the whole sequence, d5 has 561 spikes from cell 2738 and s7 has 608. The two trains are equal again from 14,344.6 ms: the last 327 spikes, to the end of the sequence at 35,133 ms, are identical. Before that, d5 has 234 spikes and s7 has 281.
- **measured.** s7 has more spikes than d5 in every 1 s bin from 0 to 14 s: 21 against 16, 16 against 13, 38 against 34, 17 against 15, 16 against 12, 19 against 18, 19 against 17, 13 against 9, 21 against 14, 11 against 10, 24 against 19, 16 against 13, 20 against 17 and 27 against 24.
- **measured.** For each of d5's 234 spikes before the re-sync, the nearest s7 spike is earlier in 181 cases, at the same time in 26 and later in 27, by −7.3 to +5.9 ms. 41 s7 spikes have no d5 spike within 5 ms, and 3 d5 spikes have no s7 spike within 5 ms (`recount/a2.log`).
- **measured.** rate-r1's train for cell 2738 is identical to d5's, all 561 spikes.
- **measured.** The X_ON totals are equal in s7 and d5 (1,882,087 spikes). The X_OFF totals differ by 47 (1,238,217 in s7, 1,238,170 in d5), which is cell 2738's difference.
- **measured.** Every cortical train differs between s7 and d5 over the whole sequence. The first cortical difference is at 75.2 ms (layer 4 excitatory cells; 77.1 ms in layer 4 inhibitory, 80.0 ms in layer 2/3 excitatory and 79.0 ms in layer 2/3 inhibitory cells), after cell 2738's first differing spike at 72.7 ms. Before 504 ms many cortical trains are still identical: 3,292 of 6,000, 138 of 1,500, 1,466 of 6,000 and 11 of 1,500 in the same order (`runs/d5-s7repeat/vs-s7.json`). **inferred**: the cortex differs because of cell 2738, the only input that differs.

Figure 1 in `../figures/` shows these trains.

## 4. The lead-in input does not depend on the sequence or the seed

- **measured**, by recomputation outside the simulator. Mozaik's own filter classes, run without NEST, gave lead-in input entries that are bitwise identical for all 1,800 cells of each LGN sheet. This held at seed 5's and seed 7's cell positions, with s7's sequence and with the pilot's, and with 1 and 4 threads of the linear algebra library. The steady X_OFF value is −59.671 pA.
- **measured.** The first input entry that depends on a cell's position is index 72 (504.2 ms). The first that depends on the sequence is index 123 (861 ms).
- **measured.** In the lead-in the input is the luminance term alone: the contrast filter's output is between about 4e-21 and 5e-19 in size. The luminance filter's output is steady at about −1.81e-6 from index 28 (its peak, −5.15e-6, is at index 5), against a gain constant of 1e-8, so the luminance gain is saturated, with a ceiling of 60 pA (corrected after review: the earlier text gave the range −4e-6 to −5e-6).
- **measured.** 18 exports other than the pilot's own, which is the reference's source and passes by construction, passed the lead-in identity exactly: 16 retained exports (14 at seed 5 and 2 at seed 6; four of them ran on the laptop, G4 and three 5-step runs) and the repeats d1 and d5 (corrected after review: the earlier count of 19 included the pilot's export).

## 5. What the fault could have been

- **inferred.** The deviation is not a fixed result of seed 7, s7's sequence, the files, the recorded environment, the chain of commands or the run length: the same declared inputs gave the reference lead-in in d1 and d5, and rate-r1 equals d5.
- **inferred.** A shift in the simulator's random streams is ruled out: the 899 other LGN cells that the simulator runs on cell 2738's thread are unchanged, and cell 2738 itself is equal again after 14,344.6 ms, which needs the same noise.
- **inferred.** The fault acted only on cell 2738's own input or state. No connections project onto LGN cells (from reading the model's code). It began after the 150 ms blank before the sequence, which is identical in all runs, and by 72.7 ms into the sequence. It ended between about 14,280 and 14,344.6 ms. Throughout, the cell was more depolarised: its spikes came mostly earlier, and it had extra ones.
- **inferred, found after the fact.** Input frame k drives the cell from 7k + 0.2 ms. A memory page on this laptop is 16 KiB, which holds 2,048 numbers of 8 bytes. The first page of one cell's input buffer covers the cell's input up to 14,336.2 ms (Mozaik's buffers) or 14,343.2 ms (the copies handed to the simulator). Both ends fall inside the window above. That fits with no free parameter, but it is one event, found after looking, so it suggests and does not prove.
- **inferred.** What fits the data best is that cell 2738's luminance term read as zero for input indices 0 to 2,047. That would add about +59.7 pA of drive, about +2 mV at rest, and leave the grating's modulation intact: earlier and extra spikes, still locked to the grating, as seen.
- **untested.** Where the zeros came from. Three candidates fit equally well (corrected after review: the earlier text named only the first):
  - the first page of the cell's long-lived luminance buffer lost its contents;
  - the first page of one of two temporary arrays of the same size inside the gain computation, the product and the quotient, read back as zeros;
  - a data pointer or offset moved by 16 KiB onto a page of zeros.
  The long-lived buffer sat unused for most of the run while most of the process's memory was, by inference from the memory records, compressed or swapped. That favours the first. The temporary arrays point instead to memory being reused.
- **untested.** How a page could come back as zeros: (a) macOS's handling of freed and reused memory, where the memory wrapper sets `MallocSpaceEfficient=1`, a setting that exists only on the laptop's run path; (b) the memory compressor or swap. A hardware memory error is unlikely (**inferred**): a whole 16 KiB page of exact zeros looks like a virtual memory event, not a flipped bit (corrected after review: the earlier text listed hardware as an equal candidate).
- **measured, with a limit.** Buffers of these sizes started on a page boundary in 400 of 400 allocations each. That test ran in a small fresh process with only `MallocNanoZone=0` set, not with `MallocSpaceEfficient=1` and not inside a model run (corrected after review).
- **untested.** A test that sets cell 2738's luminance term to zero over indices 0 to 2,047 in a short run is designed and has not run. If its spikes matched s7's, it would confirm that the luminance term was zero over those indices. It would not show that a page was lost (corrected after review).

## 6. How often

- **measured.** Full-length runs on the laptop that were checked spike by spike: G4 (against the server pilot, all 18,600 trains): clean; s7: a fault; d5 and rate-r1 (against each other): clean. **1 fault in 4 runs.**
- **inferred.** With 1 in 4, the per-run rate is close to unknown: a 95 % interval (Clopper-Pearson) runs from 0.6 % to 81 %.
- **measured.** Of the 7 laptop runs of the 2 mm model whose lead-in was checked, 1 was refused. Four of them were 5-step runs, whose input buffers are smaller than one memory page, so under the page candidates they are not comparable (**inferred**; corrected after review: the earlier text used 1 in 7 as the rate).

## 7. What the early stop did

- **measured.** The lead-in check caught a real execution fault: s7's export is not what s7's declared inputs produce.
- **inferred.** The check sees a fault only if it changes an LGN cell before 504 ms. Under the page candidates it sees about one of six live input pages per LGN cell (the first page of the luminance term) and none of the cortex's memory. The same kind of event elsewhere would pass unseen.
- **measured.** The early-stop rule took the refusal as a fault shared by every instance. The design knew this could be wrong and accepted the cost. Risk 14 of the spec reads: "A fault of that one instance (an export damaged as it was written, say) would then end the protocol where the replacement rule could have replaced it; the rule accepts that cost for the server time it saves when the fault is shared." (corrected after review: the earlier text said the design had assumed determinism).
- **inferred.** The measured fault rate makes that cost higher than the design expected. The spec's sentence "the simulation is deterministic, so they change only the time" is about heat and other work changing the run time. Its premise does not hold for every run on this laptop.
- **inferred.** NOT INTERPRETABLE stands. s7 would not be a valid instance under any reading: its LGN input differed from the declared computation for 14 s.

## 8. Options for a new declared revision

The owner decides. This summary lists the options without a recommendation. The diagnosis's own draft recommended the first. The review found that, as drafted, it would pass runs from a run path that sometimes corrupts input unseen, that it wrongly said the checker file would stay unchanged, and that it used a fault rate taken over the wrong runs.

**Shared by every option.** Seeds 5, 6 and 7 are excluded: 7 is now built and its output seen. Seeds 8 to 14 are unbuilt, and their declared sequence files can be reused byte for byte. The seed search would run again over the laptop's run folders. The spec's determinism sentence would give way to the measured statement, and this diagnosis would be cited as exploratory history, with its procedural gap. One full run took 1.04 to 2.00 hours on this laptop (d5 1.04, s7 1.10, rate-r1 1.54, G4 2.00), so six take about 6 to 12 hours, one at a time under the 3.7 GiB cap.

| Option | What changes | Model time | What it catches and misses |
| --- | --- | --- | --- |
| A. One rerun after a lead-in refusal | A refused first attempt gets a second with identical inputs; a refused second attempt stops the queue, as does a third refusal in all. The queue, the launcher, the summary and the checker file itself change (its early stop and summary live there), with tests and a review. | 6 to 12 h, plus 1 to 2 h per rerun | Stops on a fault that recurs. A fault that comes and goes passes whenever it misses the lead-in, which sees about one of six input pages per LGN cell and no cortex. |
| B. Every instance twice | Each instance runs twice with identical inputs, and the two exports must agree on all 18,600 trains; a third run breaks a disagreement. A frozen comparison tool is added. | 12 to 24 h, plus 1 to 2 h per disagreement | Catches a fault anywhere in the spike output, and measures the rate. About 1.4 GB of disk for 13 runs unless the duplicates' model folders are deleted, which would have to be declared. |
| C. Check every LGN input array | Each cell's input arrays are computed offline and compared with hashes taken inside the run by a read-only observer. | about 2 h for an identity run, then 6 to 12 h, plus offline time (untested) and reruns | Checks the buffers that the page candidates name. Adds code to the run path, and faults on the cortex side stay unseen. |
| D. Tolerate a few differing lead-in cells | The lead-in check passes when at most k cells differ. | 6 to 12 h | Analyses runs known to carry a fault; k would be chosen after seeing s7; a systematic fault that touches few cells passes. |
| E. Find and remove the cause first | The zeroing test, then a change such as running without `MallocSpaceEfficient=1` (which may not fit under the cap, untested), then an identity run and enough repeats to show the fault gone. | tens of hours at the measured rate | Removes the fault if found; slow, and any change to the wrapper needs a new identity proof. |

**inferred.** Under option A, if every fault showed in the lead-in, and runs failed independently at a fixed rate, the protocol would end unread in about 39 % of attempts at 1 fault in 4 runs, and 59 % at 1 in 3. If fewer faults show in the lead-in, it stops less often and more faults pass unseen. Under option B at 1 in 4, about 4 pairs in 10 would need a third run.

## Records

The records behind this summary are in `runs/` (receipts, lead-in checks and comparisons of d1, d5, rate-r1 and rate-r2), `recount/` (the recount of cell 2738 and its script), `tools/compare_prefix.py` (the comparison script) and `../figures/`. The rest of the diagnosis, its offline recomputations, its census of retained exports and its working notes, stays in the project's private workspace, unchanged. Paths in these records read `<home>` where they held the laptop's home folder (`../README.md`, Redactions).
