# Felisyn full-model experiment package 006

13 October 2026. Two completed preparatory steps: source compatibility investigation and a frozen full-model experiment contract. No full-model neural simulation has been performed with this package.

## What was checked

- All 82 released parameter sets pass the pinned framework's schema/type checks without constructing a network.
- Four synthetic cases exercise the unchanged historical and current connector weight methods. `self_connections=True` preserves the old unsuppressed diagonal rule; `False` zeroes the diagonal only for the same named population. Synthetic weights are not biological observations.
- The runtime audit passes 78 checks, including 187 installed Python files whose hashes match the Mozaik, PyNN and imagen source trees. NEST 3.4 and the step-current module import successfully. This does not verify compiled binaries byte for byte.
- Ten contract tests and twelve existing metric tests pass. The existing four-CPU, approximately 8 GiB host is correctly rejected by the hardware preflight, before simulation.
- The workflow command line is parsed by the actual pinned framework without invoking a model.

The initial runtime checker assumed editable installs and rejected three package locations. Inspection showed ordinary copied installations. The corrected checker compares every installed Python source file with its corresponding pinned checkout file. `checks/runtime-attempt-01.json` preserves the rejected assumption; `checks/runtime.json` contains the passing source-hash check. A source inspection also corrected the new wrapper to read the stored `simulation_log` annotation, rather than call a nonexistent getter. Neither correction involved neural simulation.

## Compatibility decision

The inspected history adds the required flag in Mozaik commit `c8261569a858ebbd3b91621b2d863d8059c89f37` on 21 October 2024. Its parent is `d2ea60ce38f824d12ff0886e051ab3ad73939079`; the retained June 2023 file has the same old weight behavior. The new `False` branch explicitly replaces the diagonal weight with zero. We retain the released `True` setting. This identifies a code-level relationship, not the exact framework revision used by the authors' archived March 2024 run. It does not establish full-network or dynamical equivalence.

The unmodified released parameters contain PyNN seed 5 and `mpi_seed=1023`. The pinned controller does not forward that field as `mozaik_seed`, so its effective Mozaik default is 513. Adding a new top-level field would violate this model's exact schema. We preserve the source and document the effective seed pair instead. Trial 0 is a stimulus label; this experiment has one replicate.

## The frozen experiment

`run-contract.json` is the machine-readable specification. It hashes every supplied upstream model file and pins five runtime source checkouts and the observed Python packages. The `model/` directory is the unchanged LSV1M v1.0 tree with its MIT license. Full precision reference arrays, exact historical runtime, complete earlier stimulus history and the 4.09-versus-3.56 nS conductance discrepancy remain unresolved.

Use **general `param/defaults`**, whose four cortical recorder definitions match the inspected archive, with `create_experiments_spont`. The alternative `param_spont` is retained as upstream source but is not the declared recipe. One 40,320 ms InternalStimulus is measured in four cortical populations at a 0.1 ms integration step and 1 ms analog sampling. Spike-count correlation uses 4,032 bins of 10 ms. The pinned workflow inserts an initial 150 ms blank, a 150 ms reset blank before the final stimulus, and a 150 ms terminal blank: 40,770 ms total if it completes. This total is source-derived; no full-run timing has been observed. Do not trim an additional burn-in from the analysis window.

The six measurements are firing rate, ISI CV, within-population spike-count correlation, signed membrane voltage, excitatory conductance and inhibitory conductance. Compute population mean and SD using the original conventions. Retain eligible-cell and pair counts; pairs that share neurons are not independent replicates.

## Environment and preflight

Use the fresh-host build recipe, source pins and transitive lock in [execution-003](https://www.felisyn.com/research/execution-003/README.md). That consolidated installation recipe has not yet been executed end to end on a second clean machine. OS repositories are not snapshotted. This package is not a container or a byte-identical OS image. Use `sources-resolved.json` from execution-003; its older `source-commits.tsv` contains the rejected initial PyNN candidate.

On an approved host, place this package in `/opt/felisyn/protocol-006`, owned by the unprivileged research user. As that user:

```bash
source /opt/felisyn/venv/bin/activate
source /opt/felisyn/venv/bin/nest_vars.sh
export MPLBACKEND=Agg
cd /opt/felisyn/protocol-006
python -m unittest -v test_contract.py test_metrics.py
python verify_runtime.py --sources /opt/felisyn/sources --output runtime-check.json
python check_compatibility.py --model /opt/felisyn/protocol-006/model --output compatibility-check.json
mkdir -p /opt/felisyn/full-runs
python launch.py --output /opt/felisyn/full-runs
```

Default preflight checks hardware only and does not simulate. Exit 2 means the machine falls below a declared resource floor. Hardware readiness alone does not verify runtime/source compatibility. `--execute` separately verifies source checkout IDs, tracked-file cleanliness, packaged model hashes, installed versions and package source hashes, then parameter schemas. It creates a fresh output directory and retains its contract hash and command.

Only after the paid allocation is approved, use:

```bash
python launch.py --output /opt/felisyn/full-runs --execute
```

The wrapper requests 16 MPI ranks and one thread each. It runs the unchanged model and spontaneous experiment factory, and saves raw data before analysis. Original heavy plotting is deferred. GNU timeout bounds the simulation process to three hours, followed by a 60-second termination grace period. Exit zero alone is insufficient: the saved workflow marker must report no explosion. The framework's firing-rate guard is evaluated after a segment; it is not a continuous early-abort monitor. No promise of full-run success or resource sufficiency follows from passing preparation checks.

## Analyze a completed run

First inspect `launch.json`, the saved parameter record and runtime receipts. Confirm the requested model recipe and seed/MPI configuration actually ran. Replace `DATASTORE` below with the newly created directory that contains `datastore.recordings.pickle`, and `RUN_ID` with that run's actual identity. Do not substitute a reference fixture.

```bash
python validate_datastore.py --run-label RUN_ID \
  --datastore DATASTORE --figure-source model/visualization_functions.py \
  --framework /opt/felisyn/sources/mozaik --output analysis-RUN_ID
python compare_reference.py --receipt analysis-RUN_ID/analysis-parity.json \
  --output analysis-RUN_ID/printed-reference-comparison.json
```

The analysis is derived from validation-005. This version requires an explicit run label, uses neutral saved-data provenance, and rejects nonfinite comparisons even if both calculations are undefined. It does not invoke a simulator or save into the input datastore. It compares the original numerical statements of Figure 4 with independent helpers and records before/after file hashes. No full paper figure is regenerated. Pickle files require the matching environment; inspect provenance before loading them.

The reference comparator requires the full 40,320 ms window, 4,032 bins, four populations, unchanged data and 48 finite agreeing mean/SD comparisons. It rejects short-run fixtures, missing or duplicate metrics, and failed checks. It reports 24 mean differences against `reference-means.csv`, transcribed from printed Figure 4 labels in benchmark-004. The half-last-printed-unit flag describes rounding proximity. It is **not** a biological equivalence threshold or paper-reproduction pass. The returned scientific pass remains null even for a perfect printed-label match.

A generalized-analysis regression is retained under `checks/analysis-regression/`, using the previously saved 105 ms small-model run. It verifies software behavior only. That fixture cannot pass the full-window reference comparator.

## Compute and stop policy

Proposed temporary separate host: DigitalOcean Memory-Optimized Regular, 16 vCPU, 128 GiB RAM and 400 GiB SSD. The [published rate](https://www.digitalocean.com/pricing/droplets) checked on 13 October 2026 is USD 1 per hour (USD 672 monthly cap). Six hours allocated to setup, simulation, analysis and verified export would cost about USD 6 before tax and extras. Availability is not yet checked. This is an estimate, not a measured runtime or automatic spending cap.

Hardware floors are engineering choices for that allocation: 16 CPUs, 120 GiB reported total RAM, 110 GiB available RAM and 60 GiB free disk. They are not empirical minimum requirements. The paper reports using 16 processes and 128 GB; our cloud hardware differs. Never infer full-model memory demand from the tiny regression.

The process timeout does not stop cloud billing. DigitalOcean [bills powered-off Droplets until destruction](https://docs.digitalocean.com/products/droplets/details/pricing/). At the six-hour review point, preserve logs and results, verify an off-host copy, then destroy only the temporary machine or obtain approval for an extension. Cloud lifecycle operations are intentionally absent from the launcher. No new paid host has been created, resized or authorized by this package.

## Evidence and reuse

`manifest.json` records hashes and sizes; `protocol-evidence.zip` contains the public reproducibility package. `draw_protocol.py` produces the 1920 × 1200 PNG and SVG from the contract and check receipt. It is a sequence/method diagram, not neural activity or anatomy; timeline widths are not proportional.

Sources: [Antolík et al. 2024](https://doi.org/10.1371/journal.pcbi.1012342), [LSV1M v1.0](https://github.com/CSNG-MFF/mozaik-models/tree/04bf11b72ff5f78e66f28e037016e8e33b09c12e/LSV1M), [Mozaik v0.4.0](https://github.com/antolikjan/mozaik/tree/f5c09a0eccc5786daf78881d883bbce586a00e6d), and [the flag-introducing commit](https://github.com/antolikjan/mozaik/commit/c8261569a858ebbd3b91621b2d863d8059c89f37). Model MIT and Mozaik CeCILL license text are included. Historical files, commit metadata and original retrieval URLs are retained in `sources/`. Felisyn diagrams are new explanatory drawings; no source-paper image is reproduced.
