[vincentchoqueuse]

Seven commits that tell one story: what the tutorials needed was a storage convention, not an engine. The branch keeps the full record — an Experiment object was built, measured, extended, and removed the same day — because the measurements are what justify the destination.

The destination

Every simulation section in the tutorials now reads in the same three parts, under the same markers, in the same order:

# --- metrics, pre-allocated ---
# --- simulation loop ---
# --- results: tables and figures ---

The convention, written into the tutorial skill and applied to nine sweeps across the MIMO, OFDM, optical and AWGN pages:

• the methods and the metrics are declared first, as the ordered dictionaries they are;
• storage is pre-allocated to zeros, one np.zeros(n_points) per (metric, method), indexed by name on both levels — never a 2-D array indexed by position, whose countable column offset is the classic silent bug and was live in one of these scripts (ber_data[index_snr, index_comp+1]);
• the loop draws one child seed per point from a master seed (SeedSequence.spawn, D6/D35) and fills the arrays — the loop stays visible, because the loop is the pedagogy;
• the display comes last, from the same dictionaries the loop filled — each inner dict is exactly the curves that print_data and plot_data render.

The detour, and why it is in the history

An Experiment(config, parameter=, values=, seed=) object ran the loop behind a simulate callback, per a full specification: auto-kept seeds, save= selection, grouped observations, result.print()/result.plot(). Measured against the four studies it ran: −2, +5, +11, −16 lines — shorter nowhere — and it hid exactly the loop the tutorials exist to teach. Removed before any release.

What survives from it is what mattered:

• Three studies were reproducible by nobody, their author included — the two MIMO detector sweeps and the CD compensator comparison ran with no seed at all. All are seeded now, in the script where the reader can see it.
• The migration is verified to the digit: the DBP page's tables are identical between the object and the loop, because the in-script seeding reproduces the object's derivation exactly.

What the review pass found on the way

• A list comprehension in one_shot_ofdm.py that the tutorial rules ban, hiding a monte_carlo call inside an expression.
• The last append-and-convert accumulations, a hand-rolled aligned print loop that print_data replaces, a dead five-line dict-to-array conversion, and a bits-per-symbol column now computed by np.log2(orders).astype(int) — NumPy over iteration wherever the loop taught nothing.
• A drift: the OFDM page still pasted the runtime table in a format the script had stopped printing.

The launch amplifier also moves before the upsampler in the three optical scripts that had it after the pulse shaping — verified equivalent to 1.6e-14 on the symbols, and the fourth script already did it — so the amp/1/amp pair now acts on the same quantity at both ends of the chain.

Verification

All five re-run tutorial scripts produce physics identical to the digit; wall-clock tables are re-pasted from the runs. ruff check ., 901 tests + 1155 subtests, doctests, pyright against NumPy 2, the literalinclude guards, and sphinx-build -W all pass.

---

Generated by Claude Code