# Logbook — 2026-04-13 — Preflight results **Context.** Follow-up to `2026-04-13-preflight-kickoff.md`. Executes the §4a protocol of [../README.md](../README.md) (v0.3). Scope: single-anchor cross-engine validation at (δ₀ = 0, |α| = 0, φ_α = 0); resolution of the Q3 and Q5 residual questions; S1/S2/S3 budget estimate. **Verdict.** *Preflight passes, with one engine reclassified.* The HDF5 "adaptive-learner" dataset is marked synthetic (`export_version = "0.4-synthetic"`) and produces an unphysical Bloch vector (|Bloch| = 1.032 > 1) at the anchor; it is therefore excluded from the preflight gate as a non-engine artefact rather than failed. The JSON-uniform legacy and the fresh candidate run agree to < 10⁻⁶ because they are the same engine — this is a baseline reproducibility check, not an independent cross-check. Main sweep is cleared to proceed on the candidate engine; the HDF5 mismatch is reclassified as a documentation/data-provenance issue handled separately (§4). ----- ## 1. Environment and engine identification - Python: `/Users/uwarring/opt/anaconda3/bin/python` (Anaconda). NumPy 1.23.5, SciPy 1.10.0, h5py 3.7.0. QuTiP absent (not imported — the module docstring of [../../scripts/stroboscopic_sweep.py](../../scripts/stroboscopic_sweep.py) lists it but the code uses `scipy.linalg.expm` only). - Candidate engine: [../../scripts/stroboscopic_sweep.py](../../scripts/stroboscopic_sweep.py) CODE_VERSION `0.8.0`, imported as a library. **No modifications.** - Legacy HDF5: [../../data/alpha_0/detuning_scan.h5](../../data/alpha_0/detuning_scan.h5). - Legacy JSON: [../../data/runs/alpha_0_default.json](../../data/runs/alpha_0_default.json). ## 2. Q1 / Q5a — Bloch-observable exposure **Finding.** `stroboscopic_sweep.py` **already** computes and returns ⟨σ_x⟩, ⟨σ_y⟩, ⟨σ_z⟩ at every detuning point. Evidence: - `compute_observables()` at [../../scripts/stroboscopic_sweep.py:205-240](../../scripts/stroboscopic_sweep.py#L205-L240) computes `sx = 2·Re(ρ_{du})`, `sy = −2·Im(ρ_{du})`, `sz = ρ_{uu} − ρ_{dd}` from the full spin–motion state vector after the N-pulse sequence. - `run_single()` at [../../scripts/stroboscopic_sweep.py:397](../../scripts/stroboscopic_sweep.py#L397) serialises all three into the output payload. - The JSON legacy file confirms this — `payload.data` contains `sigma_x`, `sigma_y`, `sigma_z` arrays of length 201. **Important notational note.** The script's `coherence` field is the total Bloch-vector length √(σ_x² + σ_y² + σ_z²), not the WP's transverse coherence |C| = √(σ_x² + σ_y²). Driver code must compute |C| explicitly: `|C| = (sx**2 + sy**2)**0.5`; do not use `coherence`. **Consequence for Q5.** *No extension of `stroboscopic_sweep.py` is required for anchor-point work.* For the slice sweeps (§4 deliverables), a thin Python driver calling `run_single()` in a loop over the outer axis (|α| or φ_α) is sufficient — the detuning axis is already a native output. Decision recorded in §7 of this entry. ## 3. Q3 — Lamb–Dicke closed-form prediction **Finding.** A systematic grep across the repository (pattern `lamb[\s_-]*dicke|small.?eta|eta\s*=\s*0\.0|linear.*coupling|LD.limit`, case-insensitive) returned matches only in narrative markdown/HTML ([../../reference.html](../../reference.html), [../../framework.html](../../framework.html), [../../ideal-limit-principles.md](../../ideal-limit-principles.md), etc.) **No Python closed-form implementation exists.** **Decision.** R1 will be implemented **numerically at small η**: reuse `stroboscopic_sweep.py` with `eta = 0.04` (10× reduction; quadratic correction suppressed to 0.2% of linear term, well below the 20% at η = 0.397 noted in dossier §1.2). No extrapolation; η = 0.04 is close enough to linear that the residual nonlinearity is ≪ the other residuals being probed. If needed, a second reference run at η = 0.02 provides a convergence cross-check. R2 (instantaneous pulse) is more delicate — see §7. ## 4. Three-engine comparison at anchor Measured at (δ₀ = 0, |α| = 0, φ_α = 0), nominal parameters per v0.3 §2. ### 4.1 Candidate engine (fresh run) Narrow detuning grid {−0.1, −0.05, 0, +0.05, +0.1} × ω_m, 5 points, wall time 40 ms total (7.9 ms/point, nmax = 30). ``` σ_x = +0.000000000 σ_y = +0.916565752 σ_z = +0.128438170 |C| = √(σ_x² + σ_y²) = 0.916565752 arg C = +90.000° (to within floating-point precision) |Bloch| = 0.925521011 (< 1, physical) max_fock_leakage = 9.6×10⁻⁵¹ (converged = True) ``` ### 4.2 JSON-uniform legacy 201-point scan over det ∈ [−6, +6] × ω_m, anchor at index 100 (det = 0). ``` σ_x = +0.000000 σ_y = +0.916566 σ_z = +0.128438 arg C = +90.000° |Bloch| = 0.925521 ``` **Agreement with candidate:** identical to JSON print precision (6 decimals) across all three Bloch components. The two are the same software (CODE_VERSION 0.8.0 of `stroboscopic_sweep.py`); this is a baseline reproducibility check, not an independent cross-check. ### 4.3 HDF5 adaptive-learner legacy — reclassified File attributes: | attr | value | |---|---| | `eta` | 0.397 | | `omega_m` | 1.3 | | `omega_rabi` | 0.3 | | `omega_eff_carrier` | 0.277266 | | `alpha` | 0.0 | | `n_thermal` | 0.001 | | `aom_setup` | fast | | `n_points` | 120 | | **`export_version`** | **`0.4-synthetic`** | | `contrast_x` | 0.519188 | | `contrast_y` | 0.164944 | | `contrast_z` | 0.293080 | Detuning range **[−0.25, +0.25] × ω_m** (narrow carrier zoom — not a Doppler-scale scan). Anchor at det ≈ 0.002 × ω_m (no exact zero on the grid): ``` σ_x = +0.988951 σ_y = +0.025695 σ_z = +0.294671 |Bloch| = √(0.989² + 0.026² + 0.295²) = 1.032 ← UNPHYSICAL (>1) contrast_z (stored) = 0.293080 = (max−min)/2 over the scan ``` **Verdict.** The HDF5 file is a **synthetic placeholder**, explicitly labelled as such via `export_version = "0.4-synthetic"`, and produces an unphysical Bloch vector. It cannot serve as a preflight comparator. Additional anomalies consistent with synthetic/placeholder origin: - `σ_x ↔ σ_y` approximate swap relative to the candidate/JSON (rotation axis appears around +Ŷ rather than +X̂); a real engine difference of this kind would be a 5th candidate cause ("Raman phase convention") but the synthetic label short-circuits this diagnosis. - `σ_z` differs by a factor ≈ 2.3 from the candidate (0.295 vs 0.128) — consistent with a hand-tuned synthetic dataset rather than a real run. - The dossier §1.4 claim of "HDF5 adaptive-learner contrast_z at α=0 = 0.61" does not match this file's stored `contrast_z = 0.293`. Either the dossier refers to a *different* HDF5 file we do not have, or the numbers in the dossier are themselves out of date. *Action:* flag for dossier update in the follow-up task (logged, not executed). **Gate consequence.** Guardian's four-candidate disambiguation (§4a) does not apply — the HDF5 mismatch is not an engine disagreement but a non-engine data artefact. The three-way gate collapses to: - Legitimate comparator pair: (candidate, JSON) — agree to < 10⁻⁶. - Non-engine artefact: HDF5, excluded. Main sweep **is cleared** to proceed. ### 4.4 Note on the per-pulse Bloch vector requirement §4a of v0.3 specifies recording the per-pulse Bloch vector at indices 1 … 22. The legacy JSON and HDF5 store only the *final* Bloch vector after N = 22 pulses (the script's `run_single()` output is final-only). Retrofitting per-pulse tracking requires a minor modification to `run_single()` (capture `psi` inside the pulse loop at [../../scripts/stroboscopic_sweep.py:422-423](../../scripts/stroboscopic_sweep.py#L422-L423)). **Decision.** Not done in preflight. The final-only three-engine comparison already disambiguated: JSON ≡ candidate (by construction); HDF5 is synthetic and excluded. Per-pulse tracking is retained as a diagnostic tool for the main sweep if residuals against R1/R2/R12 demand it — recorded as an open action, not a blocker. ## 5. Guardian flag closures | Flag | Closure | |------|-------------------------------------------------------------------------------------------------------------| | 1 (budget) | Per-point wall time 7.9 ms at nmax = 30. S1 = 4.8 s; S2 = 3.1 min; S3 = 5.1 s. Well under budget; no triage required even at 64 φ_α points. Scaling with nmax is quadratic in matrix-exp cost; anticipate ~2–3× slowdown at nmax = 60 (needed for |α| = 5). S2 at nmax = 60 still under 10 min. | | 2 (Jacobian) | Deferred to main sweep; finite-difference step plan per kickoff entry unchanged. | | 3 (tolerance) | `‖ΔC‖ < 10⁻⁴` satisfied trivially (candidate ↔ JSON < 10⁻⁶). | | 4 (injectivity) | Deferred to main sweep; decision "cond(J) on raw C, not normalised" stands. | | 5 (Q3, Q5) | Q3 resolved (§3 of this entry): R1 = numerical small-η (η = 0.04). Q5 resolved (§2): no engine patch; thin driver. | ## 6. nmax convergence note The JSON run used nmax = 30 for |α| = 0. For |α| up to 5, ⟨n⟩ = 25 and the tail populations require nmax ≥ 3·⟨n⟩ ≈ 75 to stay under 1% Fock leakage. The engine's own `max_fock_leakage` diagnostic will signal this during main sweep. Plan: **nmax scaling** with |α|: | |α| | ⟨n⟩ | nmax | rationale | |----|-----|------|---------------------------------------| | 0 | 0 | 30 | more than sufficient | | 1 | 1 | 30 | sufficient | | 3 | 9 | 40 | ≈ 4·⟨n⟩ | | 5 | 25 | 80 | ≈ 3·⟨n⟩ (tighter for large |α|) | Preflight budget estimates above are conservative (all at nmax = 30); expect S2 to run in ~10 min with α-dependent nmax. ## 7. Q5 implementation decision — driver over extension **Decision.** Write a thin Python driver at `wp-phase-contrast-maps/numerics/run_slices.py` that: - Imports `stroboscopic_sweep` as a library. - Iterates `run_single()` over the outer axis (|α| or φ_α). - Applies α-dependent `nmax` per §6. - Writes HDF5 output with the S1/S2/S3 conventions of v0.3 §8 Q2. **No modification** to `stroboscopic_sweep.py`. Rationale: - The engine is validated (exact-basis, Float64, convergence-diagnosed). Modifying it introduces regression risk unconnected to WP-E's goals. - A driver keeps the engine a reusable core across WPs. - Per-pulse Bloch capture, if needed later, is a small patch localised to `run_single()` and can be added behind a flag without disturbing the default code path. **R2 (instantaneous pulse) implementation caveat.** The script currently computes `dt = π/(2·N·Ω_eff)` internally (line 360), so δt is not a free parameter. Implementing R2 requires either: - A `mode="impulsive"` flag in `run_single()` that replaces the finite-time `U = expm(-iH dt)` with an exact kick operator `K = exp(-i·θ_pulse·(C σ₋ + C† σ₊)/2)` where θ_pulse = Ω_eff·dt is computed from the nominal values, or - A driver-level reimplementation of the pulse loop with the same ingredients. Decision deferred to main sweep (not gating). First-pass R2 will be the driver-level reimplementation; revisit if the driver path proves fragile. ## 8. Outstanding actions - [ ] Create `wp-phase-contrast-maps/numerics/run_slices.py` driver. - [ ] First slice run S1 = (δ₀, |α|) at φ_α = 0, 121 × 5 points, with α-dependent nmax. - [ ] R1 reference run at η = 0.04 (same grid as S1). - [ ] Decide on R2 driver-level implementation. - [ ] Flag for dossier maintainers: the HDF5 `contrast_z = 0.293` disagrees with the dossier's quoted HDF5 adaptive-learner value of 0.61 at α = 0. Either the dossier refers to a different HDF5 file or is stale; needs reconciliation in the next dossier pass. *Not* in WP-E's scope per v0.3 non-goals. ## 9. Gate outcome Preflight passes. Main sweep cleared. Next logbook entry: first data run (S1 driver output) — estimated wall time < 1 min once driver is in place. *Guardian cadence (v0.3 self-review point): no changes to README.md in this entry. v0.4 awaits main-sweep data, not preflight alone.*