Stored SSB fits carry a schema, and only the current one is read - 2026-10-11

Historical: describes the code as of 2026-10-11; the mean_phase estimator is retired, and later schema versions replaced the one described. See the current fit methods.

Question

A screening folder keeps each acquisition’s fitted SSB values in live/screen/<dataset>/config.json under computed.ssb. Values stored before the package restructure reconstruct a different image with the current engine (two real scans: inverted contrast on one, a lattice peak six times weaker on the other, while a fresh fit reproduces the earlier pictures). Which stored fields changed meaning, and how does a reader tell a current fit from an earlier one?

Setup

One synthetic scan, simulated in the probe script itself: a periodic weak phase object with positive phase at the atom columns, 300 kV, 30 mrad, defocus -10 nm, 64 x 64 positions at 0.5 A, 128 x 128 detector at 0.6152 mrad per pixel. The same array went through the engine before the restructure and through this tree, each with its own fresh CuPy kernel cache, phase_estimator="mean_phase", 200 trials, seed 42, no 180-degree check. Scripts: scratch probes/engine_compare.py (one tree), probes/engine_diff.py, probes/engine_cross.py.

Numbers

Fixed values C10 -10 nm, C12 2 nm, phi12 0.4 rad:

session	earlier engine loss	this tree loss	phases
det_sampling=0.6152, rotation 0, CoM as measured	0.000520867	0.000520867	identical, bit for bit
det_sampling=0.6152, rotation 0, CoM reversed	0.000984409	0.000984409	identical
det_sampling=0.6152, rotation 25, CoM as measured	0.000892663	0.000892663	identical
det_sampling=0.6152, rotation 25, CoM reversed	0.000982096	0.000982096	identical
automatic detector sampling, rotation 0	0.000331408	0.000521682	correlation 0.17
this tree at det_sampling = 2 x 30 / 49		0.000331408	identical to the earlier automatic output

Fit with automatic detector sampling (truth C10 -10 nm):

engine	bright-field radius used	detector sampling	fitted C10	C12	phi12	loss
earlier	49 px (integer half maximum)	2 x 30 / 49 = 1.224 mrad per pixel	-4.24 nm	0.33 nm	-1.21 rad	0.000268
this tree	48.69 px (disk edge)	30 / 48.69 = 0.616 mrad per pixel	-9.90 nm	0.01 nm	-0.14 rad	0.000149

The earlier engine’s fitted values, reconstructed by this tree with automatic sampling: loss 0.000974, correlation 0.13 with the picture the earlier engine made from the same values. This tree’s own fit: correlation 0.86 with it.

What changed, by stored field

stored field	earlier engine	this tree	carries over
rotation_angle_deg, com_reversed	angle below 180 plus the flag, physical angle = angle + 180 x flag	the same	the convention does; a com_reversed chosen together with a defocus sign by an earlier fit does not
aberrations C10, C12 (nm), phi12 (rad)	chi sign, units and angle reference as now; fitted with detector sampling 2 x semiangle / half-maximum radius	fitted with semiangle / disk-edge radius	no: C10 came out at 0.4 to 0.5 of the current value
aberrations in a block without aberration_unit	Angstrom stored under an nm name (before 2026-09-24)	nm	no
rotation_angle_deg before 2026-09-27	physical angle up to 360, no flag	below 180 plus flag	the angle does, read as physical
loss	phase variance with zero aperture weight beyond half the disk; a session that had not fitted reported the batched evaluator	exact reconstruction loss	no, another quantity
bf_radius	bright-field selection radius in pixels	the same	yes
voltage_kV, semiangle_mrad, scan_sampling_A	acquisition settings	the same	yes

So the engine’s conventions did not change; the automatic detector calibration did (record), and every stored fit was made without an explicit det_sampling.

Decision

persistence.SCHEMA is 6 and persistence.FIT_SCHEMA = "quantem.ssb-fit/6" is derived from it. write_fit_metadata writes "schema": FIT_SCHEMA into computed.ssb; series.stored_fit(config) returns a block only when it carries that string, and saved_fit and acquisition_settings read through it. A saved fit record counts only at schema 6. Any other stored fit is absent: nothing in it is read, a probe reference is fitted again, and write_fit_metadata replaces such a block whole. The key and the value form follow the .qem scientific-metadata schema string ("schema": "quantem.scientific-metadata/2"), whose readers also refuse another version instead of reinterpreting it.

Schema 5 could not serve unchanged: fit records of schema 5 exist that were written with the earlier sampling ("det_sampling": 1.0909 = 2 x 30 / 55 in their signature), so 5 does not prove the current convention.

Rejected

• Converting earlier values (C10 x 2, or dividing unmarked Angstrom by 10): the ratio is not a constant, it depends on the disk radius and on which twin of the 180-degree ambiguity the earlier fit took.

• Accepting a block that declares aberration_unit: "nm": 76 of 687 surveyed blocks declare it and every one of them was fitted with the earlier sampling.

• A separate marker number beside SCHEMA: two numbers for one contract can disagree.

• Reading the acquisition settings (voltage, semiangle, scan step) of an earlier block while ignoring its fit: one block would then be half trusted, and the settings are also in dataset.yaml or the call.