# Context overview ```mermaid flowchart LR BKT[("public S3 bucket dpr-common:
S2 L1A · ESA PSD L0 SAFE · GIPP")] L1["S2B L1B
(EOPF Zarr, radiance / counts)
— reverse chain input"] GIPP["operational S2A GIPP
R2EQOG / R2DEPI / BLINDP /
R2PARA / R2CRCO (XML)"] ADF["ADF
SRF (spectral characterisation)"] E2ES["s2_msi_raw_generator reverse chain
(invert L0→L1B corrections;
CCSDS-122 + packetize)"] L0["Synthetic L0 RAW EOProduct
(Zarr, compressed-ISP streams + STAC,
PSFD names — ICD-IF-NAME)"] CAL["calibration sub-set:
synth diffuser + dark →
derived ADF (inverse-crime)"] MSI["msi-processor l0_decode
→ L1A′"] REP["validation & report:
Synthetic-Synthetic L0 vs reference ESA L0 'img'
(10/20 m ≤~4 DN) · L0plus codec
round-trip decode==L1A bit-exact"] BKT -->|"s3fetch"| L1 BKT -.->|"ESA L0 (structural ref)"| REP L1 --> E2ES GIPP --> E2ES ADF --> E2ES E2ES --> Synthetic L0 E2ES -.-> CAL L0 --> MSI --> REP L1 -.-> REP ``` **Inputs.** The primary input is a **S2B L1B** (radiance / counts) EOPF Zarr granule; the operational S2A **GIPP** (per-pixel dark + relative response/PRNU, defects, offsets, crosstalk, on-board-eq); and the **SRF** for spectral characterisation. L1A is not an input — it is an intermediate the reverse chain *produces* on the way down to Synthetic L0. **Processing.** The reverse chain runs the S2 L1B backwards through the **exact inverse** of the operational L0→L1B radiometric correction chain — invert offset, relative-response/PRNU, dark, un-bin, SWIR re-stage, defective, crosstalk, on-board-eq — to reconstruct **L1A → L0plus → Synthetic L0**. MTF-deconvolution is OFF, so PSF and noise are **not** re-applied. A separate **calibration sub-set** synthesises sun-diffuser + dark acquisitions and *derives* the calibration coefficients back — the coefficients a downstream processor would actually use (inverse-crime cure). **Output.** The reconstructed **Synthetic L0 RAW** EOProduct (the ICD-IF-L0 Zarr: 156 detector/band frames, quality masks, optional CCSDS ISP telemetry, STAC + sensor-configuration metadata); the reverse chain also emits the **L1A** and **L0plus** (CCSDS-122 ISP) intermediates en route. **Verification context.** The reconstructed **L0** is compared against the **reference ESA L0 'img'**: the 10/20 m bands agree to **≤~4 DN**. As a supporting check, the **L0plus codec round-trip** is bit-exact — `decode(L0plus) == L1A`. ## Calibration database (ADF output) Besides the Synthetic L0 RAW product, the generator also *derives* the radiometric calibration coefficients and writes them as a versioned set of EOPF **Auxiliary Data Files** — the **calibration database** — that the downstream processor (the L1PP blocks of `msi-processor`) consumes directly. This is the single shared sensor-model ADF of the E2ES ⇄ processor coupling: the generator produces the ADF; the processor keeps calibration internal. Coefficients are **derived** (synthetic diffuser + dark), not the truth ADF, so the round-trip is non-tautological. ```mermaid flowchart LR subgraph SRC["derived coefficients (per band)"] PR["NUC gain g_d"] OF["NUC offset o_d"] DK["dark k"] AA["abs gain (~1/A)"] ES["ESUN (solar irradiance)"] end subgraph DB["cal-DB — EOPF ADFs (zarr v2)"] NUC["nuc.zarr
/gain, /offset"] DARK["dark.zarr
/dark_offset"] RADO["radiometric.zarr
/gain, /offset"] SPECT["spectral.zarr
/esun"] end PR --> NUC OF --> NUC DK --> DARK AA --> RADO ES --> SPECT ``` The NUC `gain`/`offset` follow the processor's two-point convention (`estimate_nuc`); the absolute `radiometric.gain` is diffuser-derived ($\approx 1/\mathrm{cal\_gain}$); `spectral.zarr` carries the per-band **ESUN** (Thuillier 2003, S2A — ATBD §A.3) the processor's `toa` unit needs for TOA reflectance. Written by `s2_msi_raw_generator.adf_writer` (`s2_msi_raw_generator.caldb`, pipeline phase `build-caldb`).