What should be measured (so problems are caught early)

• Noise floor: RMS plus low-frequency behavior (correlation / slow shading tendency), not only peak ENOB.
• Drift: offset vs temperature/time and gain vs temperature/time, tracked per gain state.
• Banding sensitivity: does striping change with timing edges, reset phase, or reference conditions (a strong hint of correlation)?
• Lag: frames-to-recover after saturation; check dependence on temperature and gain state.

(1) Charge-to-voltage (integrator)
    Vout ≈ Q / Cf

(2) Dominant error contributors (concept form)
    σV^2 ≈ (kT/Cf) + (en^2 · BW_eq) + ( (I_leak · Tint)/Cf )^2 + (V_inj/Cf)^2 + (1/f contribution)

(3) Drift mapping (separate additive vs multiplicative)
    Vmeas = G(T,t,gain_state) · Videal + Voffset(T,t,gain_state) + εrandom
        

How to interpret the terms (what they usually become in images)

• kT/C on Cf: sets a reset-related floor. Timing (CDS) can reduce visibility, but aggressive edges and injection can reintroduce structured residue.
• en over BW_eq: drives random grain. If this term dominates, better front-end noise and bandwidth control help more than complex drift models.
• I_leak · Tint / Cf: is a practical “slow error generator.” Even tiny leakage can become measurable baseline drift across the integration window, often showing as shading or temperature-dependent offsets.
• V_inj / Cf (switch injection & feedthrough): is often correlated with timing edges, so it tends to create banding rather than benign random noise.
• Separate G(T,t,·) from Voffset(T,t,·): gain drift and offset drift require different calibration evidence and different acceptance checks (flat-field vs dark).

Practical decision rules (what to fix first)

• If banding dominates: treat it as correlation. Focus on reset/injection control, CDS phase margins, and reference/bias coupling before chasing “more bits.”
• If drift dominates: separate offset vs gain. Add temperature sensing where drift is created, and build calibration tables indexed by gain state and temperature bins.
• If random grain dominates: focus on front-end noise and bandwidth, and confirm the ADC choice supports the intended sampling and filtering without passband ripple artifacts.
• If lag dominates: validate frames-to-recover after saturation and adjust reset strategy, node leakage, and “memory” contributors until recovery is within spec.

Core relations (concept form)

1) Charge-to-voltage:
   Vout ≈ Q / Cf

2) Droop from leakage during integration:
   ΔVdroop ≈ (Ileak · Tint) / Cf

3) Sensitivity to injection/feedthrough:
   ΔVinjection ≈ Qinj / Cf   (or a timing-locked step on the summing node)
        

• Dynamic range vs sensitivity: smaller Cf increases conversion gain (Q→V), improving low-dose sensitivity but also amplifying injection and residual steps.
• Reset-related floor: Cf directly impacts kT/C behavior and how much reset residue can survive into sampling windows.
• Correlation risk: if a repeatable charge step is injected every row/phase, smaller Cf makes that step more visible as banding.

• Droop scales with Tint: if shading worsens roughly in proportion to integration time, leakage (or bias-current paths) is a prime suspect.
• Uniform vs spatially varying leakage: uniform droop can look like a baseline shift; spatial leakage gradients create non-uniform residuals that calibration must track.
• Guarding is a stability tool: guard rings and controlled return paths reduce the chance that leakage becomes unpredictable or temperature-sensitive.

• Fast edges can couple into the summing node: parasitic feedthrough and switch injection create a repeatable step.
• Banding happens when the step is sampled: if Sample A or Sample B is too close to reset, settling residue becomes a fixed-pattern error.
• Practical knob: enlarge settle margin after reset and validate with a phase sweep (move Sample A later and see if banding falls).

• Memory shows up after large steps: strong exposure or saturation can leave a slow tail that contaminates the next frames.
• Not an ADC problem: if frames-to-recover depends on temperature or reset strategy, the integrator node is a likely root cause.
• Validation: step-to-recovery curves (frames-to-baseline) across temperature bins and gain states.

• Change Tint: does shading scale roughly with Tint (leakage signature)?
• Sweep Sample A phase: does banding change with phase (injection/settling signature)?
• Modify reset edge: does banding respond to edge shaping (feedthrough signature)?
• Run step-to-recovery: does lag depend on temperature or gain state (memory signature)?

• Guarantee a settle margin after reset before Sample A.
• Check banding sensitivity vs Sample A delay (phase sweep).
• Repeat the sweep per gain state and temperature bin.
• Confirm that the chosen A–B interval does not weaken drift cancellation.

• Define an overlap region: adjacent gain states must share a usable input window for continuity checks and stitching.
• Keep “forbidden zones” away from switching: avoid boundaries near saturation and near the noise floor where any mismatch becomes visible.
• Index calibration by gain state: offset and gain corrections should be stored and verified per range (not one global map).

• Hysteresis: use separate up-switch and down-switch thresholds so the system does not bounce at the boundary.
• Hold-off: after switching, freeze the gain decision for a minimum number of rows/frames to guarantee stable sampling.
• Trigger on robust statistics: base switching on a windowed metric (peak count, mean, saturation flags), not a single sample.

• Define frames-to-recover: after saturation or a large step, measure how many frames are required to return to the baseline envelope.
• Protect the switching window: during recovery, suppress gain switching or force a conservative state until settling is confirmed.
• Watch for memory tails: a slow tail that depends on temperature or gain state is a strong indicator of node memory, not “random noise.”

• Sampling must occur after settling: if the driver, ADC sample-and-hold, or reference has not settled, the residue is sampled and can become banding.
• Phase sweep is the fastest diagnosis: move the sample instant later and check whether seam/banding changes (a signature of settling residue).
• Repeat per gain state: the hardest state often differs by load, swing, and reference dynamics.

• In the overlap region, verify Δ(Out) between adjacent gains stays within limits across temperature bins.
• Verify no boundary thrash: hysteresis and hold-off prevent repeated switching on similar scenes.
• After switching, verify settling margin at the chosen sample instant (phase sweep sensitivity low).
• After saturation, verify frames-to-recover and block switching during recovery if needed.

• Why it can fit: digital filtering can suppress wideband noise and support strong low-frequency behavior and consistent linearity.
• What must be managed: group delay (latency) and passband ripple risks. Poor filter choices can introduce structured slow texture or response quirks tied to system cadence.
• Validation focus: verify the filter response under the system’s timing cadence and confirm low-frequency residuals do not become shading patterns.

• Why it can fit: low latency and predictable sampling behavior, useful for tight timing and high per-channel throughput.
• Primary risk: strict requirements on driver and reference settling. If sampling occurs before settling, the residue is captured and can become banding, especially under switching.
• Validation focus: phase-delay sensitivity, step response at the input/driver, and reference transient checks across gain states and temperature bins.

• Need very low latency? SAR is often favored if settling can be proven.
• Low-frequency stability is the top KPI and latency is acceptable? ΣΔ is often favored with disciplined filter validation.
• Gain switching is frequent? prefer the option whose switching-settling verification is stronger and easier to guarantee.
• Complexity budget: ΣΔ shifts complexity to filtering/verification; SAR shifts complexity to driver/reference/settling control.

• Gain drift signature: the error scales with signal level. Mid-gray flats shift proportionally and appear as contrast or global shading changes.
• Offset drift signature: low-signal and dark regions shift more obviously. Baseline moves and fixed-pattern residue becomes visible near the floor.
• Practical split test: compare drift behavior on a dark frame vs a mid-gray flat. Proportional change points to Vref/gain; additive shift points to offsets/bias.

• Source layer: Vref (and its buffer output), bias rails, analog rails (AVDD), and any rail that can modulate references.
• Chain layer: PGA output, ADC driver node, ADC reference pins/decoupling neighborhood, and any switched node that can inject residue.
• Result layer: gain/offset estimators, flat-field residual metrics, shading/banding indicators aligned with temperature logs.

• Reference TC matters twice: it changes gain directly and it can also shift bias points that influence offsets.
• Rails can masquerade as “drift”: a rail moving with load or temperature can modulate Vref/bias and create slow structured changes.
• Goal: make drift stable, monitorable, and modeled. Uncontrolled coupling leads to unpredictable residuals that look like shading.

1. Run dark + mid-gray flat: decide additive vs multiplicative behavior.
2. Align temperature vs time: check monotonic drift and warm-up shape.
3. Correlate Vref and key rails with the measured gain/offset estimators.
4. Check gain states: see whether a specific range magnifies the issue.
5. Confirm with controlled perturbations (phase/edge/rail): a real root cause responds predictably.

• Near reference and bias networks: track what changes gain and offsets directly.
• Near AFE/ADC hot spots: capture local self-heating and gradients that a corner sensor would miss.
• Multiple points: use at least a small set of sensors so gradients can be modeled, not guessed.

• Per-zone: different regions see different gradients; apply zone-indexed corrections to avoid over/under-compensation.
• Per gain state: gain ranges have different sensitivities and loading; store tables per range to prevent seams.
• Prefer bin tables over a single curve: temperature bins make validation and rollback straightforward.

• Early drift is fastest: initial self-heating and reference stabilization can dominate the first minutes.
• Use guarded modes: during warm-up, increase residual monitoring and avoid aggressive auto-switching if it magnifies artifacts.
• Enter “stable mode”: apply the tightest compensation after the temperature slope falls below a defined threshold.

• Residual monitors: track flat-field residual, shading metric, and seam metric in the overlap region.
• Gated updates: if residuals exceed thresholds, roll back to a safe table or trigger re-calibration.
• Version lock: tables must be tagged by TableID, BinID, ZoneID, and GainState to support traceability and rollback.

• Temperature sensing covers reference + AFE/ADC hot spots (not only board average).
• Corrections are indexed by zone + gain state + temperature bin.
• Residual monitors are logged and thresholds are defined (seam + shading + warm-up slope).
• Calibration tables are versioned and rollback-ready (TableID/BinID/ZoneID/GainState).

• Dark map (offset): removes additive fixed pattern. It must not capture warm-up slope or short-lived drift.
• Flat map (gain): removes multiplicative non-uniformity after offset is removed. It must not include offset residue.
• Defect map: flags unstable pixels/rows/columns for replacement. It must not confuse random spikes with permanent defects.
• Linearity/LUT: reduces structured nonlinearity. It must not “fit” noise into a curve.

1. Acquire dark set → compute dark map: remove the additive layer first so gain is computed on the right baseline.
2. Acquire flat set → compute flat map: compute multiplicative correction on offset-corrected frames.
3. Build defect map: identify stable defects using statistics after dark/flat corrections reduce confusion.
4. Apply linearity model (where defined): keep linearization disciplined and validate by residuals, not by perfect fitting.

• Too few frames: single-frame or small-sample maps bake random noise into fixed correction.
• Warm-up not finished: early drift becomes a “map feature,” creating slow shading later.
• No outlier handling: transient spikes become defects or gain distortions and show up as banding.
• Mixed conditions: reusing one map across different gain states or temperature bins creates seams and discontinuities.

• Dark residual: check for row/column structure that indicates offset instability or drift baked into the map.
• Flat residual: confirm low-frequency residual decreases without creating new stripes or periodic patterns.
• Seam residual (overlap): verify adjacent gain states agree in overlap regions to prevent visible steps.
• Temperature consistency: confirm residuals stay bounded within each temperature bin; large bin-to-bin jumps require re-binning or model fixes.

• Freeze: tag tables with TableID, temperature bin, GainState, acquisition conditions, and timestamp.
• Update gate: allow updates only when residual thresholds are exceeded consistently or after service events.
• Rollback: keep last-known-good tables for immediate rollback if new tables increase structured residuals.

• Integrator node memory: saturation drives nodes into regions where recovery is slow or nonlinear.
• Charge injection / trapping: switch edges leave residual charge that decays over multiple frames.
• Dielectric absorption: capacitive elements can release stored charge slowly, creating a long tail.

1. Apply a controlled step: dark → bright (saturate) → return to dark.
2. Record offset residual per frame: measure how far baseline stays from the dark target.
3. Compute frames-to-recover: number of frames required to return under a defined residual threshold.
4. Repeat across conditions: temperature bins and gain states reveal worst-case recovery behavior.

• Residual threshold: set a limit tied to the dark noise envelope so “invisible” tails remain below it.
• Time/frames limit: define the maximum allowed frames-to-recover under worst-case conditions.
• Fail action: if gates fail, enforce recovery policy (blanking/flush) and block risky switching during recovery.

• Blanking / discard window: drop a defined number of frames/rows after saturation so tails do not enter images.
• Stability hold: freeze gain switching and critical timing changes until recovery gates are satisfied.
• Reset discipline: enforce a controlled reset/settle routine to minimize injection-driven residue.

• frames-to-recover is measured and logged for each gain state and temperature bin.
• pass/fail gates are defined (residual threshold + frame limit).
• recovery policy (blanking + hold) is applied automatically on gate failure.

• Rows = test items that commonly trigger late-stage rework.
• Columns = conditions (temperature bins, gain states, stimulus levels, time) that reveal worst-case behavior.
• Cells = output metrics that are computed, logged, and gated (PASS/FAIL) for release readiness.

• Freeze a calibration set only after PASS: TableID, TempBin set, GainState set, acquisition conditions, timestamp.
• Define fail actions: rollback to last-known-good tables, block risky switching modes, enforce recovery blanking.
• Re-test rules: after any table update or service event, re-run the matrix items that touch the modified map or mode.

These part numbers are practical reference points for building stable rails, references, switching, sensing, and logging that help verification results remain reproducible across temperature bins and gain states.