Pass 3 calibration runs update (for records)

[divine7022]

Report status and follow-up — 20 October 2026

Historical Pass 3 diagnostic. Parameter-bound behavior and the proposed input/pool diagnosis are retained as evidence, not a final parameterization.

Implementation: parameter/target and covariance specification #245, multisite fit #247, and documentation #249. Later calibration artifacts and kCN/litter tests #275 record subsequent work. The specific request below to re-score saved Pass 3 outputs with shared-year covariance has not been verified; link its result or explicitly supersede it with a documented covariance analysis.

---

Quick calibration status update: Pass 3 runs is complete and the audit passed. The additional turn_over_time parameter did not materially improve the fit -- RMSE only changed from 6.67 to 6.63, while the residual for the compost treatments remained around +8 Mg C/ha.

More importantly, the parameter behavior suggests we're now pushing the parameters to compensate for a structural/input mismatch rather than estimating them cleanly. turn_over_time moved close to its upper bound (1.2), and fracLitterRespired remained close to 1.0. som_respiration_rate also moved well outside its original bety prior. So I don't think these should be treated as final calibrated parameter estimates.

The current question is whether the remaining mismatch is coming from (1) the target definition -- the model target includes both soil and litter carbon, while the Salinas observations appear to represent 0–30 cm soil carbon or (2) the compost carbon input, where the Salinas record implies ~35% C but doesn't document that carbon fraction.

Rather than doing another large calibration run, I'm going to do two diagnostic forward runs using the Pass 3 posterior: one control and one compost treatment. I'll retain the litter/soil carbon outputs and decompose the modeled compost effect between the soil and litter pools. That should tell us which of those two issues is driving the residual.

So the calibration itself is paused for now. The useful outcome from Pass 1-3 runs is that we've ruled out simply needing another loss-rate parameter; So it seems like next step is to resolving the input/target definition before producing a final parameter set

[divine7022]

One additional diagnostic came up from the Pass 3 results though: a large fraction of the residual variation appears to be shared by campaign year across the eight Salinas systems. Our current Sigma treats those measurements as independent, which may be overstating the amount of independent information in the target and could be contributing to the parameter boundary behavior. Before doing another calibration or changing the compost/input specification, we're going to re-score the existing Pass 3 output with a properly derived shared-year covariance term. This requires no new model runs and will tell us how much of the apparent residual is actually observation/campaign variability. The compost C fraction and soil-vs-litter target definition remain unresolved separately