Sensitivity Report
fresh-salvage sensitivity-run writes a Markdown report that summarizes a
one-at-a-time rolling-horizon sensitivity analysis. For each configured
parameter, the driver changes only that parameter and preserves all other
base values. It automatically includes the base value when it is absent
from the parameter’s configured value list.
Running The Report
Update the external stands_path, yields_path, and bridge_path
locations in examples/sensitivity_tsa29.yaml before running it. Then run:
fresh-salvage sensitivity-run examples/sensitivity_tsa29.yaml --json --strict
The report is written to
outputs/sensitivity_tsa29/reports/tsa29-sensitivity-results.md. The
--strict flag returns a non-zero exit status if any ordinary sensitivity
scenario fails. Do not interpret a partial report as a complete comparison.
Report Contents
Analysis Summary records the number of configured parameters and ordinary
sensitivity scenarios, their success status, and wall-clock time. Effective
Baseline Settings is the resolved control configuration used for every
one-at-a-time comparison.
Scenarios Examined lists every parameter value, whether it is the
baseline, and any scenario failure code. Each Results: <parameter> section
then contains:
Outcome By Setting: green harvest, burned salvage, burned area, principal objective, and agent objective at every tested setting.Change From Baseline: the signed difference from the baseline scenario; a positive value means the metric increased.Response Summary: baseline, minimum, maximum, range, and observed direction for each recorded metric.
Subsidy Flip Points
The TSA29 example’s flip_point_search block additionally requests an
adaptive search for the subsidy at which burned salvage becomes positive. A
separate search is run for every configured value of each non-subsidy
parameter. The report’s Subsidy Flip Points table has one row per search:
Inactive loweris the highest observed subsidy with burned salvage less than or equal to the configured threshold.Active upperis the lowest observed subsidy with burned salvage greater than the threshold.Widthis the uncertainty interval between those two tested subsidies.Probescounts midpoint runs, excluding the two endpoint checks.Convergedisyeswhen the bracket meets the configured tolerance.
The table omits subsidy_rate_per_m3 because it is the value being searched,
not a fixed condition that can move its own flip point. See
Flip-Point Search for the binary-search algorithm and bracketing rules.
not_bracketed means the configured interval did not contain a transition.
For example, the fire-free burn_rate_multiplier: 0.0 control has no burned
wood to salvage, so its upper subsidy endpoint remains inactive. This is an
expected sensitivity finding, not a failed ordinary sensitivity scenario.
Operation-Cost Sensitivity
The TSA29 example includes burned_harvest_cost: [56.0, 67.2, 78.4]. These
values represent the calibrated prompt-salvage operating cost and 20% and 40%
increases, in $/m3. This is the marginal operating cost for harvesting burned
wood, so it directly changes the subsidy needed to activate salvage. Compare
the corresponding burned_harvest_cost rows in Subsidy Flip Points:
a higher active-upper bound means more subsidy is needed to make salvage
privately viable.
This is intentionally separate from green-harvest and transport costs. Add
green_harvest_cost, green_transport_cost_per_m3, or
burned_transport_cost_per_m3 to parameters when those distinct cost
mechanisms are the question; each will receive its own one-at-a-time outcome
and flip-point rows.
Evidence Files
The report is a readable summary. Use its evidence files for analysis or audit:
data/tsa29-sensitivity-observations.jsonlhas the ordinary scenario-level metrics and rolling-horizon manifest locations.data/tsa29-sensitivity-comparisons.csvhas the response summaries.data/tsa29-sensitivity-flip-points.csvhas the flip-point brackets, convergence flags, and flip-search manifest locations.manifests/tsa29-sensitivity-sensitivity-manifest.jsonrecords the effective configuration and paths to all report artifacts.
Every flip-point row can require two endpoint rolling-horizon runs plus up to the configured midpoint limit. The flip-point section therefore adds material runtime beyond the ordinary sensitivity scenarios.