One workflow for four data-availability cases

The versioned gateway tools/dztr_run_v8.py adds two stages to the existing manager: no_observations generates and runs fixed profiles; compare_routes joins them with frozen joint, release-only and storage-only results. Older configuration versions continue to use their original runners.

& .\tools\Set-ProjectCache.ps1
python -B tools/dztr_run_v8.py plan --config configs/run_manager_availability_paper_v1.json
python -B tools/dztr_run_v8.py run --config configs/run_manager_availability_paper_v1.json --output outputs/my_four_route_run
python -B tools/dztr_run_v8.py run --config configs/run_manager_availability_paper_v1.json --output outputs/my_four_route_run --resume

plan performs no simulation or optimization. The paper run requests 32 monthly reservoirs and retains all 37 paper inventory rows, including Bull Lake's separate interval route and the four source-gated cases. This delivery runs fixed profiles only; it does not resume existing optimizer jobs.

Parameter generation without storage or release

For ungauged_mean_flow, storage targets are fixed to 20%, 30% and 97% of the adopted capacity. The explicitly configured active-storage adaptation uses dead_storage + fraction * (capacity - dead_storage) instead; the paper configuration preserves that existing assumption for Dickson, E. B. Campbell and Yellowtail. Neither capacity nor dead storage is adjusted to make targets valid. An invalid target set remains gated.

Release targets use the existing shared mean-flow prior: the target's verified GRAND DIS_AVG_LS divided by 1,000 gives m³/s. Where explicitly permitted, a volume-weighted mean effective inflow over intervals wholly within calibration is the fallback. The mean is multiplied by median external-donor P5/P30/P97 to mean ratios, excluding the target. These three annual targets repeat across all twelve months. Channel limiting and donor provenance remain explicit.

Initial storage is the assumed normal target; initial release is the first effective inflow clipped to zero and channel capacity. The estimator receives an inflow-only frame. Storage and release observations are kept separately and read for assessment after parameters have been frozen. The original mixed-year masks and at least 65% calibration support are retained. State remains continuous across years. No optimizer is available for this no-response recipe.

Inflow, capacity and channel metadata may themselves depend on historical observations. External donor simulations represent one region and are not independently certified. This is an exploratory conditional benchmark, not proof of global ungauged or forecast skill.

Automatic recommendation and explicit method choice

The recommendation depends on usable calibration observations and source gates:

Usable data	Recommended method	Fitted target responses
Storage and release	observed_monthly	Both
Prepared level-derived storage and release	level_geometry_monthly	Both, with the geometry provenance
Release only	flow_only	Release
Storage only and a usable release prior	storage_only	Storage
Neither response and a usable release prior	ungauged_mean_flow	None

Every numerical route also needs inflow and valid physical inputs. A source gate remains a gate. The recommendation selects an already prepared profile; an absent recommended profile is reported explicitly. It does not switch to another route because assessment scores look better, and it does not start optimization implicitly. Level-derived profiles still need their existing reference, timing and conversion checks before being usable.

The comparison's route_registry.json links all available comparison profiles to their original sealed registries. The default application selects the recommended method. An explicit weaker-data comparison uses method:

{
  "id": "application",
  "type": "simulate",
  "config": {
    "registry": "@comparison.registry",
    "forcing_csv": "inputs/new_prepared_inflow.csv",
    "reservoir_id": "my_reservoir",
    "method": "ungauged_mean_flow"
  }
}

Omit method for the automatic recommendation. A verified grand_id can replace reservoir_id. A different forcing start requires an initial_state JSON with storage_m3 and release_m3s; the legacy clock restrictions remain active. The simulation gateway expects prepared SI forcing, preserving the declared source transformation and physical inflow correction. Responses are not read for application. Missing GRAND IDs stay blank and remain selectable by exact reservoir ID.

Comparison and interpretation

The report recalculates metrics from saved trajectories on the same observations, native calendars and original masks. It rejects changed response data, year selections, profile identities and physical-balance checks. Missing cached support is reported without rerunning old models. Per-case phase scores, paired differences, seasonal biases, plots, parameters and failures are saved.

Available-case medians retain each route's actual membership; a separate common-successful-case summary uses identical membership across all four routes. Profiles and simulation success are separate fields. A valid assumed profile can still fail at a timestep; that failure is not converted into a successful run or removed from the inventory.

The existing joint calibration used a 512-evaluation NSGA-II search; the two single-response routes use the 64-evaluation DDS checkpoint of a frozen 512-evaluation horizon. The no-response route uses zero optimizer evaluations. Different data access, initial states and target dimensions prevent attributing performance differences to the optimizer alone. Convergence remains unproven.

The parameter table has one row per reservoir and method, with profile source paths for full provenance. The route inventory has one row per reservoir and the recommended method. Existing registries and earlier scientific outputs remain unchanged. Completed jobs and known failures have persistent receipts; --resume verifies and reuses them. All project caches and temporary files remain on D:.