# v0.13: observed weather and complete care

## Development evidence and choice

The exposed calibration audit used the twelve v0.11/v0.12 base seeds. All 24 delays were correct. Maximum relative gain error was approximately 0.000025%; maximum relative evaporation error was approximately 0.0000065%. The short adaptive calibration history therefore did not justify a calibration rewrite on these cases. That audit charged 4,104 physical actions and 15,624 model transitions.

The common-action audit collected 480 ticks of ordinary ensemble care for each of those seeds under all five conditions. At ticks 0, 12, …, 432 it forecast the same recorded next 48 actions and scored endpoints at 1, 6, 12, 24 and 48 ticks. Fitted forecasts receive only memory, current observed weather and those supplied actions. Offline references successively grant current state, current coefficients, recorded future weather and changing coefficients. Their differences are conditional, order-dependent diagnostics, not additive causal attributions or live foresight. The weighted moisture forecast is also distinct from the planner's weighted nonlinear loss.

The audit charged 32,904 physical actions and 154,321,320 model transitions. At 48 ticks, mean moisture error with exact current state and coefficients was 5.37 percentage points; supplying recorded future weather reduced it to 1.05 points. The full dynamic reference reproduced the physical outcomes exactly. These exposed results motivated one weather forecaster, rather than a calibration or planning-horizon change.

A fixed pilot on the first three exposed seeds, all five conditions, compared the same fitted-model mixture under constant versus temporal weather. Mean 48-tick moisture error changed from 6.97 to 6.01 points; 12-tick error changed from 1.76 to 1.64 points. It charged 8,226 physical actions, 36,353,242 physical-model transitions and 40,115 weather-recurrence transitions. No forecaster settings were tuned after this pilot. Software tests use exposed seed 90101 and synthetic signals, not the fresh panel below.

## Candidate

Only the weather forecast changes. Local observations still supply current heat and rain each tick. Calibration, protected normal knowledge, candidate families, validation and switching, ordinary action menu, loss, 48-tick planning horizon and three-model weighting remain unchanged.

The candidate retains at most 96 consecutive weather observations. At a new plan, with at least 24 observations, it fits a second-order linear recurrence independently for heat and rain. Values are centered and scaled. A ridge of 0.001 on lag coefficients is centered on persistence. Unstable or uninformative fits are rejected. The last eight observations are held back for an eight-step recursive check against persistence. A channel qualifies only if mean squared error is at least 10% lower and persistence error exceeds 1e-12. A qualifying recurrence is refitted on the available history; unstable refits also fall back to persistence.

The first forecast step uses the actual current weather. Subsequent values are bounded by the observed range extended by two observed standard deviations, with a zero lower bound. No period, simulator phase, future weather, change label or change schedule enters the candidate. It forecasts 96 ticks to cover the main horizon and tail look-ahead probes. The existing reserve calculation continues holding the final forecast weather constant beyond the simulated horizon. Fallback is separate for heat and rain.

Regression rows and weather recurrence steps are counted separately. The common model-transition ceiling counts physical-state predictions and weather recurrence steps together. A recurrence step is not assumed to cost the same CPU time as a two-bed physical forecast; active worker time is reported separately. Least-squares arithmetic is reported through fit-row counts rather than being mislabeled as simulator transitions.

## Frozen fresh comparison

Base seeds: 114101, 114103, 114107, 114111, 114113, 114127, 114131, 114137, 114143, 114149, 114157, 114161. These seeds have not been used for development. Scenario seed is base plus 12,100,001. Conditions: steady, pipe, dry, supply and compound. Two arms: weather and ensemble. Each completes 480 physical ticks, including all travel, observations and operations.

Each base seed shares the unchanged 342-action calibration. Thus 12 × (342 + 5 × 2 × 480) = **61,704 physical actions**, reserved in twelve chunks of 5,142 actions. Each season has a twelve-million model-transition ceiling, using the same two-million planning reservation and 100,000 update reservation. Limited seasons continue through 480 physical ticks and remain in the result. The Mac service has a fifteen-minute active-worker ceiling. Interrupted reservations are conservatively charged by the service.

Primary outcome: weather minus ensemble complete-season both-bed viability, averaged over five conditions within each base seed and then over twelve base seeds. Report a paired 95% t interval using 2.201 with eleven degrees of freedom. Promote only if mean gain is at least one percentage point and the lower interval is above zero. A partial or budget-ended study cannot authorize promotion. Lower forecast error alone does not meet this rule. No noninferiority claim or alternative primary outcome will be selected after results are observed.

The first base seed's steady and compound seasons are retained in full for synchronized visual replay and independent reproduction, regardless of outcome. Replays are illustrative, not extra independent cases. Prior scientific implementations and reports remain frozen. Development diagnostics, the pilot and verification work are reported separately from the fresh care comparison.