# v0.19: saved computation helps, but does not repair the planner

Charging only executed model transitions raised complete-season both-bed viability from **71.60% to 76.25%**, a **+4.65 percentage-point** improvement over the old logical allowance. Budget-limited seasons fell from **10 of 15 to 1 of 15**. Ensemble care still achieved **82.86%**: the candidate remains **6.61 points behind** and fails the frozen development gate. It is not promoted and no fresh cases were consumed.

This is a controlled accounting change within three exposed seeds, not evidence of fresh generalization or a new learning capability. The exact reuse optimization remains useful; extra available planning is insufficient by itself.

## Complete-season comparison

Each version ran the same fifteen 480-tick habitats: three seeds crossed with steady, pipe, dry, supply and compound conditions. The legacy and executed arms use the same shared two-decision planner. Only the count read by its allowance changed. All arms retain the 12-million per-season ceiling, two-million search reservation and 100,000 update reservation. The comparator is the unchanged live ensemble controller.

| Version | Both beds viable | Budget-limited seasons | Executed model transitions | Searches |
|---|---:|---:|---:|---:|
| Shared planner, logical allowance | 71.60% | 10/15 | 99,953,699 | 1,476 |
| Shared planner, executed allowance | 76.25% | 1/15 | 112,630,672 | 1,701 |
| Ensemble care | 82.86% | 0/15 | 36,189,496 | 2,175 |

The accounting change allowed 225 additional searches and spent 12,676,973 more actual transitions than the shared legacy control. It improved eight matched habitats and left seven viability scores unchanged; none declined on this panel. The candidate still used about 3.11 times ensemble's model transitions. Search counts alone do not compare computational depth: each two-decision search is substantially larger than an ensemble search.

All fifteen legacy rows and all fifteen ensemble rows reproduced the v0.16 care metrics and logical charges exactly. Every executed-allowance action, physical frame and cumulative work ledger matched its legacy counterpart before the first legacy-limited action. The five never-limited legacy seasons matched throughout. This isolates the benefit of changing the charged count; it does not repair earlier forecast or decision errors.

| Exposed seed | Legacy allowance | Executed allowance | Ensemble |
|---|---:|---:|---:|
| 93101 | 45.17% | 45.17% | 50.00% |
| 93103 | 88.21% | 94.79% | 99.92% |
| 93113 | 81.42% | 88.79% | 98.67% |

Every seed mean remains below ensemble. Mean performance by condition also varies: executed allowance reaches 83.89% in steady conditions, 78.75% with a pipe fault, 72.99% with drying, 81.25% with supply faults and 64.38% with compound faults. The corresponding ensemble means are 81.46%, 86.04%, 81.88%, 80.00% and 84.93%. The largest remaining condition-level gap is compound care. These correlated development habitats do not support an independent fifteen-sample confidence claim.

## What the retained replay shows

The two retained cases were fixed before outcomes: the first seed's steady and compound seasons. The steady shared-planner versions match for all 480 ticks, with 52.08% both-bed viability and no budget stop.

In the compound case, the legacy allowance stops planning at **tick 437**. Actual-work accounting avoids that stop entirely and actions first differ at tick 437, but complete-season viability remains **43.96%** in both versions. Before the stop (ticks 1–436), both achieve 48.39%, versus ensemble's 60.78%. Both shared versions and ensemble have zero joint viability over ticks 437–480. The extra actions arrive after substantial earlier losses and do not restore viable care in the remaining window.

The only executed-allowance stop is seed 93113's compound case, at tick 440 instead of legacy tick 301. Care improves from 64.79% to 75.21%, while ensemble achieves 100%. Its summary and common-window results are in the complete report; it was not selected for a visual trajectory after outcomes.

Open **What saved work buys** beside the nursery controls, or `/allowance.html`. The page exposes all 45 season rows, seed/condition means, synchronized recorded habitats, actual/logical work curves, budget-stop and first-difference jumps, and exact comparison exports. Playback performs no simulation. The nursery remains on ensemble care.

## Accounting and verification

Development charged **22,626 physical actions**, including 1,026 shared calibration actions; **248,777,773 executed model transitions**; **340,117,470 logical requests**; and **217.89 active seconds**. Reuse supplied 91,339,697 requested transitions across both shared arms. No uncertain actions or outstanding reservations remain. The declared ceilings were 22,626 physical actions, 600 million actual transitions and 600 active seconds.

Search accounting includes every selected and rejected evaluation. Outside-search model updates were separately counted and checked against the shared ledger: 104,538 transitions for legacy, 104,756 for executed and 105,325 for ensemble. Root counts distinguish selected and rejected roots; model-transition totals do not attribute costs to individual tree branches. All shared projection requests stayed within twelve ticks, no fallback occurred, and the memo retained at most two scalar answers per tail. Allocation, regression arithmetic, copying and memo bookkeeping are included in active time but are not themselves model transitions. This run does not establish a new speed or energy claim.

The new wrapper rejects invalid work ledgers and keeps logical counters intact, subtracting reuse once when checking the allowance. No planner source, score, fitted model, parameter, horizon or live engine changed. All 63 frozen scientific sources and previous evidence are preserved. All 138 tests and both HTTP integration suites pass. Independent verification exactly reproduced four retained executed/ensemble seasons and their ledgers, charging another 2,262 physical actions, 17,339,956 executed transitions, 22,912,029 logical requests and 17.58 active seconds. Software, independent replay, HTTP and browser verification are recorded separately in the delivery verification (full source archive).

## Decision and next bounded question

Retire actual-work accounting as a sufficient repair for this care policy; keep exact reuse available for subsequent offline planners. Do not consume fresh evaluation seeds or promote the candidate after a failed gate.

The next question is whether replanning rejects a useful follow-up because its ordinary search compares a different future continuation. Freeze an audit at the already exposed pre-budget forks: score the promised next operation and current ordinary alternatives from the same current fitted beliefs, using a common continuation rule and matched execution. Include the v0.18 cases where commitment harmed successful care. Keep actual local readings, and distinguish a changed model estimate from a changed way of scoring future actions.

Only after that diagnostic identifies a consistent improvement should we freeze a bounded rule for retaining or replacing a planned follow-up and run another full-season comparison. Blanket script commitment is already contradicted by the controls. Increasing duration, machines or the computation ceiling is not the next supported intervention.

Method: [frozen v0.19 protocol](v19-method.md). Evidence: complete development record (full source archive), audit (full source archive). This remains Stage 1 of the [roadmap](../ROADMAP.md).