# The Conservatory 0.8 — results from this Mac

All four bounded runs completed with zero uncertain actions. The 26-file scientific source was frozen before the reserved seeds were evaluated. The earlier v0.7, v0.6 and v0.5 scientific sources and previously recorded reports still match their hashes. Full evidence is indexed by the v0.8 audit (full source archive); the [method](v8-method.md) states the limits and research connections.

## Image steering: movement helped; recalibration cost more

| Controller | Both criteria met | Mean physical actions | Mean calibrations |
| --- | ---: | ---: | ---: |
| First calibration, fixed control |48/60|9.73|1.00|
| Recalibrate, fixed control |48/60|17.37|2.82|
| Move control, retain calibration |60/60|8.45|1.20|
| Recalibrate and move |60/60|13.42|2.35|

All 12 shifted cases failed with a fixed control and succeeded when the measured footprint could move. All other evaluated conditions succeeded under all four policies. Adding recalibration lowered mean target error, but did not improve the success rate and increased physical actions. This supports movement for this particular misalignment family; it does not establish general image-editing ability.

Every final background met the preservation thresholds in all 240 trials. Temporary probes and rejected edits reached maximum outside change 0.05886 before undo, above the 0.025 final tolerance; these excursions remain in the trace. Alignment search is computationally expensive: 247.46 million predicted pixel-channel values for movable only, versus 6.22 million for fixed calibration. Fewer physical edits do not mean lower total computation.

Local photo import, 128×128 center cropping, rectangular selection, exact reload replay, and JSON/PNG export are now available. Browser verification used a generated PNG fixture. These procedural results do not validate photographic quality, arbitrary target edits, or human perceptual preservation.

## Nursery: fewer investigations, no established care advantage

| Controller | Both beds viable | Tests per season | Test actions per season |
| --- | ---: | ---: | ---: |
| Decision-value investigations |86.88%|1.28|17.47|
| Same ensemble, no investigations |86.92%|0|0|
| Ordinary care |86.05%|0|0|
| Previous disagreement investigations |84.78%|4.47|66.73|

The primary paired contrast, decision-value minus ordinary care, is **+0.83 percentage points**, with 95% interval **−0.63 to +2.29** across 12 habitat means. That does not establish an advantage. Against the same ensemble with investigations disabled, the secondary contrast is **−0.04 points**, interval **−1.07 to +0.99**. There is no demonstrated benefit attributable to the new questions themselves.

The chooser took 77 investigations and 1,048 investigation actions, versus 268 and 4,004 for disagreement-based questions. It asks fewer questions while accounting for travel and future decisions. However, its care forecasts used 210,029,302 model transitions versus 162,209,234 for the ensemble without questions and 46,458,463 for ordinary care. That additional planning computation is part of the cost. A test without switching explanations was counted separately; confirmation may still be useful, so this count is not a waste metric.

These are new habitats and a different comparison. Do not interpret the higher percentages than v0.7 as a before/after improvement on identical cases. Protected normal memory and the existing validation gate remain unchanged.

## Maze: timings transfer and cached routes reduce search

| Controller | Goals reached | Mean actions | Evaluation search transitions |
| --- | ---: | ---: | ---: |
| Retained timings |48/48|20|16,848|
| Fresh timings |48/48|25|15,312|
| Retained, replan each action |48/48|20|56,592|
| Wrong timings |12/48|125|196,032|

Training used 36 single-material rooms. The 48 evaluation rooms per arm combine two gates and multiple crates; copper requires continuous pressure, brass/slate have learned release delays. Connections are discovered anew in every room. Retained timing rules saved five actions per room relative to fresh learning, while using somewhat more search. Caching reduced search compared with replanning every action and preserved the observed success/actions. All 48 separately checked witness routes were found.

The grammar is small. Seeds vary reflection and plate indexing; nearby chamber wiring and material classes are authored. The near-identical outcomes across seeds reflect this narrow variation. This is compositional transfer within supplied mechanisms, not unrestricted reasoning about unseen physics.

The separate POET-inspired curriculum generated 18 children, retained all 18, and recorded 4 winning donor transfers, producing 21 archived pairs including the three founders. Every resident and donor attempt was charged. Since every generated child passed admission, this run did not expose a difficult curriculum boundary. It demonstrates the generation/transfer loop, not sustained open-ended growth. Its selected training cases are not held-out evaluation evidence.

## Recorded computation and verification

| Run | Physical actions | Other measured computation | Active worker time |
| --- | ---: | --- | ---: |
| Image comparison |2,938|507,838,464 predicted pixel-channel values|28.35s|
| Composed gate comparison |9,672|290,208 controller +39,696 witness search transitions|11.37s|
| Gate curriculum |1,594|26,220 search transitions|1.23s|
| Nursery comparison |119,304|463,569,193 model transitions including calibration|273.43s|

The bookkeeping total is 133,508 physical actions, with 0 uncertain actions. An image edit, maze move and nursery tick are different units; their sum is not a scientific efficiency score. These counts do not measure FLOPs, energy, setup overhead or browser rendering. Development, tests and interactive journeys are outside these totals.

71 automated tests passed, including eight new v0.8 tests for observation isolation, measured movement, exact undos and replay, protected knowledge, investigation eligibility, learned delays and full cost accounting. The new tests were rerun after the final decision-status correction. Both HTTP suites passed: durable checkpoints, restart/stop behavior, private-file isolation, origin checks and worker authorization. Browser QA covers playback, navigation, imported selection and reload, and desktop/mobile panel widths. The source/evidence audit passed. No persistent experiment remains running.

## What this points to

The demonstrated gains are specific: move a misaligned image control, reuse measured gate timings, and avoid redundant route searches. The nursery's next scientific target is a test that reliably changes a consequential care decision and improves outcomes after its cost. The maze needs substantially more varied geometry and connections before more seeds add much information. Photos need perceptual assessment and richer edit controls before numerical preservation can support broader accreditation. These are directions for another frozen experiment, not capabilities established by this release.