# The Conservatory 0.8 — decision value, movable controls and composed mechanisms

The new nursery is at `/`, composed gates at `/mechanisms.html`, and raster steering at `/raster.html`. Earlier pages and their evidence remain available. Everything runs on this Mac. Running longer does not by itself establish intelligence or sentience.

## What changed

Every new domain presents the same decision record (also available under Follow the evidence in the original maze): what I expect, what I am testing, what happened, and what I changed. Predictions are formed before the action; outcomes follow it. The labels distinguish missing measurements, failed predictions, and limits of a bounded model or search. “Model limited” is not a proof that a task is impossible.

### Image steering

The two measured image controls can now be recalibrated after an accepted edit. A separate option searches translations of the measured local response, moves the control, and performs fresh charged probes. A 2 × 2 comparison separates these abilities: static, recalibrate only, movable only, and both. Every arm uses the same guard, eight-proposal limit and 64-command budget. Alignment search counts its predicted pixel-channel values.

The target is +0.12 per linear RGB channel inside the selected mask, clipped at 1, with original pixels outside it. Success requires target RMSE≤0.015, outside RMSE≤0.008 and maximum outside change≤0.025. Probes and rejected edits can exceed those limits temporarily; their exact undos and excursions are in the trace. Five procedural conditions cover local, spilling, shifted, nonlinear and image-dependent responses. Twelve predeclared image seeds are evaluated under all four arms. These are correlated conditions within each seed, not240 independent images.

PNG/JPEG/WebP import stays in the browser. A centered square crop is resampled to 128×128, converted from sRGB to linear RGB, and shown as the working original. Draw a rectangular selection on Original; this starts a fresh trial. The PNG export is the 128×128 working image, not a full-resolution photo editor. JSON export contains originals, masks, targets, commands, predictions and outcomes. A synthetic file fixture verifies import/selection/reload; it is not photographic quality evidence. Quota failures leave JSON export available. Procedural benchmarks do not validate subjective quality on photographs.

This concretizes the image-steering-accreditation work's separation of intended change, collateral change and measured capability. It remains a local numerical test with authored controls. GIE-Bench separately assesses editing correctness and content preservation; that motivates separate criteria here, not a claim that our tolerances reproduce its human or learned assessment: [GIE-Bench](https://arxiv.org/abs/2505.11493).

### Nursery investigations

The decision-value planner considers up to three currently supported hypotheses. It evaluates ordinary care journeys and, after suspicion, possible investigations. A test is eligible only if predicted readings split into different follow-up care decisions and the contingent care forecast, including the test's travel and operation costs, beats the best ordinary journey. There is no fixed curiosity bonus. Its 48-tick horizon, finite follow-up choices, approximate weights and measurement grouping make this an approximation. Future weather is held constant. The selected investigation is not guaranteed to help in the actual habitat.

Four equal 480-action arms run on five conditions for each of 12 new habitats: value, the same ensemble without investigations, the previous ordinary planner without investigations, and the previous disagreement-based investigation planner. Shared 182-action calibration and 160-action normal validation are charged once per habitat. Total 119,304 physical actions. The primary contrast is value minus ordinary, with the five-condition mean paired within habitat and a t95 interval over 12 differences. Value minus ensemble is secondary and isolates investigations within the new planner. Report test counts, test actions and model computation separately. A test without a hypothesis switch can confirm a decision; it is not automatically wasted.

The protected normal checkpoint and validation-before-commit mechanism remain the frozen v0.7 implementation. “Dual Model Predictive Control with Gaussian Process” connects informative control actions with future control performance; our finite-hypothesis approximation does not implement its GP model or inherit its guarantees: [paper](https://arxiv.org/abs/2511.08542).

### Composed gates and curriculum

Copper gates require continuous pressure. Brass and slate close after 3 and 5 actions following release. These timings are learned by leaving an activated plate and observing closure. Material colors, movement, crate pushing and the candidate timing grammar are authored. The full board is visible; hidden source bindings and timers do not enter observations. Connections are inferred from plate activations and gate changes. Dotted lines appear only after discovery.

Train on each individual material, then evaluate four unseen two-gate combinations with retained timings, fresh timings, retained timings with replanning every action, and wrong timings. Each room has a 160-action cap. Retained knowledge is copied into each evaluation, never updated across test rooms; connections start unknown. Cached routes are discarded after a gate prediction mismatch. Search is capped at 18,000 transitions per query and depth 70. It protects crates near untested plates from premature pushing. The separate true-rule witness counts search but supplies no controller information. Twelve seeds vary horizontal reflection and plate indexing; they do not provide rich geometry diversity. Maximum 36,480 physical actions including training.

The separate curriculum mutates an archived room by adding/replacing a material, evaluates its resident and up to three donors, and retains solvable pairs requiring 10–120 actions. Three seeds × six generations, up to three gates, maximum 12,000 physical actions including initial training and every donor trial. Transfer moves a small measured material-delay dictionary, with online learning of previously unknown materials. It is a bounded implementation inspired by POET's generation, optimization and transfer loop, not a replication of POET's optimizer or evidence of open-endedness: [POET](https://arxiv.org/abs/1901.01753). Selected curriculum cases are exposed training data, separate from the composed-gate comparison.

## Reproducibility

`experiments/frozen/conservatory-v8/manifest.json` declares the source hashes, seeds, primary contrasts and budgets before evaluation. The service records source and protocol hashes, timestamps, measured and uncertain actions, model/search/pixel-channel computation, and active time. Pausing completes the current bounded seed checkpoint; stopping interrupts it and charges its reserved actions as uncertain. No uncertainty should be silently discarded.

Development seeds and bug fixes are recorded separately under `experiments/development/`. The original coarse alignment search was changed to integer-pixel shifts so odd misalignment could be tested; the maze planner gained a crate-protection exploration rule after a development deadlock. The final nursery correction passes the remaining post-investigation horizon to resource forecasts. These are development changes, never retroactive changes to a scored frozen test. Repeated evaluation on the same seeds reproduces evidence; it adds no independent validation.