# v750 research workflow guide

Upload the complete ZIP contents together. The existing seven lab routes and apparatus graphics remain. The main Torus is a full-window Observatory; its original notebook is also retained. Scientific studies are separate from decorative scene layers.

## Collector I–V example

Open DirectConvert → Research tools → Measured collector sweep. Import `examples/collector-fields/synthetic-iv-metadata.json` with the metadata file selector, then choose `synthetic-iv.csv` as Raw I–V CSV. Inspect the populated sign, calibration and electrical-boundary fields; explicitly check the steady-DC declaration, then Import and inspect sweep. The file is synthetic. Signal current and power are different from total measured boundary current and power.

Load `synthetic-frozen-prediction.json` as the prediction packet, set the declared reveal time to `2026-09-11T00:00:00Z`, then Evaluate held-out sweep. Retain the signed residuals and shared covariance. A matching hash checks record integrity; it does not establish when a prediction was made or authenticate an experiment. Export the raw/normalized record and comparison separately.

CSV columns and SI units are displayed in the panel. Original bytes, sign, paired background, shared gain/offset uncertainty and exact boundary are retained. Other noise correlations or pulsed measurements need a different adapter.

## Torus field measurements and scientific studies

Run the default analytic circular field. Open Sensors → Measured field comparison. Import `examples/collector-fields/manufactured-observatory-metadata.json`, select `manufactured-observatory-diagnostics.csv`, then Import and compare field diagnostics. These are manufactured points with an intentional 0.0001 T x-component offset, not real measurements. Using a different field configuration changes their residuals.

Sensors and Compare also contain Uncertainty, evidence and reference studies. Expand a study, edit its labelled fields, and calculate. Study edits discard old output. A later parent run preserves the previous study as historical and retains its native run ID and input hash. A prepared external-reference job is a portable contract; an external solver still has to be selected, run and independently reviewed.

## AlphaTrack registry and evaluation

Open Research tools → Calibrated plate workbench → Registry files and manufactured workflow. Load the manufactured workflow fixture. It contains three generated 8×8 images and synthetic site labels; it is not a three-laboratory study.

In Lock evaluation and freeze method, lock the split and freeze the baseline from shipped source. In Run frozen baseline and build evaluation report, run the imported pixel batch, then build the report. The small example yields one eligible evaluation image with precision, recall and F1 of 1. Physical validation remains false.

Export the complete workbench bundle to retain registry, calibration declarations, annotations, adjudication, split lock, frozen method, predictions and pixel manifest. Import it to reproduce the report. Retain the registry hash independently if mutation detection against a previous record matters.

For real images, use the typed acquisition/calibration controls, independent labels and explicit data permissions. The strict cross-site lock disallows plate/session/site/raw-byte/pixel overlap. A one-site campaign cannot pass that protocol. Generated shapes and empty or failed images do not become measured evidence.

The batch UI handles at most 100 images and one million decoded pixels in portable JSON (16 MiB). The underlying sequential decoder API supports a bounded full-image buffer and tiled segmentation; this is not TIFF/OME streaming. Unsupported large/cut components fail explicitly. Pixel morphometry is not species identification.

The independent Binned detector response study evaluates a declared response matrix, exposure and background. Rank deficiency, nonunique realized-count likelihood or nonconvergence leave incident rates unavailable. The two-bin default is manufactured. Actual image energy/species inversion remains unavailable without validated calibration.

## Heating, uncertainty and direct conversion

Reactor and Virtual inputs include wall-plug-to-absorbed heating efficiency and heater power already included in auxiliaries. Read conditional net and heating-adjusted net together. Required additional heating is debited once; this is a balance of included terms, not evidence of maintained burn.

Torus source studies can export a paired particle-flux record to DirectConvert. Direction, energy, position/time, species, normalization, energy allocation and fate remain paired. Displayed particle count is never substituted for a physical escaping flux.

Joint uncertainty uses the declared covariance and affine model; numerical error, model alternatives and unknown physics are separate. Experiment ranking is a one-step Gaussian design exercise with explicit costs and baselines, not a validated autonomous campaign.

## Replay and continuity

`examples/uncertainty-replay/` contains a frozen historical fixture with 22 hashed artifacts and an independent analytic Python reference. Run `node cli.mjs replay .` inside that folder. Its historical app/kernel tags are retained intentionally.

Use `RESUME-CHECKPOINT.md`, `docs/v750-requirements.json` and `docs/v750-verification/README.md` for the current state. Start from this source tree; do not replace it with an older upload. The matrix distinguishes implemented software, partial research tools, missing data, external validation and deferred specialist work.