# LGASieve 1.1.1 research validation bundle

Snapshot date: 2026-08-01

This archive supports independent review of LGASieve's source code, automated
tests, and corrected schema-v2 synthetic/software-development results. It does
not contain an orthogonally characterized clinical cohort and cannot validate
clinical sensitivity, specificity, limit of detection, PPV, NPV, or patient
results.

## Five-minute integrity and software check

From the extracted archive root:

```text
python verify_release.py
python -m pip install -r requirements-validation.txt
python -m pytest -q -p no:cacheprovider
```

Expected test result in the recorded environment:

```text
162 passed, 3 skipped, 7 subtests passed
```

The three skips require real sequencing libraries that are not present in the
recovered source. The test suite exercises algorithms, serialization, strict
model loading, cohort separation, provenance, result vocabulary, and failure
states; it is not an analytical validation.

## Quick deterministic simulation check

PowerShell:

```text
.\run_quick_validation.ps1
```

POSIX shell:

```text
sh run_quick_validation.sh
```

The quick profile uses `--verification-seed-offset 1000` and writes new files
under `validation/rerun/`. Scientific metrics are deterministic for the same
software environment, but the complete JSON bytes are not: timestamps,
runtime, and absolute paths are recorded by design.

## Reproduce the frozen full simulation

The archived full result used verification seed offset 300:

```text
python -m scripts.clinical_panel_simulation \
  --verification-seed-offset 300 \
  --output validation/rerun/full.json \
  --model validation/rerun/full-model.json
```

The archived anchors are:

- held-out same-generator 80%-overlap recovery, intention-to-test: 0.8300;
- held-out same-generator completed recovery: 0.9887844801454987;
- held-out same-generator completed no-event referral rate: 0.026170105686965275;
- positive events completed: 3,299 of 4,000;
- within-assay stress 80%-overlap recovery, intention-to-test: 0.4750;
- changed-assay positive and no-event completion: 0.

## Run the research CLI from source

The package metadata references a `LICENSE` file that was absent from the
recovered project. For that reason, this validation snapshot is run directly
from its extracted source directory rather than advertised as an installable
or open-source release.

```text
python -m lgasieve.clinical_depth_cli --help
python -m lgasieve.clinical_depth_cli fit --help
python -m lgasieve.clinical_depth_cli screen --help
```

The broad-panel CLI starts from a complete target count matrix. It does not
include a frozen clinical raw-read-to-count extractor. A real site must supply
complete assay-wide unique-fragment or UMI counts and fit a new model using
disjoint, assay-matched normal cohorts.

## Archive structure

- `lgasieve/`: complete recovered Python package
- `scripts/`: analysis and simulation programs
- `tests/`: complete automated test suite
- `reference/` and `reference_grch37/`: synthetic reference fixtures
- `analysis/`: corrected frozen model and result artifacts
- `work/exomedepth_development_v2/`: small fixtures required by tests
- `results/`: reviewer-friendly tables used by the HTML report
- `CHECKSUMS.sha256`: SHA-256 for every archived file except itself
- `verify_release.py`: no-dependency hash and result-invariant verifier

## Result vocabulary

- `REFLEX_POSITIVE`: research referral hypothesis requiring independent
  confirmation.
- `NO_REFLEX_FLAG`: no referral from the frozen research model; not a clinical
  negative.
- `NO_CANDIDATE_PARTIAL_COVERAGE`: incomplete reportable coverage; no negative
  interpretation is permitted.
- `NO_CALL_QC`: sample-level QC failure.

## Rights and use status

No explicit license file was present in the recovered project. This archive
does not assert an open-source license or grant rights beyond access authorized
by the project owner. It is research/assay-development software, is not a
medical device, and must not be used for patient reporting.