# MutationSieve 0.1.0.dev1 release notice

Decision date: 2026-08-01

## Decision

**GO for deterministic software verification only. NO-GO for synthetic-read,
analytical, clinical, diagnostic, treatment-selection, or patient use.**

MutationSieve 0.1.0.dev1 passed its prespecified deterministic software-control
gates. It is **not analytically or clinically validated**. It must not be used
for patient reporting, diagnosis, treatment selection, or to declare a sample
negative. A `REVIEW` record is a research hypothesis, not a reportable result.

The authoritative frozen files record `overall_pass: true` in
`benchmark/benchmark.json`, `benchmark/validation.json`, and
`validation/validation.json`. That boolean means only that all 13 internal
software-control gates passed. It does not mean that an assay, sequencing
workflow, caller, or clinical use has been validated.

## What the frozen control run showed

- 340 tests passed, 3 platform-specific tests were skipped, and 3 unittest
  subtests passed; the JUnit record contains 346 test cases with zero failures
  and zero errors.
- 28 of 29 intention-to-test constructed control alleles were recovered. The
  one miss was the prespecified off-target no-call challenge. All 28 callable
  constructed controls were recovered.
- All 28 selected calls matched a constructed truth allele, and the one small
  deterministic calibration-negative fixture produced zero selected false
  positives across 64,000 fixture-callable bases.
- Every declared class was exercised: SNV 10/10, MNV 2/2, insertion 7/8
  intention-to-test (7/7 callable), deletion 7/7, and complex 2/2.
- All 28 designed evaluation controls were realized in their included runs,
  and all eight evaluated outputs retained a partial-no-call state.
- Automatic `PASS` and `HIGH` result tiers remained disabled; selected records
  were `REVIEW` only.

These are fixture proportions and control-recovery counts, not sensitivity,
specificity, PPV, NPV, LoB, LoD95, precision, reproducibility, or
transportability estimates.

## Mandatory control-construction disclosure

MutationSieve's simulator constructs aligned BAM records directly. It does not
emit FASTQ and run a pinned production aligner against the full intended
reference. Therefore these files are hand-constructed aligned-BAM
unit/regression fixtures, **not synthetic-read validation**.

All 25 positive simulator alleles had a declared three-primary-unit floor, and
24 of 25 required at least one conditional forced top-up. Across alleles, the
ledger records 67 forced top-up primary-unit observations, 8 ordinary-selected
floor contributions, 75 floor-prioritized observations, 131 ordinary-selection
events, and 192 realized allele-primary-unit observations. These are not
necessarily unique physical fragments because one unit can cover more than one
nonoverlapping allele. The three SpikeForge controls were edited by a separate
structural edit-output implementation, but they still inherit the declared
hand-constructed baseline alignment context. That control ran in a fresh
isolated Python process from a private byte-verified copy; all 12 loaded
`spikeforge.*` modules were bound to the approved SpikeForge 1.1.0rc2 source
fingerprint. This code-origin control is not an operating-system sandbox or an
independent biological experiment.

Conditional top-up deliberately makes a control allele observable. Recovery
from that control proves a deterministic code path was exercised; it cannot be
interpreted as biological or analytical sensitivity.

## Work that was not run

Mutect2, Strelka2, LoFreq, DeepSomatic, Lancet2, hap.py/vcfeval, and GIAB or
SEQC2/CASTLE public-truth comparisons were not run in this environment. No
assay-matched FASTQ-to-production-aligner study, physical reference material,
replicate study, interference study, LoB/LoD95 study, cross-platform study, or
blinded clinical lockbox was performed. The external-comparator harness records
checkpoint provenance but is not an OS sandbox or proof of caller identity,
truth blinding, or scientific independence.

## Permitted use

- Source review, unit/regression testing, deterministic software experiments,
  and design of independently governed validation studies.
- Inspection of `REVIEW` hypotheses, rejected candidates, no-call territory,
  manifests, hashes, and control-generation ledgers as research artifacts.
- Reproduction of the same software-control protocol with the exact frozen
  seed and inputs.

## Withheld use

- No automatic patient-reporting tier is available.
- No analytical or clinical performance claim may be derived from the
  software-control pass.
- A future analytical release requires assay-matched physical material,
  FASTQ-to-pinned-aligner evidence, external callers and truth comparisons,
  prespecified denominators, and independent governance.

The canonical archive schema is
`mutationsieve.software-verification-release.v1`. Historical Python APIs whose
names contain `validation` are compatibility aliases only and confer no
certification.