Add sanitized Phase-0 research archive under docs/research

Research notes R1-R4, documentation-study files A-E (source of the functional spec's `src:` tags) and PGx coverage CSVs, rewritten to use neutral naming and current project decisions (plan §2).
2026-09-27 16:27:34 +01:00
1 parent 336a6b8226
commit 854cf95029
@@ -0,0 +1,648 @@
+# R1 — the reference product: public-source product dossier
+
+> **Archived Phase-0 study (sanitized copy, see `README.md`).** Written 2026-09-27, before the decisions in
+> `SIMURGARRAY_MASTER_PLAN.md` §2. The analysed vendor product is called "the reference product" and its CLI
+> `refcli`; its documentation URLs are replaced by `[ref-docs:…]` tokens. Where this file speaks of a clean room,
+> legal clearance, or running/benchmarking the vendor binary, plan §2 supersedes it: personal research; MIT/BSD/Apache
+> code may be reused with attribution; GPL tools are subprocess oracles only; vendor binaries are never run for
+> comparison, inspected or decompiled; oracles are published vendor-produced GTCs, gtc2vcf and BeadArrayFiles.
+
+Compiled 2026-09-27 for a reimplementation feasibility study.
+**Only public sources were used.** No binaries were downloaded, run, or inspected.
+
+Tag legend:
+- **[PUBLIC DOCUMENTATION]**: stated in an Illumina (or other named public) source. The URL is given.
+- **[INFERRED]**: a direct logical consequence of public statements. The reasoning is shown.
+- **[HYPOTHESIS]**: a plausible guess that still needs validation.
+
+### Primary source: the official reference-product documentation site
+
+- Site: [ref-docs:home]. It is also mirrored under [ref-docs:mirror].
+- Every page also has a Markdown copy: append `.md` to the page URL. The site index is [ref-docs:llms.txt].
+- Versioned docs live under `v1.0` to `v1.3`. The root path is v1.4, the latest release.
+
+Short aliases used in this report:
+
+| Alias | URL |
+|---|---|
+| `[DOC-HOME]` | [ref-docs:readme] |
+| `[DOC-APPS]` | [ref-docs:our-features] |
+| `[DOC-LOCAL]` | [ref-docs:local-analysis] |
+| `[DOC-IN]` | [ref-docs:input-files] |
+| `[DOC-OUT]` | [ref-docs:output-files] |
+| `[DOC-QC]` | [ref-docs:local-analysis/qc-report] |
+| `[DOC-FAQ]` | [ref-docs:faq] |
+| `[RN-1.0]` | [ref-docs:release-notes-1.0.0] |
+| `[RN-1.1]` | [ref-docs:release-notes-1.1.0] |
+| `[RN-1.2]` | [ref-docs:release-notes-1.2.0] |
+| `[RN-1.3]` | [ref-docs:release-notes-1.3.0] |
+| `[RN-1.4]` | [ref-docs:release-notes-1.4.0] |
+| `[DOC-REV]` | [ref-docs:document-revision-history] |
+| `[DOC-V10-LOCAL]` | [ref-docs:v1.0/local-analysis] |
+| `[DL]` | [ref-docs:downloads] |
+
+---
+
+## 1. What the reference product is
+
+### 1.1 Identity
+
+- **Name and scope.** "[product name] secondary analysis is a … bioinformatics software for Illumina Infinium array-based assays." [PUBLIC DOCUMENTATION] `[DOC-HOME]`
+- **Two packages.**
+  - A local package with a command-line interface: "no specialized server or hardware required".
+  - A cloud package with a GUI, running in the vendor cloud platform.
+  - [PUBLIC DOCUMENTATION] `[DOC-HOME]`, `[DOC-FAQ]`
+- **Binary name.**
+  - The CLI executable is **`refcli`** (`refcli.exe` on Windows). It sits in the `refcli` subfolder of the unpacked installer.
+  - Version check: `refcli version`.
+  - Example install path in the docs: `/usr/local/bin/<refcli>-linux-x64-DAv1.4.0/<refcli>/`.
+  - [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, Illumina Knowledge 000010127 ([ref-docs:kb-000010127])
+- **Installer file names** (from the download page):
+  - `…<Product>v1.4LinuxInstaller-1.4-<refcli>-linux-x64-DAv1.4.0-rc2-sha.9b051a4b….tar.gz` (about 60 MB)
+  - `…-<refcli>-win-x64-DAv1.4.0-rc2-sha.9b051a4b….zip` (about 57 MB)
+  - Both dated May 4, 2026. Downloading requires a MyIllumina login and a customer or instrument serial number.
+  - [PUBLIC DOCUMENTATION] `[DL]`, Knowledge 000010127
+
+### 1.2 Implementation technology
+
+- **.NET application.**
+  - "[product] v1.4 is a self-contained .NET 9 application … Linux users may need to install native OS-level dependencies." [PUBLIC DOCUMENTATION] `[RN-1.4]`, `[DOC-LOCAL]` (Linux Dependencies section)
+  - A config file `refcli.runtimeconfig.json` exists. The docs suggest setting `System.Globalization.Invariant=true` in it as a workaround. [PUBLIC DOCUMENTATION] `[RN-1.0]`…`[RN-1.4]`
+- **Internal namespaces.** Stack traces published in every release note show:
+  - `ArrayAnalysis.CLI.App`
+  - `ArrayAnalysis.Core.GtcFileLoader` (source path `/src/ArrayAnalysis.Core/GtcFileLoader.cs`)
+  - `ArrayAnalysis.Services.GtcFactory`
+  - `ArrayAnalysis.Services.GtcToVcfService`
+  - [PUBLIC DOCUMENTATION] `[RN-1.0]`
+  - Inference: the `refcli` code base descends directly from the vendor's earlier array CLI (ACLI) code base. [INFERRED]
+
+### 1.3 Release history
+
+| Version | Date | Key content | Source |
+|---|---|---|---|
+| **v1.0.0** | Dec 2023 | Local CLI: genotyping (IDAT→GTC→VCF), PGx CNV, native PGx star-allele caller with quality scores and CPIC/DPWG annotation. PGx validated only on GDA-ePGx (`GDA_PGx…E2`). | `[RN-1.0]` |
+| Genotyping Cloud 1.0.0 | Mar 2024 | Cloud genotyping; Final Report and Locus Summary outputs. | [ref-docs:release-notes-1.0.0-cloud-genotype] |
+| Methylation QC Cloud 1.0.0 | May 2024 | Cloud-only methylation QC: 21 control metrics, detection p-values, beta values, M-values. | [ref-docs:release-notes-1.0.0-cloud-methylqc] |
+| **v1.1.0** | Sep 2024 | EX PGx chips (GSAv4-ePGx, GCRA-ePGx). More PGx genes. A custom EGT can be used with the stock CN model (warning only). JSON schema renames. | `[RN-1.1]` |
+| **v1.2.0** | Feb 2025 | Whole-genome **cytogenetics** CNV/LOH calling, VCF output, BAF/LRR bedgraph, ISCN 2020 / cytoband annotation JSON. New `pgx` root command. New license server `[ref-license-server]`. | `[RN-1.2]` |
+| **v1.3.0** | Aug 2025 | Mosaic-fraction estimation. PAR/sex-chromosome improvements. GenomeStudio-compatible sample sheet. Failed samples stay in `gt_sample_summary`. `--smoothing` removed from gtc-to-bedgraph. | `[RN-1.3]` |
+| Methylation QC Cloud 1.0.1 | Feb 2026 | Maintenance release. | [ref-docs:release-notes-1.0.1-cloud-methylqc] |
+| **v1.4.0** | May 2026 (latest) | **QC Report** (`qc call`, `qc report`; "replacing the GenomeStudio control dashboard"). GAINLOH calling. Better mosaic detection. PGx SV star alleles and a new CSV set. `--use-infI-nc-info` for multi-allelic variant (MAV) calling. .NET 9. | `[RN-1.4]` |
+
+All rows above are [PUBLIC DOCUMENTATION]. Additional notes:
+
+- The v1.0.0 cloud pipeline versions were deprecated in Q1 2026. [PUBLIC DOCUMENTATION] `[RN-1.0]`
+- Documentation revisions run from 01 (Dec 2023) to 14 (Aug 2026). Revision 14 renames the vendor cloud platform's data layer. [PUBLIC DOCUMENTATION] `[DOC-REV]`
+
+### 1.4 Relationship to other Illumina tools
+
+**The vendor's earlier array CLI (ACLI)**
+- Illumina calls it "the predecessor to [product]". [PUBLIC DOCUMENTATION] Illumina Knowledge 000003079 (https://knowledge.illumina.com/microarray/general/microarray-general-reference_material-list/000003079)
+- "ACLI was obsolesced in December 2024, and support ended in July 2025." [PUBLIC DOCUMENTATION] (same source)
+- ACLI v2.1 subcommands: `genotype call`, `genotype gtc-to-vcf`, `copy-number call`, `copy-number train`. Options such as `--bpm-manifest`, `--cluster-file`, `--idat-folder` match `refcli`'s option names. [PUBLIC DOCUMENTATION] https://support-docs.illumina.com/ARR/ArrayAnalysisCLI/Content/ARR/IMA/RunAnalysis_Local.htm and https://sapac.support.illumina.com/array/array_software/ima-array-analysis-cli.html
+- The reference product's v1.0 also had top-level `copy-number` and `star-allele` commands. v1.2 regrouped them under `pgx`. [PUBLIC DOCUMENTATION] `[DOC-V10-LOCAL]`, `[RN-1.2]`
+
+**IAAP (Illumina Array Analysis Platform) CLI and AutoConvert**
+- These are earlier .NET GenCall front-ends (IDAT→GTC). [PUBLIC DOCUMENTATION] https://emea.support.illumina.com/downloads/iaap-genotyping-cli.html
+- Third-party gtc2vcf (MIT license) reports that "AutoConvert 2.0, IAAP-CLI, and ACLI … perform version 1.2.0 of the normalization step and seem to produce the exact same results", while the original AutoConvert uses normalization 1.1.2. [PUBLIC DOCUMENTATION – third party] https://github.com/freeseek/gtc2vcf
+- The reference product probably kept normalization 1.2.0 (or a successor) and GenTrain/GenCall-compatible calling. It still reads EGT cluster files and writes GTC v5, and GTCs from AutoConvert/Beeline are accepted by `gtc-to-vcf` as "legacy GTCs". [HYPOTHESIS], supported by `[DOC-OUT]`
+
+**GenomeStudio 2.0**
+- It is still the tool for **creating custom cluster files (EGT)**. The reference product does not cluster; it only consumes EGTs. [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, section "Optimizing cluster files"
+- The reference-product QC Report "replaces GenomeStudio control dashboard". [PUBLIC DOCUMENTATION] `[DOC-FAQ]`, `[RN-1.4]`
+- The cloud Final Report and Locus Summary are "similar to the one available in GenomeStudio". [PUBLIC DOCUMENTATION] `[DOC-OUT]`
+- For PGx cluster review, "GenomeStudio auxiliary files" (`*_GS_import*.zip`) are provided. [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[DL]`
+
+**GTCtoVCF (github.com/Illumina/GTCtoVCF, Apache-2.0, Python 2.7)**
+- Its README says: "As of May 2023, it is now recommended to use … [product] for GTC to VCF conversion." [PUBLIC DOCUMENTATION] https://github.com/Illumina/GTCtoVCF
+- `refcli genotype gtc-to-vcf` reuses GTCtoVCF's option names and near-verbatim documentation text:
+  - `--unsquash-duplicates`, `--auxiliary-loci`, `--filter-loci`, `--disable-genome-cache`
+  - the "Squashing duplicates", "Genome cache", "Auxiliary loci" and "combined genotype" paragraphs, and the dbSNP mapping conventions
+  - Inference: the refcli implementation is a port or evolution of GTCtoVCF logic. That makes GTCtoVCF (open source, Apache-2.0) a strong **legal reference implementation** for the VCF conversion layer. [INFERRED] Compare `[DOC-LOCAL]` with the GTCtoVCF README.
+
+**The vendor cloud platform**
+- The cloud edition of the reference product runs in the vendor cloud platform UI. Data is stored on the vendor cloud platform.
+- An iScan can stream IDATs to the cloud directly.
+- Analysis types in the vendor cloud platform dropdown:
+  - "[product] – Genotyping and QC"
+  - "– PGx – CNV calling"
+  - "– PGx – Star allele annotation"
+  - "– Cytogenetics analysis"
+  - "- Cytogenetics analysis + tertiary interpretation" (cloud only)
+  - "– Methylation – QC" (cloud only)
+- Tertiary-interpretation platform integration uses vendor cloud-platform SNS notifications.
+- [PUBLIC DOCUMENTATION] `[DOC-APPS]`, [ref-docs:cloud-analysis], [ref-docs:cloud-analysis/cytogenetics-analysis]
+- Cloud sample limits: max 1152 samples for genotyping, cytogenetics and methylation; max 384 for the PGx types. [PUBLIC DOCUMENTATION] `[DOC-APPS]`
+
+**Relationship to the vendor's NGS secondary-analysis product of the same brand**
+- The reference product shares only branding with it: no FPGA, a different code base, CPU only.
+- The developer article states: "The software runs on CPUs and does not require a specialized server or FPGA hardware."
+- [PUBLIC DOCUMENTATION] [ref-docs:dev-article-1.0-pgx]
+
+---
+
+## 2. Supported arrays and species
+
+All of this section is [PUBLIC DOCUMENTATION] from `[DOC-APPS]` ("Product & Analysis Compatibility") unless noted.
+
+### 2.1 Genotyping
+
+"Any Infinium genotyping array, including human (including custom and semi-custom) and non-human diploid species".
+
+Tested manifests:
+
+| Array | Manifest | Genome build |
+|---|---|---|
+| ASA | ASA-24v1-0_E2 | GRCh37/38 |
+| BovineSNP50 v3 | BovineSNP50_v3_A1 | UMD3 (non-human) |
+| CGA | CGA-24v1-0_A1 | GRCh37/38 |
+| GCRA | GCRA-24v1-0_20068388_B2 | GRCh37/38 |
+| GDA | GDA-8v1-0_D2 | GRCh37/38 |
+| GSA v3 | GSA-24v3-0_A2 | GRCh37/38 |
+| GSA v4 (Infinium EX) | GSA-48v4-0_20085471_D2 | GRCh37/38 |
+| JSA | JSA-24v1-0_B2 | GRCh37/38 |
+| PRSbooster | PRSbooster_20083382_A | GRCh37 |
+
+Species support:
+- Non-human diploid species work if the user supplies a reference FASTA. For non-human data, turning off VCF output is recommended on cloud.
+- "Genotyping only supports diploid organisms. Polyploid genotyping is currently not supported." `[RN-1.1]`…`[RN-1.4]`
+
+### 2.2 PGx (CNV and star alleles)
+
+Supported only for these commercial ePGx products, all on GRCh38:
+- GDA-ePGx `GDA_PGx-8v1-0_20042614_G4`
+- GSAv4-ePGx `GSA-PGx-48v4-0_20079540_E2`
+- GCRA-ePGx `GCRA-PGx-24v1-0_20084467_C2`
+
+Semi-custom arrays get PGx calls only for the commercial backbone content. `[DOC-LOCAL]` (semi-custom section)
+
+### 2.3 Cytogenetics (CNV/LOH)
+
+Validated arrays:
+- GDA Cyto-8 (`GDACyto-8v1-0_20047166_E2`)
+- GSA Cyto-24 (`GSACyto-24v1_20044998_C2`)
+- CytoSNP-850K v1.4, iScan (`_iScan_B2`) and NextSeq550 (`_NS550_B2`, from v1.3)
+
+Details:
+- Builds GRCh37 and GRCh38.
+- Constitutional samples only; "oncology samples not supported". `[RN-1.2]`, `[RN-1.4]`
+- The feature list says "any human genotyping array" `[RN-1.2]`, but custom products need a cyto model file from Tech Support. `[DOC-IN]`
+
+### 2.4 Methylation
+
+- **In scope, but only as cloud "Methylation QC".** It is **not available in the local CLI** as a methylation pipeline. `[DOC-APPS]`
+- Supported arrays: EPIC v1 (B5 manifest), EPIC v2 (`EPIC-8v2-0_A2`), EPIC-Xtra (custom), and MSA (`MSA-48v1-0_20102838_A1`), all GRCh38.
+- The local `refcli qc call --array-type methylation` does compute methylation **control-probe** QC metrics, and `qc report` can report them. There is no local beta-value, M-value or detection-p pipeline. [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[DOC-QC]`
+- Cloud methylation QC outputs, per sample:
+  - `{Barcode}_{Pos}_ctrl.tsv.gz` (control probes)
+  - `{Barcode}_{Pos}_cgs.tsv.gz` (beta values, M-values, detection p-values)
+- Cloud methylation QC outputs, per batch:
+  - `sample_qc_summary.{xlsx,tsv}`
+  - plots, including a PCA of beta values
+  - `[DOC-OUT]`
+
+### 2.5 Instruments
+
+- IDATs come from iScan. NextSeq550 IDATs are accepted for select arrays. [PUBLIC DOCUMENTATION] `[DOC-IN]`
+
+---
+
+## 3. Workflows and algorithms (what is and isn't disclosed)
+
+### 3.1 Genotyping (IDAT → GTC → VCF)
+
+**What the docs state** [PUBLIC DOCUMENTATION]:
+- Pipeline: `genotype call` (IDAT + BPM + EGT → GTC v5), then `genotype gtc-to-vcf` (GTC + BPM + CSV manifest + FASTA → per-sample `*.snv.vcf.gz` + `.tbi`).
+- `--gencall-cutoff` defaults to **0.15** ("GenCall score cutoff to label a NoCall"). `[DOC-LOCAL]`
+- The GTC holds genotypes, GenCall scores, BAF, LRR and sample QC. It is the "Binary proprietary format", parseable with https://github.com/Illumina/BeadArrayFiles. `[DOC-OUT]`
+- In the SNV VCF, `FILTER` is always `PASS`. FORMAT fields are `GT:GS:BAF:LRR`. Genotypes are ploidy-adjusted: haploid on Y, MT and non-PAR X in males. `[DOC-OUT]`
+- BAF and LRR are oriented REF=A / ALT=B relative to the reference. They may therefore be 1 − the GenomeStudio value. `[DOC-OUT]`
+- Multi-allelic / duplicate-site merging:
+  - Enumerate the candidate genotypes from the queried alleles.
+  - Filter out genotypes inconsistent with each assay's call, respecting design limits. For example, Infinium II cannot tell A/T or C/G apart.
+  - Exactly one survivor becomes the call; zero or more than one gives a no-call.
+  - For merged records, `GS` is the minimum GS across assays. MAV records carry no BAF or LRR.
+  - `--use-infI-nc-info` (v1.4) treats Infinium I no-calls as "missing signal" rather than "probe failure", which removes alleles. It requires a cluster file with "zeroed-out" MAV assays.
+  - `[DOC-OUT]`, section "Note on Multi-Allelic Variants"
+- Probe IDs in the `ID` field are joined with "," for SNPs and ";" for indels. `[DOC-OUT]`
+
+**Algorithms not disclosed for the reference product itself:** GenCall normalization and calling.
+- The GTC stores per-sample `NormalizationTransform` records: version, offset_x/y, scale_x/y, shear, theta. The public BeadArrayFiles code shows how they are **applied**: subtract offset → rotate by theta → shear → scale → clamp ≥0. It then computes R = x+y and θ = atan2(y,x)·2/π. [PUBLIC DOCUMENTATION] https://github.com/Illumina/BeadArrayFiles/blob/develop/module/GenotypeCalls.py
+- How the transform parameters are **fit** is only described qualitatively in the tech notes (see §8). [PUBLIC DOCUMENTATION]
+
+### 3.2 PGx copy number (targeted)
+
+**Scope** [PUBLIC DOCUMENTATION] [ref-docs:pharmacogenomics-pgx-reference/pgx-cnv-coverage]
+- 7 genes across 10 regions (GRCh38): GSTM1, UGT2B17, CYP2E1, SULT1A1, CYP2A6 (intron.7, exon.1), CYP2D6 (exon.9, intron.2, p5), and GSTT1 (on `22_KI270879v1_alt`).
+
+**Batch-based calling** [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`
+- Needs ≥24 samples, with ≥22 passing LogRDev < 0.2. ≥96 are recommended; about 500 GDA-ePGx samples fit per batch on the reference hardware.
+- Error text when too few pass: "{n} sample(s) passed QC. Requires at least {m} samples". `[DOC-OUT]`
+
+**Model file and training** [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[DOC-IN]`
+- Requires a manifest-revision-specific **CN model `.dat`**.
+- The model is trainable with `pgx copy-number train`:
+  - needs ≥96 samples, of which ≥90 pass LogRDev ≤ 0.2; ≥150 are recommended, with diminishing returns beyond 3,000
+  - needs `--platform LCG|EX`
+  - needs a `.msk` mask file next to the BPM, plus the FASTA
+- The trainer "is designed with the assumption that the copy number distribution resembles the standard population distributions".
+- v1.4 turns an EGT mismatch between GTC and CN model into a hard error. v1.1–1.3 only warned. `[RN-1.4]`
+
+**Output** [PUBLIC DOCUMENTATION] `[DOC-OUT]`
+- `*.cnv.vcf.gz` with `<CNV>` records, `FORMAT CN:NR`. CN is capped at 5 ("5 or 5+"). NR is the "aggregated normalized intensity".
+- QUAL is Phred-scaled, capped at 60. `Q7` filter below 7; `SampleQuality` filter when LogRDev > 0.2.
+- Header keys `##OverallPloidy=` and `##GCCorrect=True`, which implies GC correction of intensities.
+- Batch files: `cn_batch_summary.csv` (gain/loss/neutral per region) and a CN summary file.
+
+**Algorithm** [HYPOTHESIS]
+- Not disclosed. The developer article gives accuracy only: CYP2D6 CNV "98.3%–99.0%" versus qPCR on 288 samples.
+- Likely shape: batch-normalized, GC-corrected, probe-aggregated normalized R (the "NR" field), compared against per-region reference distributions in the `.dat` model (population-trained). Plausibly a Gaussian-mixture or likelihood model over CN states 0–5.
+
+### 3.3 PGx star alleles
+
+**Commands and licensing** [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`
+- `pgx star-allele call` (VCF folder with `*.snv.vcf.gz` + `*.cnv.vcf.gz`, plus a PGx database `.zip`) → `star_alleles.dat`.
+- Then `pgx star-allele annotate --guidelines CPIC|DPWG` → CSV and JSON.
+- **Star-allele calling is the only paid, license-server-gated local step.** It is charged per sample through `--license-server-url https://user:pass@[ref-license-server]` (v1.2+), or `[ref-license-server-legacy]` for v1.0/1.1.
+
+**Data sources** [PUBLIC DOCUMENTATION] [ref-docs:pharmacogenomics-pgx-reference/pgx-allele-definitions-and-pgx-guidelines]
+- PharmVar 6.1, PharmGKB snapshot 2024-05-16, UGT nomenclature, legacy CYP nomenclature, CPIC 1.38.0, DPWG June 2023.
+- CPIC phenotypes cover 17 genes. Activity scores are given for CYP2C9, CYP2D6 and DPYD.
+- Coverage: "2400+ targets for over 50 genes" `[DOC-HOME]`; the product page says "1700+ targets for 50+ genes".
+
+**Scoring and reporting** [PUBLIC DOCUMENTATION] `[RN-1.0]`, `[DOC-OUT]`
+- Quality score is a "negative log likelihood" in v1.0.
+- Later, "a probability score … based on the SNV and CNV inputs, input quality, and population frequencies" (0–1), plus a `Raw Score` that excludes population frequency.
+- Ranked alternative solutions; hybrid alleles; allele-specific CN.
+- Alternatives are shown if their score is ≥50% of the top score.
+- No novel alleles are called.
+
+**PGx database format** [PUBLIC DOCUMENTATION] `[DOC-IN]`
+- "each line … a single probe ID mapping to a variant's HGVS tag". The public database zip is named `DAv1.4.0-rc2-pgx-mapping-sha.b164b4aa….zip`. `[DL]`
+
+**v1.4 CSV outputs** [PUBLIC DOCUMENTATION] `[DOC-OUT]`
+- `diplotype_summary.csv`
+- `diplotype_annotations.csv`
+- `supporting_variants.csv`
+- `unresolved_variants.csv`
+- `variants_not_covered_on_array.csv`
+- plus a per-sample star-allele JSON. Its fields are listed in `[DOC-OUT]`.
+
+**Algorithm** [HYPOTHESIS]
+- Not disclosed. It is described only as "a new and native PGx star allele caller" (developer article).
+- Possible approach: a haplotype-pair likelihood search over PharmVar definitions. It would take CN constraints from the CNV VCF and error or missing-data tolerance, and use population-frequency priors, which would explain the "Score" versus "Raw Score" split.
+- The v1.x error "The model has been changed since the solution was last computed" is typical of Google OR-Tools / a MIP/CP solver. That suggests an **optimization (ILP/CP-SAT) formulation**. [HYPOTHESIS]
+
+### 3.4 Cytogenetics CNV/LOH (genome-wide)
+
+**Commands** [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`
+- `cyto call --cn-model cyto_model.dat` (GTC → `*.cnv.vcf.gz` + `cyto.cnv.dat`).
+- `cyto annotate --annotation-db CytoAnnotateData_*.zip` (→ per-sample JSON with ISCN 2020 notation, cytobands, RefSeq/Ensembl genes, QC).
+
+**Model file** [PUBLIC DOCUMENTATION] `[DOC-IN]`
+- The cyto model "primarily contains probe GC content information and is tied to the probe sequence/manifest".
+- It does **not** need retraining when the EGT changes.
+
+**Parameters** [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`
+
+| Event type | Default minimum probes | Default minimum size |
+|---|---|---|
+| DEL | 10 | 25 kb |
+| DUP | 10 | 50 kb |
+| LOH | 500 | 3000 kb |
+| GAINLOH | 300 | 500 kb |
+| mosaic DEL (mdel) | 10 | 25 kb |
+| mosaic DUP (mdup) | 10 | 50 kb |
+| mosaic LOH (mloh) | 500 | 3000 kb |
+
+Other parameters:
+- `--smoothing 5` (probes on each side)
+- `--max-mosaic-fraction 1.0`
+- `--min-qual 20` (annotate)
+- Minimum probe thresholds below 10 are reset to 10.
+
+**Behaviours documented** [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[DOC-OUT]`, `[RN-1.2]`…`[RN-1.4]`
+- CN up to 4+.
+- QUAL = −10·log10(1 − P(called state)), capped at 60; `Q10` filter.
+- LogRDev > 0.2 → `SampleQuality` flag. LogRDev > 0.3 → only whole-chromosome events are reported.
+- Mosaic events are detected down to about 15% for one-copy mosaic states. `MF` holds the mosaic fraction.
+- LOHTYPE ∈ {AOH, CNLOH, GAINLOH}.
+- LOH < 500 kb flanked by copy-neutral regions is removed (v1.4).
+- Variants must match the expected LRR sign (v1.4).
+- Effective size excludes gaps larger than 150× the median probe spacing.
+- An unknown sex is resolved by the cyto caller or else assumed male.
+- `GT` is hardcoded to `1/1`.
+
+**Algorithm** [HYPOTHESIS]
+- Not disclosed. The outputs (per-state posterior → Phred QUAL; joint LRR+BAF states including LOH, GAINLOH and mosaic fractions; GC correction; smoothing window) are consistent with an **HMM or segmentation + state-likelihood model over LRR and BAF**, in the family of PennCNV / cnvPartition / the reference product "cytogenetics modality".
+- cnvPartition is Illumina's older published recursive-partition approach (see §8).
+
+### 3.5 QC (v1.4)
+
+[PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[DOC-QC]`, `[DOC-OUT]`
+- `qc call --array-type genotyping|methylation` → `controls.raw_metrics.csv`, `controls.qc_metrics.csv`. These files are also emitted automatically by `genotype call`.
+- `qc report` → `<Product>_QC_Report_YYYY_MM_DD.html` (self-contained, interactive) and `<Product>_QC_table_YYYY_MM_DD.{xlsx|csv}`.
+
+### 3.6 Other items
+
+- **Locus Report / Final Report / Locus Summary.**
+  - Cloud-only outputs: `gtc_final_report.csv` (GenomeStudio-style) and `locus_summary.csv`, with columns Call_Freq, Het_Excess_Freq, ChiTest_P100, Cluster_Sep, AA/AB/BB θ and R mean/SD, and so on.
+  - No locus report exists in the local CLI.
+  - [PUBLIC DOCUMENTATION] `[DOC-OUT]`, `[DOC-APPS]`
+- **Cluster file generation.** Not supported; it is done in GenomeStudio 2.0. [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`
+- **BedGraph.** `genotype gtc-to-bedgraph` → `*.LRR.bedgraph`, `*.BAF.bedgraph` for IGV. [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[RN-1.2]`
+
+---
+
+## 4. Command-line interface (v1.4, exact)
+
+Syntax: `refcli [module] [sub-module] [command] [required] [optional]`. Options are case-sensitive. [PUBLIC DOCUMENTATION] `[DOC-LOCAL]`
+
+The following commands are all [PUBLIC DOCUMENTATION] from `[DOC-LOCAL]`:
+
+```
+refcli version | help
+refcli qc call        --array-type {genotyping|methylation} --csv-manifest F (--idat-folder D | --sample-sheet F) [--output-folder] [--debug] [--json-log]
+refcli qc report      --data D1[,D2..|PARENT] [--config cfg.yaml] [--label L1,..] [--output-folder] [--output-format xlsx|csv]
+                       [--threshold-callRate 0.98] [--threshold-logRdev 0.20] [--debug] [--json-log]
+refcli genotype call  --bpm-manifest F.bpm --cluster-file F.egt (--idat-folder D | --sample-sheet F) [--gencall-cutoff 0.15]
+                       [--num-threads N] [--output-folder] [--debug] [--json-log]
+refcli genotype gtc-to-bedgraph --bpm-manifest F.bpm (--gtc-folder D | --sample-sheet F) [--output-folder] ...
+refcli genotype gtc-to-vcf --bpm-manifest F.bpm --csv-manifest F.csv --genome-fasta-file G.fa (--gtc-folder D | --sample-sheet F)
+                       [--auxiliary-loci aux.vcf] [--disable-genome-cache] [--filter-loci list.txt] [--unsquash-duplicates]
+                       [--no-bgzip] [--use-infI-nc-info] [--output-folder]
+refcli pgx copy-number call  --cn-model M.dat (--gtc-folder D | --sample-sheet F) [--no-bgzip] [--output-folder]
+refcli pgx copy-number train --bpm-manifest F.bpm (mask .msk beside it) --genome-fasta-file G.fa --platform {LCG|EX}
+                       (--gtc-folder D | --sample-sheet F) [--disable-genome-cache] [--output-folder]
+refcli pgx star-allele call  --database pgx.zip --license-server-url URL --vcf-folder D [--query-license-quota] [--output-folder]
+refcli pgx star-allele annotate --star-alleles star_alleles.dat [--guidelines CPIC|DPWG] [--output-folder]
+refcli cyto call      --cn-model cyto_model.dat (--gtc-folder D | --sample-sheet F) [--no-bgzip] [--smoothing 5]
+                       [--min-{del,dup,loh,gainloh,mdel,mdup,mloh}-{probes,size} ...] [--max-mosaic-fraction 1.0] [--output-folder]
+refcli cyto annotate  --annotation-db CytoAnnotateData.zip --vcf-folder D [--min-* ...] [--min-qual 20] [--output-folder]
+```
+
+Common flags on every command: `--debug`, `--json-log`, `--help`, `--version`.
+
+Behaviour notes:
+- Output folders are overwritten on re-run.
+- A `--idat-folder` is searched recursively.
+- Duplicate IDAT pairs per barcode_position are unsupported.
+
+### 4.1 Inputs
+
+[PUBLIC DOCUMENTATION] `[DOC-IN]`
+
+| Input | Extension | Used by |
+|---|---|---|
+| IDAT (Grn/Red pair per `Barcode_RxxCxx`) | `.idat` | genotype call, qc call |
+| BPM manifest | `.bpm` | genotype call/gtc-to-vcf/gtc-to-bedgraph, pgx cn train |
+| CSV manifest (with SourceSeq, RefStrand; dbSNP-style 1-based mapping) | `.csv` | gtc-to-vcf, qc call |
+| Cluster file | `.egt` | genotype call |
+| PGx CN model | `.dat` | pgx copy-number call |
+| Cyto model | `.dat` | cyto call |
+| Mask | `.msk` (beside the BPM) | pgx copy-number train |
+| PGx database | `.zip` | star-allele call |
+| Cyto annotation database (RefSeq + Ensembl) | `.zip` | cyto annotate |
+| Genome FASTA + `.fai`; contigs "1", not "chr1" | `.fa` | gtc-to-vcf, pgx cn train |
+| QC config | `.yaml` / `.yml` | qc report |
+
+Sample sheet (CSV):
+- Keys are `SentrixBarcode_A,SentrixPosition_A`. On cloud they are `beadChipName,sampleSectionName`.
+- Optional `[Header]` with `RootFolder` and `[Data]` sections. The format is compatible with GenomeStudio sheets, and `[Manifests]` is ignored.
+- Column names are case-insensitive, and reserved names are rejected.
+- Values are limited to 500 characters.
+- Sample-name precedence: sheet `Sample_Name` > IDAT/GTC metadata > `Barcode_Position`.
+
+Public product files: GRCh37/GRCh38 FASTA zips, per-product BPM/CSV/EGT, CN/cyto model and database zips, QC YAML templates, and demo IDATs.
+- Examples: GDA-ePGx demo data set (iScan), 15.4 GB; GCRA/GSA-PGx `idats_Demo_36.zip`.
+- [PUBLIC DOCUMENTATION] `[DL]`
+- Note: model and database files are Illumina data files; their licence terms were not reviewed.
+
+### 4.2 Outputs
+
+[PUBLIC DOCUMENTATION] `[DOC-OUT]`
+
+**Per sample**
+- `.gtc` (GTC v5)
+- `*.snv.vcf.gz` + `.tbi`. Header `##source=refcli 1.4.0`, VCFv4.1. Tabix output is not byte-identical to `bcftools index`.
+- `*.LRR.bedgraph`, `*.BAF.bedgraph`
+- PGx `*.cnv.vcf.gz`
+- cyto `*.cnv.vcf.gz`; header `##source=refcli 1.4.0 Cyto`, `##product=…`
+- cyto annotation `*.json`
+- star-allele JSON
+- methylation (cloud only): `_ctrl.tsv.gz`, `_cgs.tsv.gz`
+
+**Per batch**
+- `gt_sample_summary.csv` and `.json`. Columns: `Sample ID, Sample Name, Sample Folder, Autosomal Call Rate, Call Rate, Log R Ratio Std Dev, Sex Estimate, [TGA_Ctrl_5716 Norm R], <samplesheet columns>`.
+- `controls.raw_metrics.csv`, `controls.qc_metrics.csv`
+- `<Product>_QC_Report_*.html`, `<Product>_QC_table_*.xlsx|csv`
+- `cn_batch_summary.csv`, CN summary
+- `star_alleles.dat`
+- the five PGx CSVs (§3.3)
+- `cyto.cnv.dat`
+- cloud only: `gtc_final_report.csv`, `locus_summary.csv`, `sample_qc_summary.{xlsx,tsv}`
+
+**Logs**
+- Warnings and errors go to the terminal, optionally as JSON (`--json-log`).
+- Error strings are catalogued in `[DOC-OUT]`, for example "Skipping intensity only locus", "Failed to combine genotypes due to inconsistencies …", and "Cluster file (…) is not the same as CN Model Cluster file".
+
+---
+
+## 5. QC metrics and thresholds
+
+### 5.1 Functional per-sample metrics (`gt_sample_summary`)
+
+[PUBLIC DOCUMENTATION] `[DOC-OUT]`, `[DOC-QC]`, `[DOC-FAQ]`
+- Metrics: Autosomal Call Rate, Call Rate, Log R Ratio Std Dev (LogRDev), Sex Estimate (M/F/Unknown), and TGA_Ctrl_5716 Norm R.
+- TGA_Ctrl_5716 Norm R is present on PGx products only. It is the normalized R of one probe; a value < 1 means the targeted gene amplification (TGA) failed.
+- Failed samples get call rates of 0 and NaN for LogRDev (v1.3+).
+
+**Documented thresholds** [PUBLIC DOCUMENTATION]
+
+| Metric / rule | Value | Source |
+|---|---|---|
+| `qc report` default callRate | 0.98 (PASS if Autosomal Call Rate ≥ threshold) | `[DOC-LOCAL]`, `[DOC-QC]` |
+| `qc report` default logRDev | 0.20 (PASS if ≤) | `[DOC-LOCAL]`, `[DOC-QC]` |
+| Generic genotyping YAML example logRDev | 0.15 | `[DOC-QC]` |
+| PGx CNV: per-sample pass | LogRDev < 0.2 (≤ 0.2 for training) | `[DOC-LOCAL]` |
+| PGx CNV: batch minimum | ≥ 22 of ≥ 24 passing | `[DOC-LOCAL]` |
+| PGx star-allele: review triggers | LogRDev > 0.2, call rate < 0.99, TGA control < 1.0 | `[DOC-LOCAL]` |
+| Cyto: low-quality flag | LogRDev > 0.2 → SampleQuality | `[DOC-LOCAL]` |
+| Cyto: whole-chromosome-only mode | LogRDev > 0.3 | `[DOC-LOCAL]` |
+
+**Not defined in the reference-product docs**
+- The exact **Sex Estimate** algorithm.
+  - [HYPOTHESIS] It likely mirrors GenCall/GenomeStudio "GenderEst" (X heterozygosity plus Y intensity, with an autosomal call-rate gate). A public QC SOP describes GenomeStudio as estimating sex only when autosomal call rate > 0.97: https://khp-informatics.github.io/COPILOT/GenomeStudio_genotyping_SOP.html
+  - Needs validation against GTC outputs.
+- p10 GC / p50 GC ("GC10"/"GC50") and p05/p50/p95 percentiles.
+  - These are GTC fields (see BeadArrayFiles TOC IDs 1009, 1011, 1014, 1015), but they are not reported in `gt_sample_summary`. [PUBLIC DOCUMENTATION] https://github.com/Illumina/BeadArrayFiles
+  - "p99 GC" was not found in any reference-product document.
+- LogRDev computation details, such as which loci are included and whether outliers are trimmed. The cyto JSON defines it as "Standard deviation of the Log R ratio values for all probes". [PUBLIC DOCUMENTATION] `[DOC-OUT]`
+
+### 5.2 Control-probe QC metrics (`controls.qc_metrics.csv`)
+
+[PUBLIC DOCUMENTATION] `[DOC-OUT]`
+
+Definitions:
+- `bkg_grn` = max green intensity of Extension(A|T).
+- `bkg_red` = max red intensity of Extension(C|G).
+- Offset x = 3000.
+
+Genotyping metrics:
+
+| Metric | Formula |
+|---|---|
+| RestorationQC | Restoration green / (bkg_grn + x) |
+| StainingGreenQC | Biotin(High) / Biotin(Bgnd) |
+| StainingRedQC | DNP(High) / DNP(Bgnd) |
+| ExtensionGreenQC | min Ext(C,G) green / max Ext(A,T) green |
+| ExtensionRedQC | min Ext(A,T) red / max Ext(C,G) red |
+| Hyb High/Med and Hyb Med/Low | green ratios |
+| TargetRemovalIQC | (bkg_grn + x) / TargetRemoval green |
+| StringencyQC | String(PM) red / String(MM) red |
+| NonSpecificBinding Green / Red | (bkg + x) / max NSB(Bgnd) |
+| NonPolymorphic Green / Red | min NP(C,G) / max NP(A,T) green; min NP(A,T) / max NP(C,G) red |
+
+Methylation adds:
+- TargetRemovalII
+- Specificity I green/red, Specificity II, Specificity II background
+- Bisulfite Conversion I green/red (with background variants), Bisulfite Conversion II (with background variant)
+- that is, the 21 BeadArray Controls Reporter metrics.
+
+Genotyping YAML template example values `[DOC-QC]`:
+
+| Threshold | Example value |
+|---|---|
+| staining | 5 |
+| extension | 5 |
+| targetRemovalI | 5 |
+| hyb ratios | 1.25 |
+| NP | 5 |
+| stringency | 1.5 |
+| NSB | 5 |
+| tgaControl | 1 |
+| restoration | null (1 when the FFPE Restore kit is used) |
+
+- Chemistry-specific templates: `config_genotyping.yaml`, `config_genotyping_EX.yaml`, `config_methylation_EPIC.yaml`, `config_methylation_MSA.yaml` (downloads on `[DL]`).
+- Methylation cloud defaults (EPIC / MSA) are tabulated at [ref-docs:cloud-analysis/methylation-qc]. Includes BgCorrectionOffset 3000 and PvalThreshold 0.05.
+- Methylation algorithm [PUBLIC DOCUMENTATION] from tech note M-GL-02644 ([ref-docs:methylation-qc-tech-note-M-GL-02644]):
+  - **Noob** background correction
+  - **nonlinear** dye-bias correction
+  - detection p-values using "Negative controls for null, log transformed signal"
+  - GenomeStudio, by contrast, uses background subtraction, linear dye bias, and a negative-control null.
+
+---
+
+## 6. Hardware, OS, performance
+
+[PUBLIC DOCUMENTATION] `[DOC-LOCAL]`, `[DOC-FAQ]`, developer article
+
+- **CPU only.** "does not require a specialized server or FPGA hardware". The local package is "CPU-based".
+- **Recommended machine.**
+  - 8 cores.
+  - Memory: **32 GB** in v1.4; 16 GB in v1.0 `[DOC-V10-LOCAL]`.
+  - ≥30 GB disk; `cyto annotate` needs ≥5 GB free.
+- **Operating systems.** Windows 10+ (win10-x64); CentOS 7+ / Ubuntu 20.04+ (linux-x64).
+  - .NET 9 self-contained, needing native .NET dependencies.
+  - Should not run as admin.
+  - No ARM build is listed. [INFERRED] from the runtime IDs listed (win10-x64, linux-x64).
+- **Docker.**
+  - No official Docker image is documented for the reference product. [INFERRED] from its absence in the docs.
+  - The cloud runs in the vendor cloud platform; container use there is not documented.
+  - IAAP had a Docker-based "Orchestrated Workflow". [PUBLIC DOCUMENTATION] https://emea.support.illumina.com/downloads/iaap-genotyping-orchestrated-workflow.html
+- **Performance claims.**
+  - "Genotyping VCF in as little as 35 seconds per sample". `[DOC-HOME]`
+  - "all PGx analysis results in ~1 minute per sample". `[DOC-HOME]`
+  - Accuracy >99.5%. `[DOC-HOME]`
+  - PGx CNV >95%. `[DOC-HOME]`
+  - CYP2D6 call rate >97%. `[DOC-HOME]` The product page says >98%.
+- **Threading and scale.**
+  - `--num-threads` exists on `genotype call`.
+  - More than 1K samples per sample sheet slows things down (known issue).
+  - Windows and Linux star-allele results differ slightly: 6 field differences in 1,576 samples, with concordant top solutions. `[RN-1.3]`
+- **Network requirement.**
+  - Internet access is needed for the license check, **only for `pgx star-allele call`**.
+  - Everything else runs offline. [INFERRED] Only star-allele call is quota-gated per `[DOC-LOCAL]` and `[DOC-APPS]` ("Local: No cost download" for genotyping, CNV and cyto).
+
+---
+
+## 7. Licensing / EULA
+
+- **Product-specific EULA text** was not found at a public URL. It is presumably shown in the installer or on MyIllumina download. [HYPOTHESIS] Action: obtain it legally and have counsel review it before any benchmarking with the real binary. *(Superseded: the plan does not run, benchmark or inspect the vendor binary; see plan §2.)*
+- **Commercial model.**
+  - Local genotyping, PGx CNV and cyto are a "No cost download".
+  - Local star-allele calling uses per-sample quota.
+  - Cloud uses vendor cloud credits, plus a per-sample PGx charge and a vendor cloud-platform subscription.
+  - [PUBLIC DOCUMENTATION] `[DOC-APPS]`, product page [ref-docs:product-page]
+- **Representative vendor EULA language** [PUBLIC DOCUMENTATION] the vendor's sequencing-pipeline v4.3 EULA, [ref-docs:vendor-sequencing-eula-v4.3]. This is the same family and is likely similar for the reference product [HYPOTHESIS]:
+  > "Licensee may not modify, improve, translate, reverse engineer, decompile, disassemble, or create derivative works of the Software or otherwise attempt to … (b) derive the source code or the underlying ideas, algorithms, structure, or organization form of the Software."
+  - The same EULA says Licensee may not "publish" the Software; the restriction is on the Software itself, not on results.
+  - No explicit benchmarking or publication-of-results clause was found in the passages reviewed. [PUBLIC DOCUMENTATION] same PDF.
+- **Implications** [INFERRED]
+  - A clean-room effort should rely only on public docs, public tech notes, open-source Illumina code, and observed input/output formats.
+  - It must not decompile the .NET assemblies. .NET is easily decompilable, which makes this contractually sensitive.
+  - Running the licensed binary to generate comparison outputs is probably permitted "internal use", but the "derive … underlying ideas, algorithms" clause is broad. That needs legal review. *(Superseded: the plan does not run, benchmark or inspect the vendor binary; see plan §2.)*
+- **Open-source Illumina pieces usable as references** (§9):
+  - BeadArrayFiles: BSD-2-style licence text in its README.
+  - GTCtoVCF: Apache-2.0.
+
+---
+
+## 8. Illumina public algorithm documents (URLs verified live, HTTP 200)
+
+| Document | URL | Relevance |
+|---|---|---|
+| Illumina GenCall Data Analysis Software (tech spotlight) | https://www.illumina.com/Documents/products/technotes/technote_gencall_data_analysis_software.pdf | Per-array self-normalization (channel intensity, background, dye crosstalk). GenTrain clustering with neural-net estimation of missing clusters. Bayesian calling. GenCall score = GenTrain score × data-to-model fit, penalised by the DNA score. GC10/GC50. Sensitive range 0.2–0.7. |
+| Infinium Genotyping Data Analysis (tech note) | https://www.illumina.com/Documents/products/technotes/technote_infinium_genotyping_data_analysis.pdf (alt: https://support.illumina.com/content/dam/illumina-marketing/documents/services/technote_infinium_genotyping_data_analysis.pdf) | Normalization, clustering, QC workflow, and cluster-editing practice. Cited by the reference product docs `[DOC-LOCAL]`. |
+| GenTrain 2.0 | https://www.illumina.com/Documents/products/technotes/technote_gentrain2.pdf | Three-step θ/R clustering: preliminary, model proposals, HWE-aware scoring. |
+| GenTrain 3.0 (M-GL-01258) | https://www.illumina.com/content/dam/illumina/gcs/assembled-assets/marketing-literature/gentrain-tech-note-m-gl-01258/gentrain-tech-note-m-gl-01258.pdf | Normalization "bins" (Inf I and Inf II separate). A two-parameter model for Inf I bins with < 192 loci, versus six parameters otherwise. Limits the penalty on near-axis homozygous clusters. |
+| X-chromosome SNP cluster generation | https://www.illumina.com/documents/products/technotes/technote_x_chromosome_snp_cluster_generation.pdf | Sex-aware clustering. |
+| DNA Copy Number Analysis Algorithms (cnvPartition basis) | https://www.illumina.com/documents/products/technotes/technote_cnv_algorithms.pdf | LRR/BAF definitions; recursive-partition CNV. |
+| Cytogenetic analysis tech note | https://www.illumina.com/Documents/products/technotes/technote_cytoanalysis.pdf | LRR/BAF cytogenetics. |
+| cnvPartition in GenomeStudio (training PDF) | https://support.illumina.com/content/dam/illumina-support/courses/GenomeStudio-cnvPartition/story_content/external_files/GenomeStudio_cnvPartition.pdf | |
+| CNV plug-ins for BeadStudio | https://www.illumina.com/content/dam/illumina-marketing/documents/products/technotes/technote_cnv_plug_ins.pdf | |
+| Custom cluster file creation for improved copy number analysis (M-GL-02142) | https://www.illumina.com/content/dam/illumina/gcs/assembled-assets/marketing-literature/custom-cluster-file-tech-note-m-gl-02142/custom-cluster-file-tech-note-m-gl-02142.pdf | Cited by the reference product for EGT/CN-model retraining. |
+| [product] Methylation QC (M-GL-02644) | [ref-docs:methylation-qc-tech-note-M-GL-02644] | Noob, nonlinear dye bias, detection-p method. |
+| iSelect / Infinium I vs II design | https://www.illumina.com/Documents/products/technotes/technote_iselect_design.pdf | Needed for MAV logic. |
+| Infinium Controls training guide | https://support.illumina.com/content/dam/illumina-support/courses/eval-inf-controls/story_content/external_files/Infinium_Controls_Training_Guide.pdf | Control-probe semantics. |
+| Developer article: [product] 1.0 PGx | [ref-docs:dev-article-1.0-pgx] | Accuracy numbers; CPU-only. |
+
+- Tech-note summaries in the table are [PUBLIC DOCUMENTATION], extracted from the PDFs with pdftotext.
+- **Key gap:** none of these gives equations for the normalization-transform **fitting**, the GenCall data-to-model score, or GenTrain's neural-net cluster imputation.
+- Published third-party descriptions and reimplementations of Illumina-like calling exist: gtc2vcf (MIT), which re-implements BAF/LRR from EGT; and the WARP Illumina genotyping pipeline, https://broadinstitute.github.io/warp/docs/Pipelines/Illumina_Genotyping_Arrays_Pipeline/README (it uses iaap-cli). [PUBLIC DOCUMENTATION]
+
+---
+
+## 9. Illumina GitHub repositories relevant to arrays
+
+The GitHub API was rate-limited (unauthenticated), so repos were checked through their public HTML/raw pages. A complete org listing still needs to be done (`gh api orgs/Illumina/repos --paginate` with a token).
+
+| Repo | Licence | Relevance |
+|---|---|---|
+| **Illumina/BeadArrayFiles** — https://github.com/Illumina/BeadArrayFiles | BSD-style (text in README, "Copyright (c) 2020, Illumina") | Python parsers for **GTC (v5 spec: `docs/GTC_File_Format_v5.pdf`)**, **BPM**, **EGT**, and LocusAggregate. `NormalizationTransform.normalize_intensities` and `rect_to_polar` give the exact application of the stored affine normalization. The README recommends the reference product for making GTCs. [PUBLIC DOCUMENTATION] |
+| **Illumina/GTCtoVCF** — https://github.com/Illumina/GTCtoVCF | Apache-2.0 | Semi-archived since May 2023 in favour of the reference product. Same option set and merge semantics as `refcli genotype gtc-to-vcf` (§1.4). Includes unit tests (`tests/test_class.py`) on genotype merging. [PUBLIC DOCUMENTATION] |
+| Illumina/IlluminaHumanMethylationEPICv2manifest, Illumina/IlluminaHumanMethylationMSAmanifest | not checked | R/Bioconductor-style manifests for EPIC v2 and MSA; useful for methylation QC. [PUBLIC DOCUMENTATION] (the repos exist) |
+| Illumina/Cyrius | "Other" licence (not SPDX-identified) | CYP2D6 caller for **WGS**, not arrays. Conceptual reference for CYP2D6 star-allele and CN logic only. |
+| Illumina/canvas, Illumina/SMNCopyNumberCaller, Illumina/Gauchian, Illumina/ExpansionHunter, Illumina/hap.py, Illumina/witty.er | "Other" licence (not SPDX-identified) | NGS tools. hap.py and witty.er are useful for benchmarking VCF concordance. |
+
+Related non-Illumina open source:
+- freeseek/gtc2vcf (MIT): IDAT/BPM/EGT/GTC → VCF with BAF/LRR recomputation. https://github.com/freeseek/gtc2vcf
+
+---
+
+## 10. Biggest unknowns and risks for an independent reimplementation
+
+1. **GenCall internals** (normalization fitting, per-locus calling likelihood, GenCall score formula, DNA-score penalty, sex estimation).
+   - Only qualitative tech notes are public.
+   - Reproducing bit-compatible GTC genotypes and GC scores will need empirical calibration against reference outputs.
+   - The GTC stores the per-sample normalization transform, which helps with verification.
+2. **PGx CNV model** (`.dat` format and algorithm) and **cyto model** (`.dat`: "primarily … GC content"). Both formats and both algorithms are undocumented.
+3. **Star-allele caller**: scoring function, priors, solver, hybrid-allele handling. The PGx database zip format is only loosely described ("probe ID → HGVS" lines).
+4. **Cyto CNV/LOH caller**: the segmentation or HMM approach, the mosaic-fraction estimator, GC correction, and QUAL calibration are all undocumented.
+5. **LogRDev, Autosomal Call Rate and Sex Estimate**: the exact locus sets and formulas are not specified.
+6. **EULA**:
+   - The product-specific terms are not public.
+   - The vendor-family EULA forbids deriving "underlying ideas, algorithms" from the Software.
+   - Comparison benchmarking should be legally cleared. *(Superseded: the plan does not run, benchmark or inspect the vendor binary; see plan §2.)*
+   - .NET decompilation must be strictly avoided.
+7. **Clearly reproducible from public material**: VCF conversion (GTCtoVCF), GTC/BPM/EGT I/O (BeadArrayFiles), control-probe QC formulas (fully tabulated in `[DOC-OUT]`), methylation Noob/dye-bias/beta/M-value methods (open literature plus the tech note), and output schemas (VCF headers, CSV/JSON fields).