Scope, out-of-scope and roadmap

ProofPack computes and records the performance statistics you declare criteria for, and renders them as attachments. This page is the list of things it deliberately does not do — permanently, not “not yet” — plus the guidance documents it maps outputs to and the dated roadmap. Every item below is also stated in the T8 run manifest of every pack.

Permanently out of scope

Area	ProofPack does not do
Study and evidence design	Study design; Sample-size calculation; Selection of acceptance criteria, margins or thresholds; Adjudication / reference-standard construction; MRMC / reader studies; Clinical-outcome evidence; Literature review; Valid clinical association.
Submission sections we do not author	Benefit-risk; Risk management; Cybersecurity documentation; Human factors; Explainability; 510(k) Summary authorship (21 CFR 807.92).
Post-market	Complaints / vigilance; Incident trend reporting (MDR Art. 88; GB SI 2024/1368 reg. 44ZN).
Company scope	Hospital buyers; SOC 2 in year 1; DE/AT cold email; LinkedIn scraping; Any LLM-composed customer- or prospect-facing text.

Reporting conventions are not acceptance criteria. The vendor-set numbers in ProofPack are k-suppression (n<10, events<5, non-events<5), AUROC “not evaluable” below 10 positives or 10 negatives, low-n tiers (n<10, n<30), the 200/200 calibration-curve flag and the bootstrap B. Each is labelled “ProofPack reporting convention” wherever it appears, and none of them produces a status word.

Guidance map v1

Packs resolve section anchors through a versioned guidance map, so that a renumbered draft changes the map rather than the pack. Every anchor to a draft document renders the words “per draft guidance (Jan 2025), not for implementation”.

Document	Version / date	Status	Used for
FDA AI-enabled device software functions: lifecycle management and marketing submission recommendations	January 2025	draft	T1 evidence attachment set — §VIII data management, App. C performance, App. E model card
Predetermined Change Control Plans for machine-learning-enabled device software functions	December 2024, updated August 2025	final	T2 PCCP performance-evaluation report — §VII.B(3)
Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests	2007	final	Proportions, PPA/NPA routing, both-way indeterminates, subgroup reporting

Section numbers in guidance map v1 are transcribed from the published PDFs. The machine-readable map ships with the engine.

Public roadmap

Release	When	Contents
v1.0.0	Target 30 September 2026	Binary classification only. CSV in, customer-authored criteria.yaml, twelve HALT gates, proportions / discrimination / calibration / subgroups / fairness gaps / paired version comparison, T1, T2, T7, T8, T9, T12 documents, pip package plus one linux/amd64 image, Ed25519 offline licence.
v1.1	Dated when v1.0 ships	Multi-class, regression and segmentation ingest; drift (PSI) and the T3 monitoring document; robustness (LOSO, threshold sensitivity, score noise); intersectional subgroups; AUPRC; BCa bootstrap; optional model card; agreement statistics; analytic clustered AUROC variance once the citation is verified.
v1.2	Not dated	T5 GMLP and T6 crosswalk documents. No promise is made here about FDA finalising the January 2025 draft, and no ProofPack date depends on it.