Tool Reviews

Review status

Last reviewed: 2026-09-16

Evidence level: Expert judgment

Source quality: Unsourced / illustrative

Snapshot scope: Editorial guide to the sample reviews. Individual reviews retain their own source-review dates.

These pages are worked examples of the project's tool evaluation framework. They are not an awesome list and not endorsements. Each one applies the tool review template to a single, defensible representative tool from one PET family, so you can see how the framework turns "this tool exists" into "this tool fits (or does not fit) a concrete architecture, threat model, and benchmark."

How to read these

Every review answers the same six questions the framework asks:

What does this help build? A pattern or architecture, not a vague PET category.
What protected asset does it cover? Inputs, updates, prompts, embeddings, outputs, logs, or weights.
What adversary is in scope? Honest-but-curious, malicious, colluding, platform, inference, or side-channel.
What evidence supports the claim? Documentation, a proof, a benchmark, a deployment, or an audit.
What breaks in production? Setup, keys, logs, latency, upgrades, debugging.
What benchmark should be run first? A review should produce a first measurement, not a vibe.

Each review ends with a fit label (strong / promising / narrow / poor / unknown), an evidence level, a first benchmark to run, operational risks, when not to use the tool, and a verdict (recommend / watch / reject for the stated use case).

Evidence levels

Reviews use the Evidence Policy. Evidence levels apply to individual claims; a supported claim does not validate every claim on a page. Source quality is recorded separately:

Level	Meaning
Measured	A benchmark states its workload, environment, date, and limits.
Deployment-backed	Named, maturity-labelled production deployments support the claim.
Literature-backed	A specification, security proof, or peer-reviewed paper supports the claim.
Expert judgment	A reasoned conclusion drawn from documentation and design, without independent measurement.
Needs evidence	The claim is plausible but unverified here; treat numbers as uncertain.

Versions, performance numbers, and "latest release" facts drift. Where a fact could not be confirmed as of the review date, it is labelled Needs evidence rather than guessed.

The sample reviews

One representative per PET family:

PET family	Tool	Review	Pair with
Federated learning (FL)	Flower (`flwr`)	flower-fl.md	FL + secure aggregation
Differential privacy (DP)	OpenDP	opendp.md	FL + differential privacy
Homomorphic encryption (HE)	Microsoft SEAL	microsoft-seal.md	HE private inference API
Confidential computing (TEE)	AWS Nitro Enclaves	confidential-computing-platform.md	Confidential RAG

Contributing a review

Want to add one? Read the tool evaluation framework first to confirm the tool earns a page (it must help a reader decide on a concrete architecture, threat model, or benchmark), then copy the tool review template and fill in every field. A review that lists features without saying when not to use the tool is not finished.