v0.2.1

[0.2.1] 2026-11-09

Documentation

• Complete citation metadata and bump it on release (#2)
• Name current small models in the canary ADR alternatives (#4)
• Map the threat model to OWASP LLM 2026 and MITRE ATLAS v2026.10 (#5)
• Diagrams on white in the shared theme, make diagram for every diagram, and a self-contained Definition of Done (#7)

Build and CI

• Lock transformers 5.22.2 (was 4.57.6) and confirm the canary benchmark reproduces exactly (#6)

Maintenance

• Latest toolchain and CI versions (#1)
• Release script keeps hard-coded version strings in step (#3)

v0.2.0

[0.2.0] 2026-11-07

Features

• Add membership inference results against DP-SGD on Purchase100 (#12)
• Add DP-SGD privacy/utility curves on Purchase100 (#15)
• Add canary exposure and extraction results on Pythia-70M (#16)

Performance

• Bound the canary benchmark to five-digit secrets and release MPS memory (#10)
• Make the canary benchmark resumable and score sequences in batches of 8 (#11)
• Fine-tune the canary model in chunks of 4 to stay under the MPS memory cap (#13)
• Checkpoint canary model fine-tuning every epoch and halve the enumeration batch (#14)

Maintenance

• Set the version to 0.2.0 and fix the citation metadata (#17)

v0.1.0

[0.1.0] 2026-11-07

Features

• Add RDP, PRV and GDP accountants validated against Opacus (#1)
• Add membership inference attacks, ROC metrics and DP-SGD (#2)
• Add LiRA reference scoring, canary exposure and LM extraction (#3)
• Add HTTP service and attack report command (#4)
• Add Purchase100 results, DP-SGD families and figures (#6)
• Add accountant parity with Opacus and Rust core speed benchmark (#7)

Documentation

• Add ADRs and the architecture diagram (#5)
• Write the README, evaluation report, figures and demos (#9)

Maintenance

• Scaffold project
• Add release tooling, changelog and agent notes (#8)