Aegis
A localized medical imaging de-identification engine that handles the cleansing of identifiable metadata in DICOM files, OCR detection of burned-in pixel text, and NER entity classification. It outputs converted files and a queue for manual review, ensuring that protected health information (PHI) does not cross the local trust boundary.
Timeline (last 90 days)
EasyProver's 52.7% pass rate on the Aegis stratified sample shows that implementation details such as fixed-width arithmetic and explicit failure remain difficult, and that the system's performance on larger codebases and longer proof chains, as well as the quality distribution of Aegis-generated specifications, still require independent analysis and verification.
Aegis is a new benchmark for real-implementation verification, and its construction pipeline is: first write a safe Rust solution, then translate it into Lean with Aeneas, then generate specifications, and finally annotate each frozen statement with a kernel-checked proof or refutation.
Aegis generated 2,045 differentially validated Rust-Lean program pairs and 11,533 candidate specifications.
Aegis tasks carry implementation-level obligations, including fixed-width arithmetic, explicit failure, generated loops, type casts, and functionalized mutable state.
On a stratified sample of 2,000 statements, EasyProver proved 1,054 statements, or 52.7%, within a time budget of 30 minutes per statement.
Aegis's frozen statements are annotated with proofs or refutations checked by the Lean kernel, and such kernel-check results can serve as verifiable reward signals for reinforcement learning.
The authors of EasyProver point out that benchmarks such as PutnamBench, IMO 2026, NTP4VC and VERINA are approaching saturation, and have therefore turned to building a new benchmark, Aegis.
Aegis is a local de-identification engine that, combined with Claude's MCP tool-calling capability, forms a privacy-first medical imaging AI Agent architecture, in which the model is only responsible for process orchestration while image pixels and PHI always remain within the local trust boundary.
For DICOM files, Aegis cleans identifiable metadata and checks for burned-in text in pixels that metadata rules cannot remove; OCR detects candidate text regions, NER helps classify sensitive entities, and policy rules determine which content needs to be masked, whitelisted or sent for manual review.
Outputs allowed to flow from Aegis to Claude include aggregate counts, job IDs, cleaned filenames, state transitions and review-needed flags; prohibited outputs include image bytes, pixel crops, OCR text, DICOM tag values, raw exception payloads and re-identification keys.
2 more timeline events
All Facts (12)
EasyProver's 52.7% pass rate on the Aegis stratified sample shows that implementation details such as fixed-width arithmetic and explicit failure remain difficult, and that the system's performance on larger codebases and longer proof chains, as well as the quality distribution of Aegis-generated specifications, still require independent analysis and verification.
50%UnverifiedAegis is a new benchmark for real-implementation verification, and its construction pipeline is: first write a safe Rust solution, then translate it into Lean with Aeneas, then generate specifications, and finally annotate each frozen statement with a kernel-checked proof or refutation.
50%UnverifiedAegis generated 2,045 differentially validated Rust-Lean program pairs and 11,533 candidate specifications.
50%UnverifiedAegis tasks carry implementation-level obligations, including fixed-width arithmetic, explicit failure, generated loops, type casts, and functionalized mutable state.
50%UnverifiedOn a stratified sample of 2,000 statements, EasyProver proved 1,054 statements, or 52.7%, within a time budget of 30 minutes per statement.
50%UnverifiedAegis's frozen statements are annotated with proofs or refutations checked by the Lean kernel, and such kernel-check results can serve as verifiable reward signals for reinforcement learning.
50%UnverifiedThe authors of EasyProver point out that benchmarks such as PutnamBench, IMO 2026, NTP4VC and VERINA are approaching saturation, and have therefore turned to building a new benchmark, Aegis.
50%UnverifiedAegis is a local de-identification engine that, combined with Claude's MCP tool-calling capability, forms a privacy-first medical imaging AI Agent architecture, in which the model is only responsible for process orchestration while image pixels and PHI always remain within the local trust boundary.
50%UnverifiedFor DICOM files, Aegis cleans identifiable metadata and checks for burned-in text in pixels that metadata rules cannot remove; OCR detects candidate text regions, NER helps classify sensitive entities, and policy rules determine which content needs to be masked, whitelisted or sent for manual review.
50%UnverifiedOutputs allowed to flow from Aegis to Claude include aggregate counts, job IDs, cleaned filenames, state transitions and review-needed flags; prohibited outputs include image bytes, pixel crops, OCR text, DICOM tag values, raw exception payloads and re-identification keys.
50%UnverifiedDuring development, synthetic samples should be used instead of real patient data to validate the tools, and both successful and uncertain results should be checked, comparing reported counts against the generated audit rows.
50%UnverifiedBoth Vaultwarden and official Bitwarden support TOTP-based two-step verification and can be used with authenticator apps such as Aegis and Authy.
50%