Exam runner v0.4.0

Exams an agent cannot recognise. Each sitting draws a random sample of the exam's questions inside the enclave, uses one of the author's phrasings, can have a TEE-hosted NEAR AI model reword it (kept only if every number survives), and frames the request like ordinary chat traffic: no fixed test instruction, a varied answer marker, and a system message only half the time. Keeps v0.3.0's 25-second cap per answer. Reports proctor_version 0.4.0.

SHA-256 4ee3e3359cd1f0d44ae2e47b3f9e287542449a19103ce56eab3ea0882a522ebe, built by GitHub Actions from PROCTOR commit 9f22371 with scripts/build-runner.sh (remapped paths) and reproduced on a separate CI run.

Exam runner v0.3.0

Each answer attempt is capped at 25 s with one retry, so a slow model no longer runs into OutLayer's 30 s per-request limit, which ends the whole run. Completion ids the registry would refuse, replies over 256 KB, and repeated ids now count as failed calls (the sitting is incomplete), so an agent endpoint cannot choose a rejected outcome. Reports proctor_version 0.3.0. SHA-256 96abb4a702c2ec060604ef4040a54dce95b57bffc872b3a091016ab57cd5ddae, built by GitHub Actions from PROCTOR commit 5652e47 with scripts/build-runner.sh (remapped paths) and reproduced on a separate CI run.

Exam runner v0.2.0

Agent endpoints (NEAR AI key sent only to cloud-api.near.ai), retry without temperature for models that refuse 0, and a reproducible build. SHA-256 9a41b2354d7c508536343dc3585e07cb3a8e579c358126ba3f01a88489b9dde4, built by GitHub Actions from PROCTOR commit 08e4528 with scripts/build-runner.sh (rustc 1.99.1, remapped paths) and reproduced on later CI runs.

Exam runner v0.1.0

SHA-256 3cf89b3425f8e8c4029d9cfa34ed9a1925f23dca7a80ddf8d7d74ca927f69c99 · 410,297 bytes · wasm32-wasip2 · built from PROCTOR commit e3ca823.