[JacobStephens2]

Summary

Experiment
  Trial → session sandbox → provider proxy
  Code Grader → offline sandbox → report packet + neutral repository
  Model Grader → session sandbox → blind packet + provider proxy

Graders cannot read Trial logs, Arm mappings or the Experiment's other files.
Unavailable Trial/Code Grader sandboxes produce Void Trials; an unavailable
Model Grader pauses grading so resume can reuse completed Trials.

Seams under test: the whole Experiment/Trial binary against fixture Harnesses,
and the existing bwrap launcher tests. Ordinary integration fixtures retain
the pass-through launcher; new isolation cases run real bwrap.

Closes #1014

Evidence

• Before: cargo test --test experiments a_sandboxed_model_grader_sees_only_its_packet -- --nocapture failed because the Grader could access host paths.
  After: the same test passes with real bwrap.
• Before: the Code Grader isolation test failed on host access and then its Arm-identifying origin URL.
  After: cargo test --test experiments code_graders_see_reports_and_repository_but_no_trial_logs_or_arm_paths -- --nocapture passes.
• Before: bwrap startup failures were reported as failed Trials or checks.
  After: cargo test --test trial a_bwrap_startup_failure_makes_a_trial_void -- --nocapture passes with sandbox_unavailable.
• Review regressions: reproduced the mount metadata leak, launcher exec failure, lost I/O error cause and false void caused by a check's own bwrap: diagnostic. Kept every failing regression. cargo test --test experiments sandbox:: -- --nocapture passes all four named sandbox tests; the exact startup unit test and both pass-through Trial regressions pass.
• cargo check, cargo fmt --check and cargo clippy --all-targets -- -D warnings pass.
• /usr/bin/time -p scripts/affected-tests 5536630954daf4934f6157522413b1420d331a4f passed at 4a714600: 2,535 Rust tests passed, 58 skipped; 43 site checks passed. Rust execution took 1,226.74 s; all affected checks took 1,339.87 s including compilation (50.50 s) and site checks.
• The final change only adds the existing bwrap-availability skip convention to the check-diagnostic regression. Its exact test passes with real bwrap and reports SKIPPED with bwrap absent from PATH; formatting passes.

Unaddressed findings

Standards

None. Both findings addressed: shared toolchain setup and preserved startup I/O errors.

Spec

None. All four findings reproduced and addressed: Arm paths in mount metadata, launcher exec failures, pass-through startup classification and checks whose diagnostics mention bwrap.

Security

None unaddressed. Two introduced findings reproduced with failing tests and fixed: a Trial-planted .git symlink redirecting the code Grader's origin-URL rewrite outside its snapshot (sandbox::tests::code_grader_git_rewrite_stays_inside_its_snapshot), and special files (sockets, FIFOs) in a Trial repository failing the Grader snapshot copy and aborting the whole Experiment (evals::experiment::grading::tests::grader_snapshot_skips_special_files). Both regression tests pass after the fix.

Security review outcome

No unaddressed introduced findings.

Review coverage

Independent Standards and Spec reviews used issue-977 fixed at 5536630954daf4934f6157522413b1420d331a4f. Both axes read all 18 changed files, including generated site/docs/index.html. Changed files left unread: none. Generated-output exclusions: none. Reports, combined files-read lists and verified dispositions are recorded beside the Session trace.

Merge Danger

Door: two-way

Blast Radius: Experiments

Code checks receive report copies at neutral paths; raw Trial logs are no
longer available through THIRDSHIFT_TRIAL_LOG. Model Graders require the
same bwrap, socat and login setup as Trials.

Built with codex · gpt-6.1-sol · xhigh