# E72-D/E/F — Five real missions

All five are accepted only from artifacts and test output (`artifacts/e72/real-missions.json`). Each mission uses
authorized workers; every EXECUTE is independently recomputed by a different worker (cross-field where set), and
each has a signed v5 evidence bundle that verifies offline.

| Mission | Source | Result |
|---|---|---|
| Repository engineering | isolated git worktree of this repo | reproduced a real E72 scheduling stall, independent test **failed before / passed after**, 11/11 suite passed, separate verifier worktree confirmed both |
| Defensive security | isolated 30-line Python target | found command injection + path traversal, safe marker exploit reproduced, fixed, exploit blocked, legitimate function retained |
| Research | arXiv Atom API | 6 primary entries, citations preserved, claims/hypotheses separated, lexical tensions labelled unverified |
| Data | public Iris CSV (150 rows) | schema/quality validation, means, Welch comparisons, bootstrap CI, assumptions and multiple-comparison caveat |
| Infrastructure | local host | injected isolated child kill, detection, replacement liveness, unreachable provider classified BLOCKED |

## Repository mission details

The defect: with `soft_cap == 0.9999999999`, `int(math.floor(...))` produced 0 and every task was refused with
`SOFT_CAP` (observed: 24 tasks, 0 verified, 187 refusals, 240 s stall). The detector authored the regression test;
the patcher restored `max(1, int(math.floor(...)))`. The pre-fix file is a **reconstruction of the observed defect**,
not a claim about a pre-existing HEAD bug. The test fails on it and passes on the fix; the E68 suite passes; a second
worktree independently repeats both outcomes.

## Limits

The security target was written for the exercise. The research mission reads abstracts only. The data mission is a
small classic dataset. These are real external sources, not a standardized benchmark.